From 6737756515462e265e5a9e759ce18f88ef2705e6 Mon Sep 17 00:00:00 2001 From: Holly Grimm Date: Fri, 17 Jul 2026 16:30:14 -0600 Subject: [PATCH 1/4] =?UTF-8?q?data:=20WhisperX=20transcripts=20for=206=20?= =?UTF-8?q?items=20=E2=80=94=20first=20database-free=20run?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Speaker-diarized transcripts (large-v3 + pyannote) for GuestStream_128 and Fundamentals Sessions 023-027, produced by transcribe_worklist.py from the journal-derived worklist. GuestStream_128 speakers identified via the new parts[].speakers mapping (SPEAKER_00 pending confirmation); sessions await their speaker pass. SCHEMA.md documents the mapping. Worklist: every publicly fetchable video now has a transcript. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_015NjFiMWkgrVy4foGF1Zbax --- INDEX.json | 12 +- INDEX.md | 12 +- .../GuestStream/GuestStream_128/metadata.json | 7 +- .../GuestStream_128/transcript.json | 1 + .../GuestStream_128/transcript.txt | 1310 ++++++++++++++ .../Cohort_1/Session_023/transcript.json | 1 + .../Cohort_1/Session_023/transcript.txt | 1508 +++++++++++++++++ .../Cohort_1/Session_024/transcript.json | 1 + .../Cohort_1/Session_024/transcript.txt | 1262 ++++++++++++++ .../Cohort_1/Session_025/transcript.json | 1 + .../Cohort_1/Session_025/transcript.txt | 1006 +++++++++++ .../Cohort_1/Session_026/transcript.json | 1 + .../Cohort_1/Session_026/transcript.txt | 1172 +++++++++++++ .../Cohort_1/Session_027/transcript.json | 1 + .../Cohort_1/Session_027/transcript.txt | 672 ++++++++ docs/SCHEMA.md | 5 + 16 files changed, 6959 insertions(+), 13 deletions(-) create mode 100644 data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json create mode 100644 data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.json create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.txt create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.json create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.txt create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.json create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.txt create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.json create mode 100644 data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.txt diff --git a/INDEX.json b/INDEX.json index 8a24ca391..d56356906 100644 --- a/INDEX.json +++ b/INDEX.json @@ -1625,7 +1625,7 @@ "parts": [ "vjtYYbO9jCY" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/Insights/Insights_001", @@ -3921,7 +3921,7 @@ "parts": [ "MjeQeWeyYhE" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024", @@ -3930,7 +3930,7 @@ "parts": [ "FtGnkRJwk6U" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025", @@ -3939,7 +3939,7 @@ "parts": [ "V46B4Xy1PeQ" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026", @@ -3948,7 +3948,7 @@ "parts": [ "FnYCaCkah4U" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027", @@ -3957,7 +3957,7 @@ "parts": [ "SBCl8DeaJYU" ], - "has_transcript": false + "has_transcript": true }, { "path": "data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/qwjwbiJ1LLw", diff --git a/INDEX.md b/INDEX.md index e84aa9b59..28fbd47ec 100644 --- a/INDEX.md +++ b/INDEX.md @@ -186,7 +186,7 @@ - ✓ [`GuestStream_125`](data/video/activeinferenceinstitute/GuestStream/GuestStream_125/) — 1 video record(s) - ✓ [`GuestStream_126`](data/video/activeinferenceinstitute/GuestStream/GuestStream_126/) — 1 video record(s) - ✓ [`GuestStream_127`](data/video/activeinferenceinstitute/GuestStream/GuestStream_127/) — 1 video record(s) -- · [`GuestStream_128`](data/video/activeinferenceinstitute/GuestStream/GuestStream_128/) — 1 video record(s) +- ✓ [`GuestStream_128`](data/video/activeinferenceinstitute/GuestStream/GuestStream_128/) — 1 video record(s) ## Insights (24) - ✓ [`Insights_001`](data/video/activeinferenceinstitute/Insights/Insights_001/) — 1 video record(s) @@ -449,11 +449,11 @@ - ✓ [`Session_020`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_020/) — 1 video record(s) - ✓ [`Session_021`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_021/) — 1 video record(s) - ✓ [`Session_022`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_022/) — 1 video record(s) -- · [`Session_023`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/) — 1 video record(s) -- · [`Session_024`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/) — 1 video record(s) -- · [`Session_025`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/) — 1 video record(s) -- · [`Session_026`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/) — 1 video record(s) -- · [`Session_027`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/) — 1 video record(s) +- ✓ [`Session_023`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/) — 1 video record(s) +- ✓ [`Session_024`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/) — 1 video record(s) +- ✓ [`Session_025`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/) — 1 video record(s) +- ✓ [`Session_026`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/) — 1 video record(s) +- ✓ [`Session_027`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/) — 1 video record(s) - ✓ [`qwjwbiJ1LLw`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/qwjwbiJ1LLw/) — 1 video record(s) - ✓ [`tmqPfDfQ5-s`](data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/tmqPfDfQ5-s/) — 1 video record(s) - ✓ [`Meeting_001`](data/video/activeinferenceinstitute/TextbookGroup/ParrPezzuloFriston2022/Cohort_1/Meeting_001/) — 1 video record(s) diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json index 60ccb34ff..0fbbf114b 100644 --- a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json @@ -25,7 +25,12 @@ "url": "https://www.youtube.com/watch?v=vjtYYbO9jCY", "title": "ActInf GuestStream 128.1: Active Inference as the Test-Time Scaling Law for Physical AI Agents", "duration": null, - "upload_date": "" + "upload_date": "", + "speakers": { + "SPEAKER_03": "Daniel Friedman", + "SPEAKER_05": "Omar Hashash", + "SPEAKER_04": "Christo Kurisummoottil Thomas" + } } ] } diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json new file mode 100644 index 000000000..43c36e2eb --- /dev/null +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json @@ -0,0 +1 @@ +[{"video_id": "vjtYYbO9jCY", "segments": [{"start": 13.193, "end": 14.215, "text": " Hello, welcome.", "speaker": "Daniel Friedman"}, {"start": 15.117, "end": 16.68, "text": "It's July 14th, 2026.", "speaker": "SPEAKER_00"}, {"start": 17.06, "end": 24.194, "text": "We're in active guest stream 128.1 on active inference as the test time scaling law for physical AI agents.", "speaker": "SPEAKER_00"}, {"start": 25.076, "end": 29.785, "text": "Thank you, Omar and Christo, for joining, and please take it away for the presentation.", "speaker": "SPEAKER_00"}, {"start": 32.819, "end": 33.34, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 34.882, "end": 38.145, "text": "Thank you, Daniel, for having us today in this presentation.", "speaker": "Omar Hashash"}, {"start": 40.268, "end": 44.693, "text": "Today, as you said, we'll be presenting one of our new works.", "speaker": "Omar Hashash"}, {"start": 45.334, "end": 49.579, "text": "So first of all, for the people that don't know me, I'm Omar Hashash.", "speaker": "Omar Hashash"}, {"start": 49.599, "end": 51.982, "text": "I'm a postdoc at Virginia Tech.", "speaker": "Omar Hashash"}, {"start": 52.683, "end": 60.733, "text": "And today I'll be also joined with my colleague, Christa Thomas, which is an assistant professor at WPI.", "speaker": "Omar Hashash"}, {"start": 61.372, "end": 68.162, "text": " And together we'll be presenting one of our recent works, which is active inference as a test time scaling law for physical AI.", "speaker": "Omar Hashash"}, {"start": 68.802, "end": 70.485, "text": "So it's a bit of an interesting work.", "speaker": "Omar Hashash"}, {"start": 70.685, "end": 84.645, "text": "It gives a bit of new concepts to the role of active inference and how it plays it in physical AI and why is it really necessary to be posed as a scaling law for physical AI agents.", "speaker": "Omar Hashash"}, {"start": 85.766, "end": 89.151, "text": "So just as a brief", "speaker": "Omar Hashash"}, {"start": 89.772, "end": 116.742, "text": " introduction uh so i recently finished my phd at the brand the department of electrical computer engineering at virginia tech back in december 2025 and where my focus was on wireless communications and ai including topics like world models and digital twins and edge intelligence and i'll pass it over here to crystal just a little bit to introduce himself", "speaker": "Omar Hashash"}, {"start": 118.105, "end": 140.907, "text": " uh yeah thanks for more uh i'm christopher thomas i am currently an assistant professor at uh wpi and i lead a research group uh named trustworthy resilient ai and networks um and i did my prayer to this i did my post postdoc associate at virginia tech ec department", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 140.887, "end": 151.417, "text": " And my research mainly lies at the intersection of AI, native wireless networks, semantic communication, physical AI, and mathematical foundations of AI.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 153.379, "end": 163.288, "text": "Yeah, so both of us have this kind of hybrid background on AI and wireless communications at the same time.", "speaker": "Omar Hashash"}, {"start": 163.808, "end": 170.775, "text": "And we're going to see how that plays a role in extending us more into neuroscience concepts like active inference.", "speaker": "Omar Hashash"}, {"start": 171.042, "end": 180.511, "text": " So without any further notice, I think we can start first with the fact that AI has become the most transformational technology that is changing all our worlds.", "speaker": "Omar Hashash"}, {"start": 180.932, "end": 191.102, "text": "And of course, one of those worlds is actually the wireless communications world, and especially the 6G, the next generation of wireless communications that we're expecting to see roughly in a couple of years.", "speaker": "Omar Hashash"}, {"start": 191.983, "end": 198.429, "text": "So we already started to touch on improvements with the current generations of AI that we have.", "speaker": "Omar Hashash"}, {"start": 198.489, "end": 201.032, "text": "We call these networks as AI-native networks.", "speaker": "Omar Hashash"}, {"start": 201.012, "end": 205.487, "text": " We started to see improvements in terms of efficiency and latency.", "speaker": "Omar Hashash"}, {"start": 205.747, "end": 207.272, "text": "The improvements are massive, actually.", "speaker": "Omar Hashash"}, {"start": 207.834, "end": 209.821, "text": "But that's only the good part of the story.", "speaker": "Omar Hashash"}, {"start": 210.463, "end": 214.195, "text": "But there's also, as any other story, there's also the bad and the ugly.", "speaker": "Omar Hashash"}, {"start": 214.968, "end": 216.791, "text": " So how do the bad and the ugly look like?", "speaker": "Omar Hashash"}, {"start": 217.171, "end": 218.193, "text": "Well, they look something like this.", "speaker": "Omar Hashash"}, {"start": 219.034, "end": 230.25, "text": "So the fact that we're training what we call as an AI native interface, for example, we train it to send beams to users in the network.", "speaker": "Omar Hashash"}, {"start": 230.27, "end": 233.555, "text": "So these users can be humans, they can be physical agents.", "speaker": "Omar Hashash"}, {"start": 234.196, "end": 237.28, "text": "So a part of these", "speaker": "Omar Hashash"}, {"start": 237.26, "end": 266.355, "text": " uh beams actually work very well but the fact is that the world is always non-stationary and dynamic and it's always changing so these so these kind of uh tasks hold some sort of problems and even fail at test time um so even if we look at the level of the agents themselves that are connected to the network we can see that we have plenty of examples on failures as well so for example we have autonomous vehicles that are trained on millions and millions of driving trials", "speaker": "Omar Hashash"}, {"start": 266.977, "end": 271.406, "text": " They still continue to fail in ways that we just can't comprehend in the world.", "speaker": "Omar Hashash"}, {"start": 272.428, "end": 286.054, "text": "At the same time, we have these kind of robots that are trained to win Olympics, for example, like this robot over here, but still crash into a man on an athletics track, even though it's trained to win an Olympics.", "speaker": "Omar Hashash"}, {"start": 286.996, "end": 289.761, "text": "So there's a lot of things that don't make a lot of sense over here.", "speaker": "Omar Hashash"}, {"start": 290.939, "end": 299.869, "text": " And ironically, with the current generation of AI, we do have these big, large language models that can actually win in math and math.", "speaker": "Omar Hashash"}, {"start": 299.967, "end": 317.171, "text": " So I think here now we can pose a question, which is, why is it that today's AI is good enough to win math Olympiad, but it fails at this type of Olympiad, or more generally, in very simple green world scenarios that are very easy for us, especially with this massive amount of training?", "speaker": "Omar Hashash"}, {"start": 317.912, "end": 325.502, "text": "I'm not going to answer this now, but I'm going to use it as a motivation for the rest of the presentation and to say that we're actually missing something very big over here.", "speaker": "Omar Hashash"}, {"start": 326.562, "end": 353.076, "text": " so so if i want to look now at what do we want from these ai ages so the real the physical agents the robots the the vehicle the network so what we want from these ai agents is to become adaptive to the worlds that are happening we can start from the first one the fact that the world is always changing uh we need to train it on something but there's always going to be more data that we didn't train on so we need to have a um", "speaker": "Omar Hashash"}, {"start": 353.545, "end": 357.969, "text": " a principled way that we act in the states that we didn't actually see before.", "speaker": "Omar Hashash"}, {"start": 359.231, "end": 362.894, "text": "At the same time, we want those agents to be generalizable.", "speaker": "Omar Hashash"}, {"start": 363.155, "end": 367.979, "text": "So we want them to train them on something, but at the same time, to be able to use them on something else.", "speaker": "Omar Hashash"}, {"start": 368.94, "end": 370.622, "text": "At the same time, we want to be autonomous.", "speaker": "Omar Hashash"}, {"start": 370.742, "end": 376.768, "text": "We don't want to stay in the loop or probably minimize our presence in the loop with the AI system as much as possible.", "speaker": "Omar Hashash"}, {"start": 377.709, "end": 381.493, "text": "At the same time, we want them to have this continual learning capability.", "speaker": "Omar Hashash"}, {"start": 381.625, "end": 385.329, "text": " So we wanted to learn on top and top and top of other things.", "speaker": "Omar Hashash"}, {"start": 385.349, "end": 396.36, "text": "So we're starting from a very far away part over here with the current generation of AI, which is mostly based on large language models like the GPT models that we have, which are a massive success.", "speaker": "Omar Hashash"}, {"start": 397.301, "end": 408.713, "text": "So part of the success that started with GPT, at least the new part, was introducing these reasoning capabilities.", "speaker": "Omar Hashash"}, {"start": 408.76, "end": 423.918, "text": " So I think back with the old models, the old three models, the old one model, we started to see how this reasoning can pop up into these models and actually give you a better answer for the agents.", "speaker": "Omar Hashash"}, {"start": 424.92, "end": 436.033, "text": "So for the first time, we had this new kind of test time scaling law, which started from the fact that if you give it more computed test time, you give the agent more computed test time, you can have a better answer.", "speaker": "Omar Hashash"}, {"start": 436.756, "end": 440.842, "text": " So this was followed by another kind of capability, which is called planning.", "speaker": "Omar Hashash"}, {"start": 442.024, "end": 447.532, "text": "And this is what ushered in the agentic era or AI agents that we see today.", "speaker": "Omar Hashash"}, {"start": 448.133, "end": 452.059, "text": "And this is where GPT-5 and all the clouds that we have today.", "speaker": "Omar Hashash"}, {"start": 452.639, "end": 459.169, "text": "So basically, planning is the capability of breaking down a complex task into a sequence of actions in order to achieve a goal.", "speaker": "Omar Hashash"}, {"start": 459.723, "end": 464.077, "text": " So this is where we currently roughly are right now, probably more or less.", "speaker": "Omar Hashash"}, {"start": 464.88, "end": 468.231, "text": "But we can see here that we're moving on a cognitive path.", "speaker": "Omar Hashash"}, {"start": 468.295, "end": 492.379, "text": " so what we can say is that the current forms of reasoning and planning are good for language models are probably good enough for language models but are not enough for real world agents because the world works in a different way than language so the world is driven by laws that works apart from language how language works in terms of syntax and all those stuff so but that that route seems to be right", "speaker": "Omar Hashash"}, {"start": 493.186, "end": 515.565, "text": " so if we were to build on these cognitive capabilities and the fact that they were starting to work so this means that there should be other cognitive capabilities that we're searching for in order to reach the characteristics of the AI systems that we're searching for so if you want to dig deep here and try to search for this missing cognitive capability that's needed to move forward with the world", "speaker": "Omar Hashash"}, {"start": 515.545, "end": 523.48, "text": " we can see that it needs to be a cognitive capability that is in direct touch with the world and it's focused on the world.", "speaker": "Omar Hashash"}, {"start": 523.797, "end": 529.185, "text": " And that capability is nothing but something called common sense.", "speaker": "Omar Hashash"}, {"start": 529.947, "end": 530.988, "text": "So what is common sense?", "speaker": "Omar Hashash"}, {"start": 531.189, "end": 540.644, "text": "Common sense is, as the name implies, it's the common amount of knowledge of how the world works that we humans use in order to make sense of our world.", "speaker": "Omar Hashash"}, {"start": 541.104, "end": 548.556, "text": "It's our ability to understand the world and make use of it, which is something that is found in each human of us.", "speaker": "Omar Hashash"}, {"start": 549.548, "end": 551.252, "text": " So that's great.", "speaker": "Omar Hashash"}, {"start": 551.432, "end": 553.036, "text": "How can we plug in this common sense?", "speaker": "Omar Hashash"}, {"start": 553.437, "end": 564.982, "text": "So common sense is a broadly defined term, but we kind of know if you want to introduce this common sense into AI age, we have to search for the component of the cognitive system that drives it.", "speaker": "Omar Hashash"}, {"start": 565.523, "end": 569.211, "text": "So that component is actually what we call as a work model.", "speaker": "Omar Hashash"}, {"start": 569.832, "end": 570.793, "text": " So what is a world model?", "speaker": "Omar Hashash"}, {"start": 571.054, "end": 583.772, "text": "A world model of the physical world allows humans to understand the world in terms of its real-time state, in terms of the causal structures that exist between the elements of the world.", "speaker": "Omar Hashash"}, {"start": 584.733, "end": 590.001, "text": "And at the same time, it allows us to understand the dynamical evolution of how this world will evolve with time.", "speaker": "Omar Hashash"}, {"start": 590.403, "end": 616.495, "text": " so but the introduction of the world model itself so the fact that you're taking a state of the world and trying to understand it and trying to see how it's going to evolve in the future this requires other cognitive capabilities such as perception for example so perception is the capability to actually understand the world and try to grasp it and to make meaning of it but if we try to zoom out now at this specific moment", "speaker": "Omar Hashash"}, {"start": 616.93, "end": 620.255, "text": " What are we trying to say with all of these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 620.295, "end": 628.046, "text": "So we said that we have that ladder that started with reasoning, we added planning, we added common sense for the world model, and now we're saying that we need perception.", "speaker": "Omar Hashash"}, {"start": 628.687, "end": 631.011, "text": "What are we trying to say with all these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 631.912, "end": 642.027, "text": "What we're indirectly acknowledging over here is the fact that we need a cognitive solution that includes all these cognitive capabilities on top of today's AI.", "speaker": "Omar Hashash"}, {"start": 643.255, "end": 645.239, "text": " So, and this is not something new.", "speaker": "Omar Hashash"}, {"start": 645.279, "end": 654.796, "text": "We already know that because we already know that the AI that we have today, that's roughly based on neural networks, they're called system one architectures from psychology.", "speaker": "Omar Hashash"}, {"start": 654.836, "end": 657.501, "text": "They're called system one architectures because they're reactive.", "speaker": "Omar Hashash"}, {"start": 657.937, "end": 677.177, "text": " uh but we know that we're missing these system two type of architectures so we know already that the system one it composes around roughly 95 of how we take actions in the world and this is what we've based ai systems on since the beginning and now", "speaker": "Omar Hashash"}, {"start": 677.343, "end": 689.04, "text": " What we're trying to do is to look at these other system two rational type of thinking ways and trying to see how can these come up with a new architecture in order to solve the problems that we have.", "speaker": "Omar Hashash"}, {"start": 690.083, "end": 714.37, "text": " so how does this architecture look like it looks something roughly like this it looks something like these cognitive modules that we have that i talked about the perception module and planning the world model all connected together uh of course over a network and the network acts here as a bridge uh or additional computing resource for the ages that exist in the world", "speaker": "Omar Hashash"}, {"start": 714.856, "end": 727.734, "text": " so that it benefits itself for the decisions, intelligent decisions that it can take, just like the beamforming example I was giving at the beginning, and all the other intelligent actions that it can help the agents do in the world.", "speaker": "Omar Hashash"}, {"start": 728.455, "end": 732.221, "text": "So this was based on one of our previous works.", "speaker": "Omar Hashash"}, {"start": 733.042, "end": 736.527, "text": "And here, I'll shout out to one of our colleagues", "speaker": "Omar Hashash"}, {"start": 736.507, "end": 745.302, "text": " which mentioned our presentation in one of the theoretical neurobiology groups before and", "speaker": "Omar Hashash"}, {"start": 745.805, "end": 747.327, "text": " gave us credit for this nice work.", "speaker": "Omar Hashash"}, {"start": 747.988, "end": 751.852, "text": "So this is actually us, yours truly over here with Crystal.", "speaker": "Omar Hashash"}, {"start": 752.853, "end": 771.256, "text": "And what he was presenting is that our work is kind of an intersection with other leading works in the field, like, for example, Yann LeCun's vision of modular cognitive brain architecture and his famous model of the world.", "speaker": "Omar Hashash"}, {"start": 772.197, "end": 774.9, "text": "But what we did over here is that we actually", "speaker": "Omar Hashash"}, {"start": 775.403, "end": 777.348, "text": " We push things a little bit forward.", "speaker": "Omar Hashash"}, {"start": 777.949, "end": 785.346, "text": "So what we kind of did is that we saw the other intersection with what Carfis is actually also presenting with active inference and the free energy principle.", "speaker": "Omar Hashash"}, {"start": 786.088, "end": 793.986, "text": "And we tried to combine all of these things together into one kind of coherent story around world models that we're going to see here today.", "speaker": "Omar Hashash"}, {"start": 794.557, "end": 802.567, "text": " So going back to this architecture that I was talking about, the first step to actually use this architecture is to sense the world.", "speaker": "Omar Hashash"}, {"start": 802.888, "end": 806.553, "text": "So we already, in terms of wireless networks, already studied this before.", "speaker": "Omar Hashash"}, {"start": 807.734, "end": 816.405, "text": "We have the abilities now in next generation networks to develop sensing capabilities that allow us to divide this world over wireless networks.", "speaker": "Omar Hashash"}, {"start": 816.986, "end": 818.688, "text": "We said this before in previous works.", "speaker": "Omar Hashash"}, {"start": 819.369, "end": 822.994, "text": "And I'll try to highlight here a little bit on the fact that we have these little", "speaker": "Omar Hashash"}, {"start": 822.974, "end": 843.784, "text": " red dots that i have over here in red which are called digital twins so digital twins are the copies or the replicas or the different models of the ai agents of these physical ai agents so they're the models of the agents the physical agents that exist in the world and we have the world models", "speaker": "Omar Hashash"}, {"start": 844.068, "end": 847.213, "text": " that are basically the other things around it.", "speaker": "Omar Hashash"}, {"start": 847.394, "end": 852.923, "text": "And digital twins are a part of the world model itself, along with the different elements that exist in the world.", "speaker": "Omar Hashash"}, {"start": 854.085, "end": 859.414, "text": "So basically, this is just the high-level concept how digital twins work.", "speaker": "Omar Hashash"}, {"start": 859.854, "end": 864.061, "text": "So they're based on the fact that they're a part of the world model.", "speaker": "Omar Hashash"}, {"start": 865.103, "end": 866.405, "text": "They intersect with world models.", "speaker": "Omar Hashash"}, {"start": 867.126, "end": 869.35, "text": "Both of them, they allow", "speaker": "Omar Hashash"}, {"start": 870.19, "end": 875.258, "text": " Basically, with a digital twin, you can actually move that work model forward.", "speaker": "Omar Hashash"}, {"start": 875.619, "end": 880.307, "text": "And this intersects with the notion of AI, which is called action-conditioned work models.", "speaker": "Omar Hashash"}, {"start": 881.088, "end": 884.173, "text": "So digital twins allow you to move forward in time.", "speaker": "Omar Hashash"}, {"start": 884.814, "end": 891.405, "text": "At the same time, they allow you to send some configuration back to what we call the physical twin in the physical world.", "speaker": "Omar Hashash"}, {"start": 891.723, "end": 895.347, "text": " So these configurations, we still don't know what they are, basically, in 6G.", "speaker": "Omar Hashash"}, {"start": 896.028, "end": 898.171, "text": "And this is actually what we'll be uncovering in this work.", "speaker": "Omar Hashash"}, {"start": 898.732, "end": 910.646, "text": "So basically, as I was saying, digital twins, they take real-time updates from the world, and they return some real-time optimization feedback back to the agents in the world.", "speaker": "Omar Hashash"}, {"start": 911.147, "end": 917.815, "text": "And their role is roughly somewhere around what-if analysis and doing predictions and monitoring for the AI agent.", "speaker": "Omar Hashash"}, {"start": 918.402, "end": 926.399, "text": " And we've also shown in one of our previous work before that these agents can have continual learning capabilities, but we're going to show it more in depth over here today.", "speaker": "Omar Hashash"}, {"start": 927.321, "end": 935.618, "text": "So roughly going back to that architecture that I was talking about, now we can look at that part where we talk about the agent, the digital twin.", "speaker": "Omar Hashash"}, {"start": 936.526, "end": 955.041, "text": " of the agent as going forward in time as a process of reasoning, because we're seeing it from a different lens, not from a 6G wireless communications lens, but probably from a neuroscience or AI point of view, where reasoning is actually trying to optimize some certain action for the agent in the world.", "speaker": "Omar Hashash"}, {"start": 955.443, "end": 960.032, "text": " But at the same time, what we want to do over here is similar to what LLMs do.", "speaker": "Omar Hashash"}, {"start": 960.193, "end": 967.568, "text": "So we want to reason in order to take, just like LLMs reason to give you a better answer for some harder prompt, for example.", "speaker": "Omar Hashash"}, {"start": 968.169, "end": 975.223, "text": "What we want to do is that we actually want to use this reasoning capability that's now on the network in order", "speaker": "Omar Hashash"}, {"start": 975.44, "end": 979.008, "text": " to allow the agent to reason to take a better action in the world.", "speaker": "Omar Hashash"}, {"start": 979.549, "end": 991.677, "text": "But that will require something like test time scaling, the one that was shown empirically in large language models, but is basically missing today in these physical AI systems.", "speaker": "Omar Hashash"}, {"start": 991.657, "end": 1020.162, "text": " so this is basically the goal of our work today we're going to present how these how this reasoning over the network can benefit the ai agents so that it reasons to make to take a better action and generalize in new unforeseen scenarios that it hasn't been trained on and to see how this actually results once built from first principles like active inference it results in a test time scaling law similar to the test time scanning law that we saw with large language models", "speaker": "Omar Hashash"}, {"start": 1022.2, "end": 1025.307, "text": " So this is our work that I'll be presenting today.", "speaker": "Omar Hashash"}, {"start": 1025.368, "end": 1034.269, "text": "And it has been recently released on archive for people that want to read more in depth about it.", "speaker": "Omar Hashash"}, {"start": 1035.008, "end": 1038.514, "text": " So let's start with test time scaling in the real world.", "speaker": "Omar Hashash"}, {"start": 1039.215, "end": 1043.022, "text": "So before we start, we have to look at the story from the beginning.", "speaker": "Omar Hashash"}, {"start": 1043.502, "end": 1047.269, "text": "So what we were doing before with our L agents, for example, is that we used to train them.", "speaker": "Omar Hashash"}, {"start": 1048.11, "end": 1056.865, "text": "And then after a certain time, we would reach a, for example, we're talking, of course, you were talking, for example, about some example of autonomous driving.", "speaker": "Omar Hashash"}, {"start": 1057.604, "end": 1078.228, "text": " what we used to do is that we used to train these autonomous agents and then we used to plug them in the world after some convergence of policy we just to plug them in the world and just we just hope that things work out perfectly fine but that's not always the case and ai fails once the world changes and of course we have a lot of examples about that", "speaker": "Omar Hashash"}, {"start": 1078.563, "end": 1095.458, "text": " so before before we go to solve the problem what we have to do is to actually go back to the origin so what are we actually doing over there how does the brain actually does this type of learning and to see what we are actually missing from the story", "speaker": "Omar Hashash"}, {"start": 1095.843, "end": 1099.968, "text": " So to start first, we have to acknowledge that we have two regions of the brain that are working together.", "speaker": "Omar Hashash"}, {"start": 1101.47, "end": 1109.179, "text": "Basically, we have the prefrontal cortex where all the learning, reasoning, planning, the world model, everything happens over there.", "speaker": "Omar Hashash"}, {"start": 1110.1, "end": 1118.61, "text": "And then after that, once we reach that convergence that I was talking about over here, what we do is that we take the policy and plug it back in the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1119.957, "end": 1129.11, "text": " And after that, roughly what we do is that we really don't need all the effort from the prefrontal cortex anymore because that's the smart part of our brain.", "speaker": "Omar Hashash"}, {"start": 1129.751, "end": 1134.558, "text": "We don't need to devote any big effort for that task.", "speaker": "Omar Hashash"}, {"start": 1134.818, "end": 1137.723, "text": "And it becomes somehow a subconscious task at that specific moment.", "speaker": "Omar Hashash"}, {"start": 1138.384, "end": 1141.348, "text": "And we start to rely more on our policy while working.", "speaker": "Omar Hashash"}, {"start": 1141.368, "end": 1142.91, "text": "And this is actually what we did before.", "speaker": "Omar Hashash"}, {"start": 1143.92, "end": 1151.331, "text": " But the question is now, if I were to ask you, how much do you think about driving while driving?", "speaker": "Omar Hashash"}, {"start": 1152.307, "end": 1155.872, "text": " the answer would probably be around zero.", "speaker": "Omar Hashash"}, {"start": 1156.273, "end": 1157.875, "text": "You don't think about driving while driving.", "speaker": "Omar Hashash"}, {"start": 1158.496, "end": 1170.433, "text": "And this is something we know, and this is something agrees with what the famous psychologist, Daniel Kahneman, was a Nobel Prize winner, actually identified before.", "speaker": "Omar Hashash"}, {"start": 1170.854, "end": 1178.365, "text": "So he roughly said that we use our intuition system one around 95% of the time in order to take actions.", "speaker": "Omar Hashash"}, {"start": 1178.345, "end": 1194.295, "text": " but at the same time there's this five percent that we also take actions with so when does this five percent pop up actually and why is it just five percent so one example of how we can use that is through surprise", "speaker": "Omar Hashash"}, {"start": 1194.562, "end": 1207.816, "text": " So for example, if you're here driving a vehicle, and let's say you're on a highway, and then you get a jaywalking pedestrian, for example, that suddenly wants to cross the road, you weren't expecting it.", "speaker": "Omar Hashash"}, {"start": 1208.377, "end": 1212.481, "text": "And actually, after that, you got a surprise that you didn't expect, right?", "speaker": "Omar Hashash"}, {"start": 1212.581, "end": 1216.806, "text": "Because if you didn't expect something, you wouldn't get a surprise at the beginning.", "speaker": "Omar Hashash"}, {"start": 1218.627, "end": 1222.932, "text": "And once that happens, what you do in that case is that you start to think.", "speaker": "Omar Hashash"}, {"start": 1223.468, "end": 1228.56, "text": " So if I move forward, I probably hit the person, but I don't want that to happen.", "speaker": "Omar Hashash"}, {"start": 1229.643, "end": 1240.65, "text": "So this is what you'd likely be saying, or probably saying that if I possibly go slower, the person behind me will probably bump into me and I'll head into a crash, another crash.", "speaker": "Omar Hashash"}, {"start": 1240.917, "end": 1250.809, "text": " So probably what I need to do is probably just lower my speed just a little bit to allow the person to move in front of me, cross whatever they want to go.", "speaker": "Omar Hashash"}, {"start": 1251.59, "end": 1254.314, "text": "And then after that, I can just continue my path.", "speaker": "Omar Hashash"}, {"start": 1254.914, "end": 1263.565, "text": "So this type of thinking over here is exactly necessary in order to pass that certain situation that I wasn't really trained on.", "speaker": "Omar Hashash"}, {"start": 1263.605, "end": 1265.187, "text": "I didn't really expect in front of me.", "speaker": "Omar Hashash"}, {"start": 1266.01, "end": 1272.919, "text": " So roughly, if I want to combine it now, so this is kind of how we do this type of decision making.", "speaker": "Omar Hashash"}, {"start": 1273.5, "end": 1279.929, "text": "So basically, when you're driving, if there's no prediction error in front of you, you did not expect anything surprising.", "speaker": "Omar Hashash"}, {"start": 1279.989, "end": 1282.812, "text": "Everything in the world is working precisely as you would want it to do.", "speaker": "Omar Hashash"}, {"start": 1284.815, "end": 1289.521, "text": "You basically use your Bayes of Ganglia and basically the policy that exists in your Bayes of Ganglia.", "speaker": "Omar Hashash"}, {"start": 1289.902, "end": 1292.946, "text": "But at the same time, if you do have a prediction error,", "speaker": "Omar Hashash"}, {"start": 1292.926, "end": 1310.513, "text": " this means that your policy is somehow sub-optimal and you need to go through your frontal cortex again you need to kick it in again until it handle this new situation for you so what you would do in that case is that you would reason before you take the right action at that specific moment", "speaker": "Omar Hashash"}, {"start": 1311.067, "end": 1314.841, "text": " So now I probably ask you, where did we see the system before?", "speaker": "Omar Hashash"}, {"start": 1314.901, "end": 1318.735, "text": "This kind of system that toggles between reasoning and not reasoning.", "speaker": "Omar Hashash"}, {"start": 1318.795, "end": 1323.051, "text": "This is exactly what we've seen in", "speaker": "Omar Hashash"}, {"start": 1323.993, "end": 1326.696, "text": " models and large language models like GPT-5.", "speaker": "Omar Hashash"}, {"start": 1327.117, "end": 1328.318, "text": "So that's smart routing.", "speaker": "Omar Hashash"}, {"start": 1329.2, "end": 1332.444, "text": "So this is, you could say probably this is from where it came from.", "speaker": "Omar Hashash"}, {"start": 1332.984, "end": 1346.742, "text": "I'm going to see how that connects afterwards, but basically what the agents, the soft kind of agents, digital agents, language agents that we see today, when they reason or don't reason, they follow that specific scheme that is in the brain.", "speaker": "Omar Hashash"}, {"start": 1347.683, "end": 1351.688, "text": "So basically the prediction error is the, um,", "speaker": "Omar Hashash"}, {"start": 1352.697, "end": 1356.365, "text": " is the natural trigger, rather than having a holistic router.", "speaker": "Omar Hashash"}, {"start": 1356.405, "end": 1361.957, "text": "The router is trying to approximate that trigger, natural trigger.", "speaker": "Omar Hashash"}, {"start": 1362.077, "end": 1367.85, "text": "But the question here becomes, how do we actually deal with the scaling at that specific moment?", "speaker": "Omar Hashash"}, {"start": 1368.01, "end": 1371.077, "text": "I said that we thought and took an action.", "speaker": "Omar Hashash"}, {"start": 1371.513, "end": 1373.659, "text": " how is the action actually resulting over here?", "speaker": "Omar Hashash"}, {"start": 1373.679, "end": 1375.303, "text": "How is it actually coming into place?", "speaker": "Omar Hashash"}, {"start": 1375.403, "end": 1376.245, "text": "We still don't know.", "speaker": "Omar Hashash"}, {"start": 1376.306, "end": 1382.221, "text": "And that's something still mysterious in the AI space.", "speaker": "Omar Hashash"}, {"start": 1382.562, "end": 1385.269, "text": "But you see here what I presented just now in terms of that", "speaker": "Omar Hashash"}, {"start": 1385.772, "end": 1389.076, "text": " prediction error and trying to reason or to move that prediction error.", "speaker": "Omar Hashash"}, {"start": 1389.597, "end": 1391.539, "text": "This is exactly what active inference is.", "speaker": "Omar Hashash"}, {"start": 1391.9, "end": 1398.708, "text": "So active inference is a first principle that describes how all living systems survive in the world.", "speaker": "Omar Hashash"}, {"start": 1399.349, "end": 1406.358, "text": "And returning back to the example that I presented in the beginning, that humanoid robot and that people that is crashing.", "speaker": "Omar Hashash"}, {"start": 1406.698, "end": 1410.463, "text": "So this is how that in other ways that", "speaker": "Omar Hashash"}, {"start": 1410.848, "end": 1421.334, "text": " the absence of this minimum amount of freezing on top of the AIs that we have today can actually explain why those systems are dying systems in the world.", "speaker": "Omar Hashash"}, {"start": 1421.976, "end": 1428.452, "text": "If this is how all systems survive in the world, all living systems like us, humans or animals,", "speaker": "Omar Hashash"}, {"start": 1428.432, "end": 1441.952, "text": " survival in the world, and the fact that this is absent in those autonomous agents, autonomous physical AI agents that exist in the world, you can possibly find an explanation to the fact why these systems are dying.", "speaker": "Omar Hashash"}, {"start": 1443.89, "end": 1458.986, "text": " So of course, uh, in contrast to what, uh, to what we have, uh, we can see here that what we need to do actually is to allow those systems to survive at this time, because as I said, it's the first principle of that explains how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1459.806, "end": 1467.715, "text": "So survival at that specific moment over here, when we reached the conversions, so we were reaching something in non equilibrium steady state in terms of physics.", "speaker": "Omar Hashash"}, {"start": 1468.495, "end": 1471.078, "text": "So from a genetic point of view, you need to maintain", "speaker": "Omar Hashash"}, {"start": 1471.514, "end": 1478.004, "text": " non-equilibrium steady state by preserving the policy and the world model so that you can survive in the world.", "speaker": "Omar Hashash"}, {"start": 1479.366, "end": 1482.01, "text": "So preserving that actually has two routes.", "speaker": "Omar Hashash"}, {"start": 1482.31, "end": 1488.159, "text": "The first route is what existing works do, which is that they assume stationary conditions about the environment.", "speaker": "Omar Hashash"}, {"start": 1489.261, "end": 1493.627, "text": "So in that case, the system works just in the states that you actually saw on the training set.", "speaker": "Omar Hashash"}, {"start": 1494.549, "end": 1497.753, "text": "And survival is limited to just those conditions.", "speaker": "Omar Hashash"}, {"start": 1499.015, "end": 1499.496, "text": "But", "speaker": "Omar Hashash"}, {"start": 1499.577, "end": 1506.792, "text": " There's an alternative route for the system to survive, which is by integrating the first principles of survival from the beginning.", "speaker": "Omar Hashash"}, {"start": 1507.293, "end": 1510.379, "text": "So you don't need to actually make your world frozen at this time.", "speaker": "Omar Hashash"}, {"start": 1510.64, "end": 1512.844, "text": "You can keep it dynamic and non-stationary.", "speaker": "Omar Hashash"}, {"start": 1513.195, "end": 1519.923, "text": " but give the agent the first principle of survival so that it survives in dynamic, non-stationary worlds.", "speaker": "Omar Hashash"}, {"start": 1520.124, "end": 1526.031, "text": "So in that case, survival is no longer limited just to what the agent saw in the training set.", "speaker": "Omar Hashash"}, {"start": 1526.412, "end": 1531.959, "text": "Now it actually survives even in new situations, and it allows it to generalize, which was our main goal from the beginning.", "speaker": "Omar Hashash"}, {"start": 1532.319, "end": 1536.364, "text": "What we want is an agent that generalizes in unforeseen scenarios that we haven't seen before.", "speaker": "Omar Hashash"}, {"start": 1536.965, "end": 1539.388, "text": "And in that case, by definition,", "speaker": "Omar Hashash"}, {"start": 1539.655, "end": 1547.154, "text": " Surviving in the world is the main facet of a system, of an AI agent, basically, that generalizes.", "speaker": "Omar Hashash"}, {"start": 1548.478, "end": 1552.568, "text": "So by guaranteeing survival, you're actually guaranteeing that the agent generalizes.", "speaker": "Omar Hashash"}, {"start": 1554.303, "end": 1563.952, "text": " So, roughly speaking, this is what active inferences and what Carl Friston presented it as a way to remove prediction errors from the brain.", "speaker": "Omar Hashash"}, {"start": 1563.992, "end": 1566.234, "text": "Of course, we always have these sensory information.", "speaker": "Omar Hashash"}, {"start": 1567.635, "end": 1574.722, "text": "We're seeing if we don't have a prediction error, it means that we're in the region that minimizes free energy or basically surprised and we don't need to take any action.", "speaker": "Omar Hashash"}, {"start": 1575.122, "end": 1581.188, "text": "However, if we do have a prediction error, we will try to act on the world so that we minimize that prediction error.", "speaker": "Omar Hashash"}, {"start": 1583.007, "end": 1612.525, "text": " uh so of course in this way we can see that reasoning to reduce surprise is the minimum amount of reasoning in any in any living system to survive in the world that's the minimum amount of intelligence that you need to have in order to exist in the world over a long time so what we kind of did over here is that we tried to integrate this into what we know in terms of agents that we just don't need them to just survive we want to make them usable so that they can actually perform tasks", "speaker": "Omar Hashash"}, {"start": 1612.91, "end": 1618.963, "text": " So we went back to integrate this with what we had in terms of reinforcement learning and the policies that we had.", "speaker": "Omar Hashash"}, {"start": 1619.785, "end": 1631.29, "text": "So what happens in this case is that what we saw is that if we take in states and if those states are somehow unforeseen, it means that we can go reason", "speaker": "Omar Hashash"}, {"start": 1633.16, "end": 1637.007, "text": " in order to remove the prediction error that results from this unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 1637.688, "end": 1651.995, "text": "And at the same time, if we can use that test time scaling, we can use it to alter the actions to fit the specific scenario in order to make it optimal in that specific scenario so that the agent can actually generalize in that specific moment.", "speaker": "Omar Hashash"}, {"start": 1652.38, "end": 1666.356, "text": " So roughly speaking, if you want to put it down graphically over here in the figure, what we were doing with reinforcement learning is that we were looking at those states that are highly probable and trying to focus on those states.", "speaker": "Omar Hashash"}, {"start": 1667.602, "end": 1669.824, "text": " finding a policy that fits those states.", "speaker": "Omar Hashash"}, {"start": 1670.105, "end": 1672.187, "text": "This is basically what reinforcement learning is doing.", "speaker": "Omar Hashash"}, {"start": 1673.148, "end": 1676.792, "text": "And after that, what we did is that we froze the world.", "speaker": "Omar Hashash"}, {"start": 1676.812, "end": 1678.374, "text": "We said that the world is not going to change.", "speaker": "Omar Hashash"}, {"start": 1678.954, "end": 1682.278, "text": "So we're always going to be stuck in those states over there and those policies.", "speaker": "Omar Hashash"}, {"start": 1682.338, "end": 1684.581, "text": "But we know that that fades in the real world.", "speaker": "Omar Hashash"}, {"start": 1685.061, "end": 1691.348, "text": "What's actually happening in the real world is that because it's non-stationary, it will probably push you to a state", "speaker": "Omar Hashash"}, {"start": 1691.328, "end": 1694.413, "text": " That is not a problem for you.", "speaker": "Omar Hashash"}, {"start": 1694.573, "end": 1700.343, "text": "It's not a part of the set that you've seen in your training distribution.", "speaker": "Omar Hashash"}, {"start": 1701.285, "end": 1714.006, "text": "What you need to do with that type of reasoning to minimize prediction errors is actually what pushes you back or scales you back to your characteristic set, which is the set of states where you can use your policy again.", "speaker": "Omar Hashash"}, {"start": 1714.467, "end": 1717.352, "text": "So this is basically what we mean by test time scaling.", "speaker": "Omar Hashash"}, {"start": 1718.665, "end": 1734.702, "text": " In this case, we can see, and this is a very nice viewpoint on the story, which is that every living system, in this case, shows a general objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1736.324, "end": 1742.59, "text": "And specifically, I'm saying here something general, because this is something that all living systems perform in the world.", "speaker": "Omar Hashash"}, {"start": 1743.191, "end": 1747.255, "text": "So the fact that you are a living system, it means that you have an objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1748.416, "end": 1768.348, "text": " or else you would die so in that case we are governed by general objectives like survival over here and at the same time we have narrow objectives which allow us to achieve our specific tasks and goals with subsumed within that specific region of survival that we have", "speaker": "Omar Hashash"}, {"start": 1768.953, "end": 1775.687, "text": " So I'll just give here an example, just like the jaywalking example of here, just to say how this system kind of works.", "speaker": "Omar Hashash"}, {"start": 1776.509, "end": 1782.06, "text": "So basically, what we have is the fact that I have an autonomous driving vehicle, for example.", "speaker": "Omar Hashash"}, {"start": 1782.141, "end": 1783.283, "text": "It's driving.", "speaker": "Omar Hashash"}, {"start": 1783.482, "end": 1785.004, "text": " And everything's fine.", "speaker": "Omar Hashash"}, {"start": 1785.104, "end": 1786.365, "text": "I have no prediction error.", "speaker": "Omar Hashash"}, {"start": 1786.746, "end": 1788.988, "text": "This is something that we call passive influence in that case.", "speaker": "Omar Hashash"}, {"start": 1789.108, "end": 1791.051, "text": "So we're just taking in sensory observations.", "speaker": "Omar Hashash"}, {"start": 1791.091, "end": 1792.953, "text": "Everything is working as expected.", "speaker": "Omar Hashash"}, {"start": 1793.433, "end": 1794.194, "text": "I have no problem.", "speaker": "Omar Hashash"}, {"start": 1794.935, "end": 1799.801, "text": "And then suddenly what I have is that an agent that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1800.241, "end": 1801.743, "text": "Sorry, a human that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1803.144, "end": 1812.595, "text": "In that case, the network, of course, the network is acting here as an additional computing resource in addition to the agent that we have.", "speaker": "Omar Hashash"}, {"start": 1813.3, "end": 1815.845, "text": " you don't need to focus a lot on what the network is doing.", "speaker": "Omar Hashash"}, {"start": 1815.865, "end": 1818.269, "text": "The network is similar to the agent in that case.", "speaker": "Omar Hashash"}, {"start": 1818.75, "end": 1820.894, "text": "They're kind of one system together.", "speaker": "Omar Hashash"}, {"start": 1822.396, "end": 1830.972, "text": "So in this case, the network that we have will detect that there's a prediction error, and it will switch from passive to active inference about the world.", "speaker": "Omar Hashash"}, {"start": 1831.508, "end": 1836.993, "text": " So what that means is that the network will engage in reasoning in order to minimize the surprise.", "speaker": "Omar Hashash"}, {"start": 1837.554, "end": 1846.143, "text": "And it's going to use those digital twins that I talked about at the beginning, which capture the model of the agent and the work model itself in order to reason.", "speaker": "Omar Hashash"}, {"start": 1846.543, "end": 1854.511, "text": "And what we want to do is that we want to reason, just like I was talking about, I was giving an example here, should I move forward faster?", "speaker": "Omar Hashash"}, {"start": 1854.531, "end": 1855.492, "text": "Should I move slower?", "speaker": "Omar Hashash"}, {"start": 1855.552, "end": 1861.498, "text": "So all that thinking together, we're going to use it in order to put down a value function.", "speaker": "Omar Hashash"}, {"start": 1862.255, "end": 1873.032, "text": " So the value function here, we call it as a delta v. So basically, I'm trying to measure which action allows me to resolve my prediction error in that specific state that I am in.", "speaker": "Omar Hashash"}, {"start": 1873.052, "end": 1881.165, "text": "And then what I want to do is that I'm going to take in that feedback from the network and send it back to the physical AI agent.", "speaker": "Omar Hashash"}, {"start": 1881.566, "end": 1882.828, "text": "The physical AI agent", "speaker": "Omar Hashash"}, {"start": 1883.804, "end": 1886.367, "text": " What it's going to do is that it's going to use this feedback.", "speaker": "Omar Hashash"}, {"start": 1886.888, "end": 1891.374, "text": "It's like the prefrontal cortex sending the signals down to the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1891.875, "end": 1895.84, "text": "So the policy is over here at the level of the agent of the vehicle here.", "speaker": "Omar Hashash"}, {"start": 1896.301, "end": 1899.986, "text": "And it's going to use it in order to reason and do inference.", "speaker": "Omar Hashash"}, {"start": 1900.006, "end": 1904.412, "text": "So it's going to scale its policy from pi node to pi node prime, for example, here.", "speaker": "Omar Hashash"}, {"start": 1904.712, "end": 1909.0, "text": " So in that specific case, the action that it's going to take is going to be different than its normal.", "speaker": "Omar Hashash"}, {"start": 1909.44, "end": 1916.193, "text": "It's going to be possibly to decrease its velocity in order to avoid an accident, for instance.", "speaker": "Omar Hashash"}, {"start": 1916.213, "end": 1927.192, "text": "And then what happens is that after this jaywalking pedestrian just passes, I go back to my regular cases where I have my prediction error resolved and the pedestrian passes.", "speaker": "Omar Hashash"}, {"start": 1927.232, "end": 1929.336, "text": "So I can go back just to my passive inventory.", "speaker": "Omar Hashash"}, {"start": 1930.227, "end": 1943.289, "text": " So what you can see here is that our world is basically a transition always between foreseen scenarios, unforeseen scenarios, resolving these unforeseen scenarios, and then going back to foreseen scenarios.", "speaker": "Omar Hashash"}, {"start": 1943.57, "end": 1945.152, "text": "This is how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1945.753, "end": 1948.438, "text": "We're always transitioning between those different states.", "speaker": "Omar Hashash"}, {"start": 1949.937, "end": 1953.542, "text": " So now I'm going to go into the system model.", "speaker": "Omar Hashash"}, {"start": 1954.222, "end": 1957.887, "text": "So as I said before, we have an agent that is in the world.", "speaker": "Omar Hashash"}, {"start": 1958.768, "end": 1959.97, "text": "It has a policy panel.", "speaker": "Omar Hashash"}, {"start": 1960.891, "end": 1964.275, "text": "And we have a world model that exists over the network in this architecture.", "speaker": "Omar Hashash"}, {"start": 1965.136, "end": 1971.945, "text": "One part of the world model is composed from these digital tools, these different alternative policies that the agent can have.", "speaker": "Omar Hashash"}, {"start": 1971.925, "end": 1973.948, "text": " or models of the agent.", "speaker": "Omar Hashash"}, {"start": 1974.749, "end": 1984.383, "text": "At the same time, we have other things that are called assets which capture the different things that exist in the world, just like the human, the houses, everything else besides the agents themselves.", "speaker": "Omar Hashash"}, {"start": 1985.344, "end": 1988.269, "text": "And what's going to happen is that we're going to take sensing observations.", "speaker": "Omar Hashash"}, {"start": 1988.849, "end": 1993.877, "text": "So it's not the agent itself that's doing the sensing, it's actually the network that's doing the sensing on behalf of the agent.", "speaker": "Omar Hashash"}, {"start": 1994.978, "end": 1997.282, "text": "And what we're going to do is that we're going to do perception.", "speaker": "Omar Hashash"}, {"start": 1997.422, "end": 2000.987, "text": "So we're basically going to map our observations to", "speaker": "Omar Hashash"}, {"start": 2001.49, "end": 2009.058, "text": " the set of states that we have trying to make sense of the states, trying to infer what's going to happen over here in terms of the states.", "speaker": "Omar Hashash"}, {"start": 2009.538, "end": 2018.588, "text": "And then once we have a surprise, so we have an unforeseen scenario, what's going to happen is that the network is going to engage in counterfactual reasoning to minimize surprise.", "speaker": "Omar Hashash"}, {"start": 2019.289, "end": 2025.575, "text": "So it's going to think about all these different alternative routes with the different policies that they may have over here, different alternative policies.", "speaker": "Omar Hashash"}, {"start": 2026.496, "end": 2031.261, "text": "And then, as I said, it's going to send back these digital twin configurations.", "speaker": "Omar Hashash"}, {"start": 2032.405, "end": 2038.804, "text": " back to the agent in order to scale the policy and generalize in the unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2039.577, "end": 2043.742, "text": " So mathematically, we can look at perception, of course, as an inference process.", "speaker": "Omar Hashash"}, {"start": 2043.802, "end": 2051.032, "text": "But to do that, we have to first define our generative world model in terms of the states and observations and the policies that we have.", "speaker": "Omar Hashash"}, {"start": 2052.494, "end": 2060.764, "text": "Of course, before we do perception, we have to predict what the expected observation is going to be.", "speaker": "Omar Hashash"}, {"start": 2061.465, "end": 2064.048, "text": "And then we're going to get our sensing observations from the world.", "speaker": "Omar Hashash"}, {"start": 2064.509, "end": 2066.952, "text": "And then we're going to measure if there's a surprise or not.", "speaker": "Omar Hashash"}, {"start": 2066.932, "end": 2070.28, "text": " If we do have a surprise, of course, this is how we measure the surprise.", "speaker": "Omar Hashash"}, {"start": 2070.32, "end": 2071.644, "text": "It's just the easiest surprise.", "speaker": "Omar Hashash"}, {"start": 2073.127, "end": 2082.19, "text": "And after that, if we do have a surprise, it means that we're going to cross some threshold epsilon in order to signal that there's an unforeseen scenario in front of us.", "speaker": "Omar Hashash"}, {"start": 2082.17, "end": 2105.005, "text": " so the first thing that should happen in that case is that we should actually calculate the posterior because basically what we predicted about the world is not right we have an error so we have the first thing we have to do is to make sense of our world again by calculating this posterior after we saw our observations which is basically an inference process of it just we can see before we engage in plan and reason", "speaker": "Omar Hashash"}, {"start": 2104.985, "end": 2111.657, "text": " So, the planning part, it constitutes these digital twins and work models that engage in counterfactual reasoning.", "speaker": "Omar Hashash"}, {"start": 2111.697, "end": 2117.447, "text": "So, basically, the what-if analysis that I was giving an example about, in order to plan the plausible future world states.", "speaker": "Omar Hashash"}, {"start": 2117.647, "end": 2122.656, "text": "And the goal, of course, of this reasoning is to actually minimize future surprise.", "speaker": "Omar Hashash"}, {"start": 2122.957, "end": 2128.742, "text": " So our goal is to see how much surprise is going to result from each of these policies over here.", "speaker": "Omar Hashash"}, {"start": 2128.762, "end": 2135.889, "text": "And you can see that it's going to be nothing but a measure of the entropy at each of those states in the future.", "speaker": "Omar Hashash"}, {"start": 2135.969, "end": 2140.933, "text": "So from t plus 1 until the end of the time horizon that we have.", "speaker": "Omar Hashash"}, {"start": 2140.953, "end": 2149.2, "text": "And then after that, what we're going to do is that we're going to use this reasoning that we did in order to build a value function.", "speaker": "Omar Hashash"}, {"start": 2149.721, "end": 2150.922, "text": "We call it delta v.", "speaker": "Omar Hashash"}, {"start": 2152.252, "end": 2154.935, "text": " And where the reward is actually minimizing surprise.", "speaker": "Omar Hashash"}, {"start": 2154.955, "end": 2160.563, "text": "So as I said at the beginning, we have a general objective in the brain, which is to survive.", "speaker": "Omar Hashash"}, {"start": 2161.364, "end": 2164.348, "text": "So basically it has a value function over here.", "speaker": "Omar Hashash"}, {"start": 2164.368, "end": 2174.901, "text": "And then what we're going to do is that we're going to take this value and we're going to send it back to the agents in the world in order to scale the policy.", "speaker": "Omar Hashash"}, {"start": 2175.37, "end": 2196.602, "text": " how that looks like it's a little bit similar to what perception did so we have the perception as inference of course planning is also as inference over here because planning is an extension of minimizing surprises in the future so it's also planning as inference and now we have also we have to model action as inference but here it's a different type of conference", "speaker": "Omar Hashash"}, {"start": 2196.582, "end": 2212.228, "text": " the fact that when we did perception as influence the evidence that happened in front of us was 100 true it happened in front of us in the world so we're confident about it but when you want to do action as influence it's a little bit different", "speaker": "Omar Hashash"}, {"start": 2212.208, "end": 2217.479, "text": " So it's based on imaginary virtual data in our brain, right?", "speaker": "Omar Hashash"}, {"start": 2217.499, "end": 2220.204, "text": "When we were thinking, we were actually generating data.", "speaker": "Omar Hashash"}, {"start": 2220.445, "end": 2221.888, "text": "That thing didn't happen in the world.", "speaker": "Omar Hashash"}, {"start": 2222.108, "end": 2223.471, "text": "It happened inside our brain.", "speaker": "Omar Hashash"}, {"start": 2224.152, "end": 2230.926, "text": "But luckily, we have Judea Pearl, who came up with this method of virtual evidence.", "speaker": "Omar Hashash"}, {"start": 2231.446, "end": 2236.833, "text": " So he has a nice technique in order to update the beliefs in that case.", "speaker": "Omar Hashash"}, {"start": 2237.474, "end": 2242.48, "text": "So we can see here, if we write it down, this is going to be something called a soft Bayesian update.", "speaker": "Omar Hashash"}, {"start": 2243.161, "end": 2249.629, "text": "So it's soft because you don't really need to do a complete inference update over here.", "speaker": "Omar Hashash"}, {"start": 2250.27, "end": 2259.462, "text": "So if you write it down, you can see here at the last line, you can write it in terms of a feed forward and reasoning or inference term.", "speaker": "Omar Hashash"}, {"start": 2259.83, "end": 2260.891, "text": " which is something very nice.", "speaker": "Omar Hashash"}, {"start": 2261.211, "end": 2275.066, "text": "Because if you really look at it, if you put that theta, which is basically your normalized surprise, if the world is mostly predictable, so your theta or surprise is basically equal to 0.", "speaker": "Omar Hashash"}, {"start": 2275.647, "end": 2284.116, "text": "And if you put that exponential part over here equal to 0, you basically result in the feedforward policy that we had in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2285.142, "end": 2289.269, "text": " The fact is that action is always an inference process.", "speaker": "Omar Hashash"}, {"start": 2289.569, "end": 2293.195, "text": "It's always modulated by that exponential term that we have over here.", "speaker": "Omar Hashash"}, {"start": 2293.676, "end": 2299.746, "text": "And this is what scales the policy once we have prediction errors and once we have an unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2300.427, "end": 2311.065, "text": "So the policies in today's solution, they remain limited to the case where surprise is null due to the stationary assumptions about the world, just like we do in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2311.045, "end": 2314.932, "text": " So reinforcement learning is not stationary by itself at this time.", "speaker": "Omar Hashash"}, {"start": 2315.032, "end": 2323.968, "text": "We made it stationary because we did not compensate for this factor over here in equation 5, which is clearly a special case of the equation.", "speaker": "Omar Hashash"}, {"start": 2325.03, "end": 2329.678, "text": "However, the inference at this part can be challenging if you want to really solve the problem.", "speaker": "Omar Hashash"}, {"start": 2330.03, "end": 2335.64, "text": " It can be challenging because marginalizing over all the possible states is computationally intractable in practice.", "speaker": "Omar Hashash"}, {"start": 2336.341, "end": 2341.29, "text": "So to find a better solution, we need to resort to a variational Bayesian inference solution.", "speaker": "Omar Hashash"}, {"start": 2341.53, "end": 2347.06, "text": "And I'll keep it here for Christo to present the solution.", "speaker": "Omar Hashash"}, {"start": 2349.625, "end": 2352.51, "text": "I think I'll stop sharing for Christo to share.", "speaker": "Daniel Friedman"}, {"start": 2373.854, "end": 2374.595, "text": " Thanks, Omar.", "speaker": "Daniel Friedman"}, {"start": 2376.958, "end": 2380.243, "text": "So I will go over the solution roadmap first.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2381.305, "end": 2387.193, "text": "As Omar mentioned, our solution is grounded in variational variation inference principle.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2388.815, "end": 2401.954, "text": "So this evolves as a four-step cognitive loop, starting with perception, which is about perceiving the current state by minimizing a variational free energy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2402.474, "end": 2420.297, "text": " and then uh you know we go to the future time steps where we you know estimate the expected free energy static from the inferred state and then uh you know infer policy for actions that minimize the future surprise so that's the planning phase", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2420.277, "end": 2437.392, "text": " and then in the action phase uh so there is we introduced the concept of marco blanket which acts as a statistical boundary between agent and the environment and then policy scaling will be more modeled as a gradient descent on the free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2437.372, "end": 2453.727, "text": " um and finally okay we incorporate this uh updated policy as well as uh the land posterior distributions about the states into you know updating the world model that so that again happens via vfe minimization so", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2453.707, "end": 2464.773, "text": " These four stages is grounded on the same or minimizing the same variation free energy concept.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2465.174, "end": 2467.199, "text": "That's about the solution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2467.86, "end": 2469.825, "text": "Then first, let's look at the perception.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2469.805, "end": 2479.079, "text": " So perception is modeled as inference, inferring the posterior distribution of the states S given the observations.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2479.8, "end": 2494.362, "text": "But if you expand it in terms of the Bayes rule, you can see that this is interactable, computationally interactable because of the marginal distribution of the observations which appear in the denominator here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2495.017, "end": 2509.87, "text": " so here we move to a to find a tractable posterior distribution we actually uh formulate the kl divergence between the approximate posterior q of s comma pi zero", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2509.917, "end": 2536.633, "text": " and the KL divergence to the actual or the true posterior so that's that's the concept of variational inference so where if you actually expand this KL divergence we can see that it can be written as a variation free energy times variation of free energy as well as an additional yeah plus or so basically variation free energy can be written as a KL divergence plus a surprise time which is ln p of time here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2536.985, "end": 2566.772, "text": " uh and so minimizing the scale divergence become equal to uh yeah minimizing variation and then okay once we actually once we actually find a posterior that minimizes the kl divergence along with that we all we are also computing or we are also estimating the surprise quantity which is here uh so that's the advantage of this variational variation inference formulation here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2567.528, "end": 2591.494, "text": " um yeah so yeah so once we find an optimal or optimal posterior distribution that actually means that uh the resetting variation free energy is equal to the surprise quantity uh so which surprise the estimation of the surprise is actually needed for the policy scaling that uh homer mentioned in the previous slide", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2593.162, "end": 2597.248, "text": " So now the question is how we actually go about solving this.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2597.608, "end": 2606.241, "text": "So this can be modeled as a sequential inference process, which can be formulated as a partially observed Markov decision process as shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2606.281, "end": 2610.587, "text": "And this leads to the following factor graph as is shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2611.048, "end": 2617.617, "text": "So where we can see that, okay, so we actually modeled the transition from the state's", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2617.597, "end": 2620.328, "text": " to the observations using the matrix A here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2620.389, "end": 2626.695, "text": "So we actually model all these transition probabilities as categorical distributions.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2626.827, "end": 2654.252, "text": " and we use the matrix b to actually model the transition from the previous state to the current state st uh yeah so now optimizing the or minimizing the variation of free energy this leads to an expression as is shown here uh so okay this basically leads to variational message passing uh wherein we can see that okay so there is the message one which is basically uh accounting for the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2654.232, "end": 2675.38, "text": " uh yeah basically a forward transition which basically depends upon the forward transition probability and message 2 which depends upon the backward transition probability which is the message 2 component here and then message 3 actually uh incorporates the uh observation likelihood distribution", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2676.136, "end": 2689.173, "text": " And further, in order to compute the posterior distribution, which is basically defined as the s variable here, we basically take a softmax over the logarithm tense here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2689.694, "end": 2695.461, "text": "So that's how we actually obtain the ln q of s or the posterior distribution here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2696.42, "end": 2721.668, "text": " and uh so this posterior distribution can be shown to be okay so as we mentioned we started with minimizing the variation free energy component so at converge at the equilibrium or yeah at the convergence this um optimal posterior distribution should um converge to that of the true posterior so it basically is minimizing the um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2722.475, "end": 2745.208, "text": " variational free energy component so basically it's a gradient descent on the um free energy um component so that's what basically this uh yeah perception component basically shows and we can actually show that this uh variational free energy can be written to be uh okay the posterior times an error on the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2745.188, "end": 2757.477, "text": " uh posterior posterior distribution so basically minimizing the variation of free energy is equivalent to minimizing the error on the estimated posterior or inferred posterior distributions", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2759.262, "end": 2779.414, "text": " and yeah so now uh when we go to planning so what we actually do is that uh so we look at future time steps and we we want to find the policy that minimizes the average free energy uh across time so that's basically so so for that we need expected free energy component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2779.394, "end": 2797.639, "text": " and we show that this expected free energy component can be nicely decomposed into two components so one is the risk component which actually uh measures the kl divergence between uh the you know approximate uh oh basically the inferred um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2798.615, "end": 2820.285, "text": " distribution of observations given the current policy how that so the inferred uh distribute inferred observational distribution how that is actually comparing to the uh expected or preferred observations or the preferences that we have uh for the agent so that's basically the risk component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2820.265, "end": 2845.429, "text": " plus there is an ambiguity component that uh so basically basically it's a entropy of the probability distributions which is basically the likelihood probability distribution here which so basically if you want to minimize the expected free energy you should make sure that the inferred world model is actually close to that or inferred observations are basically close to that of the preferred observations as well as", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2845.409, "end": 2855.234, "text": " we should try to reduce the ambiguity in terms of the observations that the agent observe in future time steps.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2855.314, "end": 2858.001, "text": "So that's basically the expected free energy component.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2859.297, "end": 2877.12, "text": " uh yeah so now uh the the prior over the posterior the the prior over the policies can be written as s of max of uh minus gamma g so we should yeah the policy should ensure that it should always go in the direction of minimizing the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2877.438, "end": 2897.472, "text": " and uh yeah inferred policy uh inferred action policy can be written as um yeah summation over all the policies pi uh where yeah corresponding to action a times the the prior distribution of the particular policy pi here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2897.452, "end": 2917.3, "text": " uh and we can yeah so that's that's the inferred policy and then finally as omar mentioned in the previous slides what we actually do is using this inferred policy so this inferred policy basically uh needs an estimate of the surprise so so basically inferred policy depends upon the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2917.28, "end": 2921.485, "text": " which basically is equivalent to the surprise quantity, as I mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2921.525, "end": 2930.417, "text": "So basically, you cannot compute an exact value for the inference policy, but you can only estimate using the surprise quantity.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2930.897, "end": 2940.79, "text": "So to estimate the surprise, we need to perform reasoning.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2940.81, "end": 2945.616, "text": "So how much reasoning that you need to do depends upon this parameter called theta here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2945.596, "end": 2952.41, "text": " So that is similar to that of the test time computing LLMs like Omar mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2952.43, "end": 2959.504, "text": "So that's how we perform the planning and do the test time scaling.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2961.273, "end": 2968.691, "text": " Yeah, so now the question is whether all these computations are really stable or not.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2969.292, "end": 2972.821, "text": "For that, we need to look at the concept of Markov blanket.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2973.021, "end": 2978.033, "text": "So an AI agent that generalizes is a random dynamical system", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2978.013, "end": 2997.574, "text": " that exist over time so for the a for this ai agent to be able to generalize across unforeseen scenarios so there should exist a marco blanket which basically means that uh so marco blanket is consisting of those sensors okay so the agent interacts with the environment through the sensory state", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2997.723, "end": 3019.933, "text": " and so so basically environment influences the agent through the sensory state and the agent influences the environment through the active state so this active so this marco blanket compose which is composed of the active states as well as the sensory state uh that is something which you know separates the internal agent state from the external states", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3020.369, "end": 3046.14, "text": " yeah so that's how this the closed loop of planning um perception planning action and sensing happens and the key equation here is that so we can write the states x the flow for the state x just composed of this marker blanket states as shown in as is shown in here which is composed of two temps so one is the one is a great dissipative gradient time and another is the solenoid", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3046.12, "end": 3053.953, "text": " uh time which basically both of this time depends upon the log of the model evidence which is ln p of x here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3054.018, "end": 3078.979, "text": " uh so yeah for the agent to exist over time so there should be a mod uh marco blanket which also means that it should actually maximize the this log p of x quantity or the model evidence quantity which is basically equal to minimizing a free and i mean minimizing uh my uh l minus ln p of x which is basically corresponding to uh free energy estimate", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3078.959, "end": 3101.592, "text": " uh yeah so that's what basically this um concept of marco blanket tells us and also this also gives us a principled criterion for principle stopping criterion which basically means that uh so when do when do the agent actually stop reasoning so that can be written like uh yeah that can be written as depending upon this quantity called theta that i mentioned", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3101.572, "end": 3107.6, "text": " So yeah, theta basically decides how much of reasoning that the agent should solve.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3107.62, "end": 3118.954, "text": "So the stopping criteria is written as like one minus theta times the KL divergence between the feedforward policy as well as the inferred or the tested time scaled policy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3120.176, "end": 3123.78, "text": "So yeah, so that's basically about the stopping criteria.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3124.762, "end": 3129.688, "text": "Yeah, so now we talked about perception planning and action policies.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3130.293, "end": 3138.884, "text": " So now we have to incorporate this land policies or the land posterior distributions into the world model.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3139.365, "end": 3150.84, "text": "For that, we can actually write the generative world model as is shown here, which is composed of the joint distribution of observation state policy as well as the A and B parameters here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3150.86, "end": 3153.203, "text": "We can actually write this joint distribution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3153.964, "end": 3158.089, "text": "And then, yeah, so how do we actually compute this?", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3158.423, "end": 3186.237, "text": " distributions a and b so if you do variation inference we can actually write the posterior distributions for these components a as is shown here so basically what we do is yeah in the variation free energy expression we can actually filter out all those stems that are independent of a then if you actually consider this prior distribution of a to be a Dirichlet distribution so we choose Dirichlet because", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3186.217, "end": 3214.654, "text": " so as i mentioned i mentioned that this your a a is actually a categorical distribution so in order for the posterior to be tractable uh we need to find a conjugate distribution so that's the reason why we basically choose this dirichlet distribution so traditionally distribution can actually be parameterized by this quantity called alpha ij here and we actually can find a closed form expression for this alpha ij", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3214.634, "end": 3230.787, "text": " um yeah a close to home expression for this posterior distribution as is shown here so as you can see here um so what basically this expression does is that it basically okay basically counts the number of joint occurrences of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3230.767, "end": 3254.074, "text": " uh the state um j as well as the the particular observation auto equal to i so basically but yeah so basically as we get more evidence about the world observations so we actually basically update the observation model based on that so that's what basically this observation model posterior inference does similarly", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3254.054, "end": 3282.508, "text": " uh yeah i won't go into the details of these things yeah similarly we actually compute the posterior distribution for the transition model b again we use the same procedure of minimizing the variation inference and we also assume that the b also follows the dirichlet prior distribution which basically leads to a tractable posterior distribution as is shown here here again we can see that basically what we do is basically counting the number of co-occurrence of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3282.488, "end": 3309.711, "text": " uh s store minus one and s store so that's what basically this distribution does uh so this basically leads to uh okay given that um yeah so and given that we also assume that the states um corresponding to different um physical assets are independent in the posterior distribution we can actually write the posterior distribution as a scaled factor as is shown here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3309.961, "end": 3336.711, "text": " uh and yeah so so basically this b i j times basically compute the ij value of this b matrix uh and finally we also compute the posterior distribution for the policy again considering that the prior distribution for the policy follows a dirichlet distribution leading to a tractable posterior distribution as i mentioned uh previously so that's that's how we basically", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3336.691, "end": 3364.504, "text": " um sold the transition model observation and the policy uh for the world model updates so uh to conclude uh the entire these four stages uh perception planning action policy and finally world model update occurred by minimizing a single quantity called variational free energy uh yeah so that's that's that's about the solution so i will um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3365.699, "end": 3367.428, "text": " give it back to Omar now.", "speaker": "Daniel Friedman"}, {"start": 3372.895, "end": 3373.498, "text": "Okay.", "speaker": "Daniel Friedman"}, {"start": 3381.393, "end": 3384.397, "text": " OK, thank you, Christo, for the solution.", "speaker": "Daniel Friedman"}, {"start": 3384.917, "end": 3395.71, "text": "So as Christo was saying, just to recap both parts of the story, the first part shows how we can actually model everything.", "speaker": "Omar Hashash"}, {"start": 3395.95, "end": 3409.206, "text": "So all the perception, planning, action, everything can be modeled as an inference process in this case, which is mathematically also equivalent to a gradient descent process, basically machine learning.", "speaker": "Omar Hashash"}, {"start": 3409.186, "end": 3425.344, "text": " At the same time, we can find tractable solutions through approximate Bayesian inference in order to do the perception, the planning, the action, and the learning at the end to enable this continual type of learning.", "speaker": "Omar Hashash"}, {"start": 3425.945, "end": 3433.033, "text": "So now what we're going to do is that we're going to see some simulation results around the jaywalking.", "speaker": "Omar Hashash"}, {"start": 3433.013, "end": 3453.52, "text": " example that we have been covering since the beginning of the presentation so we're going to simulate now at test time basically our solution which is a test time solution in comparison to other solutions like Q-learning for example and see how they're going to react in that specific scenario", "speaker": "Omar Hashash"}, {"start": 3453.5, "end": 3468.91, "text": " so what's going to happen is that we're going to train uh new solutions we're going to consider the base solution our solution to be a q learning on top of it we have this minimum reasoning active influence capability that we're talking about the reasoning to minimize free energy", "speaker": "Omar Hashash"}, {"start": 3468.89, "end": 3474.46, "text": " At the same time, we have a Q-learning solution that does not have that capability.", "speaker": "Omar Hashash"}, {"start": 3474.941, "end": 3482.855, "text": "So first of all, of course, when the world changes and now we're going to have a jaywalking pedestrian that the agent has not seen before.", "speaker": "Omar Hashash"}, {"start": 3483.115, "end": 3485.7, "text": "So for example, when you have a green traffic light,", "speaker": "Omar Hashash"}, {"start": 3486.575, "end": 3489.198, "text": " it always sees that there's no pedestrian over there.", "speaker": "Omar Hashash"}, {"start": 3489.278, "end": 3494.265, "text": "So it's kind of confident that it needs to move forward and just go towards its goal.", "speaker": "Omar Hashash"}, {"start": 3494.285, "end": 3495.346, "text": "This is what it was trained for.", "speaker": "Omar Hashash"}, {"start": 3495.786, "end": 3502.074, "text": "But at that specific moment, when the jaywalking pedestrian passes, it should have a prediction error.", "speaker": "Omar Hashash"}, {"start": 3502.094, "end": 3506.66, "text": "But the fact that it has no world model on top of its policy, it cannot do it.", "speaker": "Omar Hashash"}, {"start": 3506.68, "end": 3508.983, "text": "So this is exactly what happens here on the left-hand side.", "speaker": "Omar Hashash"}, {"start": 3510.064, "end": 3513.128, "text": "We can see here that the agent starts at distance 3.", "speaker": "Omar Hashash"}, {"start": 3513.148, "end": 3515.451, "text": "It starts to move towards distance 1.", "speaker": "Omar Hashash"}, {"start": 3516.359, "end": 3525.545, "text": " But then what's going to happen is that at step two, the agent, the jaywalking pedestrian appears.", "speaker": "Omar Hashash"}, {"start": 3526.147, "end": 3527.23, "text": "In that case,", "speaker": "Omar Hashash"}, {"start": 3528.34, "end": 3534.049, "text": " What's going to happen, as I said, because it relies just on its policy, it's going to move forward.", "speaker": "Omar Hashash"}, {"start": 3534.249, "end": 3538.896, "text": "And it's going to still accelerate, because this is what it was used to in the beginning.", "speaker": "Omar Hashash"}, {"start": 3539.397, "end": 3542.222, "text": "And it did not realize that its world changed in that case.", "speaker": "Omar Hashash"}, {"start": 3542.803, "end": 3546.368, "text": "So it starts to accelerate, and it crashes on the pedestrian, of course.", "speaker": "Omar Hashash"}, {"start": 3546.388, "end": 3551.496, "text": "This is typical with methods like Q-learning.", "speaker": "Omar Hashash"}, {"start": 3551.476, "end": 3558.445, "text": " But however, if we look at our solution, the scale policy one, we can see here that what happened is that the agent stopped at distance two.", "speaker": "Omar Hashash"}, {"start": 3558.465, "end": 3559.766, "text": "It did not move after that.", "speaker": "Omar Hashash"}, {"start": 3560.367, "end": 3568.718, "text": "So you can see here in the background, there's a change towards a yellow color, which means that the agent is now in a reasoning type of mode.", "speaker": "Omar Hashash"}, {"start": 3569.298, "end": 3573.103, "text": "So what happened is that unlike the Q-learning agent, it actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3573.724, "end": 3575.827, "text": "So the Q-learning agent actually sped up.", "speaker": "Omar Hashash"}, {"start": 3576.487, "end": 3578.53, "text": "This one actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3578.51, "end": 3590.577, "text": " and kept itself in its place until the jaywalking pedestrian actually left, after which the prediction error is resolved and it starts to accelerate again.", "speaker": "Omar Hashash"}, {"start": 3590.618, "end": 3593.464, "text": "Now, that's a scenario we haven't trained it on before.", "speaker": "Omar Hashash"}, {"start": 3593.798, "end": 3602.738, "text": " But the fact that it has a world model and can reason in those specific cases, it can actually figure out a way in order to keep itself surviving.", "speaker": "Omar Hashash"}, {"start": 3603.179, "end": 3611.858, "text": "And of course, after that, it continues after that specific moment from distance two until it reaches its destination at the other end.", "speaker": "Omar Hashash"}, {"start": 3612.715, "end": 3613.877, "text": " behind that intersection.", "speaker": "Omar Hashash"}, {"start": 3614.798, "end": 3625.855, "text": "So looking here at the rewards, at least, if we look at the Q-Learning rewards, of course, what happened over here is that when the agent crashed, it had a very big negative penalty over here, of course.", "speaker": "Omar Hashash"}, {"start": 3627.037, "end": 3630.522, "text": "However, in our case, what happened is that the agent was able", "speaker": "Omar Hashash"}, {"start": 3630.502, "end": 3654.214, "text": " uh to detect that there's something wrong okay there's a surprise so it detected the surprise and you can see it here from the purple line over here so the surprise went above the threshold which is this 0.2 this epsilon threshold and after that it started to reason so you can see here what happened is that the agent is still getting some kind of reward in the background for surviving", "speaker": "Omar Hashash"}, {"start": 3655.544, "end": 3664.114, "text": " And after that, what was happening in this region between the two lines over here, the agent was actually reasoning.", "speaker": "Omar Hashash"}, {"start": 3664.815, "end": 3675.288, "text": "So it was reasoning in order to adapt and scale its policy in order to avoid the crashing scenario that it may face over here, although it lost a bit of its rewards.", "speaker": "Omar Hashash"}, {"start": 3675.929, "end": 3678.011, "text": "So we're going to see this now in a different state.", "speaker": "Omar Hashash"}, {"start": 3678.031, "end": 3682.476, "text": "We're going to compare our solution with a Q-learning solution and a Bayesian RL solution.", "speaker": "Omar Hashash"}, {"start": 3682.516, "end": 3685.5, "text": "Of course, Q-learning is a", "speaker": "Omar Hashash"}, {"start": 3685.48, "end": 3691.39, "text": " is a model-free scenario, model-free solution, and Bayesian RL is a model-based solution.", "speaker": "Omar Hashash"}, {"start": 3691.971, "end": 3694.095, "text": "So in order just to draw contrast with them.", "speaker": "Omar Hashash"}, {"start": 3694.615, "end": 3697.881, "text": "So what happens here is that we're going to train the three scenarios.", "speaker": "Omar Hashash"}, {"start": 3698.222, "end": 3703.27, "text": "We can see here that roughly they have the same training reward.", "speaker": "Omar Hashash"}, {"start": 3703.25, "end": 3708.221, "text": " And then at test time, we're going to let them see the jaywalking pedestrian.", "speaker": "Omar Hashash"}, {"start": 3709.002, "end": 3717.561, "text": "What's going to happen is that the reward will drop massively for Q-learning and Bayesian RL, and also for our solution.", "speaker": "Omar Hashash"}, {"start": 3717.581, "end": 3723.874, "text": "It's going to drop, but not that severe as the Q-learning and Bayesian RL scenarios.", "speaker": "Omar Hashash"}, {"start": 3723.854, "end": 3726.958, "text": " Although it does, it really drops.", "speaker": "Omar Hashash"}, {"start": 3727.059, "end": 3729.562, "text": "However, this comes at the expense of success.", "speaker": "Omar Hashash"}, {"start": 3730.183, "end": 3741.72, "text": "So our agent, in order to scale, it had to sacrifice its narrow rewards in order to ensure the success of its long-term objective, right?", "speaker": "Omar Hashash"}, {"start": 3741.74, "end": 3743.342, "text": "Because this is what RL is about.", "speaker": "Omar Hashash"}, {"start": 3743.362, "end": 3746.166, "text": "It's about maximizing the long-term reward.", "speaker": "Omar Hashash"}, {"start": 3746.534, "end": 3768.39, "text": " uh in order to ensure success while efficiently utilizing inference so what it did over here is that it showed us that you don't actually need to do reasoning and inference all the time you just need to do it when needed so it can be adaptive not like what a lot of people think now that if you add always add reasoning on top of the models it's always better", "speaker": "Omar Hashash"}, {"start": 3768.758, "end": 3775.127, "text": " So the efficient human-like way is that reasoning is only included once you know this is human behavior.", "speaker": "Omar Hashash"}, {"start": 3775.167, "end": 3789.366, "text": "And so it showed around 36% improvement more than above the Bayesian RL solution in terms of utilizing the computing resources in that case.", "speaker": "Omar Hashash"}, {"start": 3789.386, "end": 3795.154, "text": "So it shows a trade-off between model-based and model-free solutions.", "speaker": "Omar Hashash"}, {"start": 3795.775, "end": 3809.017, "text": " And after that, what we did is that we trained the agent that we have with our solution on a jaywalking pedestrian scenario.", "speaker": "Omar Hashash"}, {"start": 3809.037, "end": 3814.426, "text": "But we allowed it to see that scenario again and again in order to see what's going to happen.", "speaker": "Omar Hashash"}, {"start": 3815.007, "end": 3821.277, "text": "So basically what happened in that case is that the first time it saw the scenario, it had a very high surprise.", "speaker": "Omar Hashash"}, {"start": 3821.257, "end": 3845.307, "text": " after that the surprise started to decrease over time because it started to update and started to make sense that i'm always going to see a jaywalking pedestrian at that specific scenario it got used to it basically and you can see it here from the fact that it starts to predict from the other graph you can start to see in the blue graph that it starts to predict that there's a pedestrian whenever i have a green light", "speaker": "Omar Hashash"}, {"start": 3845.759, "end": 3851.13, "text": " So it starts to update the probability of the agent having a pedestrian in that case.", "speaker": "Omar Hashash"}, {"start": 3852.132, "end": 3856.861, "text": "So basically, it becomes less surprising for the agent.", "speaker": "Omar Hashash"}, {"start": 3856.881, "end": 3865.438, "text": "And at the same time, at the level of the policy itself, it also updates the probability of taking that action in that specific state.", "speaker": "Omar Hashash"}, {"start": 3865.621, "end": 3877.632, "text": " So basically what's happening is that every time it's seeing the scenario, it's reinforcing that I should see a jaywalking pedestrian at that green light in the jaywalking scenario that I'm going to see.", "speaker": "Omar Hashash"}, {"start": 3878.052, "end": 3887.841, "text": "And I know very well what my action, that specific action that I was able to get through my reasoning, I know very well what that action should be and it reinforces it in the policy.", "speaker": "Omar Hashash"}, {"start": 3888.542, "end": 3894.447, "text": "So after 50 episodes, what happens is that the decision making moves from being", "speaker": "Omar Hashash"}, {"start": 3894.427, "end": 3915.321, "text": " controlled through inference through reasoning to become feed forward because basically it becomes a part of its work model and the policy after that in order to get it basically it gets reinforced over time at test time so you can see here that even with we have reinforcement learning now at test time and this is basically how the agent can become a continual learning agent", "speaker": "Omar Hashash"}, {"start": 3916.567, "end": 3924.176, "text": " So just to conclude this now, this is just a comprehensive roadmap of wireless communication and AI working together.", "speaker": "Omar Hashash"}, {"start": 3924.216, "end": 3930.504, "text": "Basically, our networks are moving forward very fast, just like AI.", "speaker": "Omar Hashash"}, {"start": 3930.544, "end": 3939.696, "text": "And what we want is that we have this ultimate intersection between the two so that we can reach what we call as AGI native networks in the future.", "speaker": "Omar Hashash"}, {"start": 3939.716, "end": 3943.3, "text": "So we're moving from AI native networks to AGI native networks.", "speaker": "Omar Hashash"}, {"start": 3943.28, "end": 3957.675, "text": " where these things build on the different technologies and systems that we talked about today, the world model, the digital twins, our presented test time scaling, and more broadly, a big cognitive system to architecture of the brain.", "speaker": "Omar Hashash"}, {"start": 3958.376, "end": 3963.721, "text": "And with that, I'll conclude the presentation, and I'll be happy to answer any questions you may have.", "speaker": "Omar Hashash"}, {"start": 3964.642, "end": 3965.503, "text": "Crystal and I, of course.", "speaker": "Daniel Friedman"}, {"start": 3968.426, "end": 3968.787, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 3969.427, "end": 3969.928, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 3972.743, "end": 3975.365, "text": " Yeah, just one short personal comment.", "speaker": "SPEAKER_00"}, {"start": 3975.505, "end": 3983.172, "text": "I think this really builds nicely from Christo's previous presentation, which was about the semantic information theory.", "speaker": "SPEAKER_00"}, {"start": 3983.192, "end": 4002.75, "text": "And this takes it and really connects it quite well to the passive, active planning, online learning setting and how those features are being implemented in a functional sense by the different recent generations of LLM.", "speaker": "SPEAKER_00"}, {"start": 4002.73, "end": 4010.245, "text": " and sort of associated systems like the skills, tools, sub-agent dispatch, chain of thought, all those kinds of things.", "speaker": "SPEAKER_00"}, {"start": 4011.808, "end": 4012.069, "text": "All right.", "speaker": "SPEAKER_00"}, {"start": 4013.732, "end": 4014.854, "text": "Krista, do you want to add anything?", "speaker": "Daniel Friedman"}, {"start": 4015.255, "end": 4016.858, "text": "And then I'll read some questions.", "speaker": "Daniel Friedman"}, {"start": 4023.301, "end": 4023.502, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 4023.522, "end": 4024.925, "text": "All right.", "speaker": "Daniel Friedman"}, {"start": 4025.306, "end": 4033.547, "text": "If anyone in live chat asks a question, I'll read it, but I'll start with a few questions that were submitted by email from Joel Robinson.", "speaker": "SPEAKER_00"}, {"start": 4035.89, "end": 4046.562, "text": " Okay, Jill wrote, I've been reading active inference as the test time scaling law for physical AI agents and had several questions that arise from my own work on regulatory architectures.", "speaker": "SPEAKER_00"}, {"start": 4047.323, "end": 4057.934, "text": "I'm asking these because the empirical literature suggests certain variables must exist in any physical or biological agent, and I'm trying to understand how they map onto your formulation.", "speaker": "SPEAKER_00"}, {"start": 4059.155, "end": 4059.636, "text": "So here we go.", "speaker": "SPEAKER_00"}, {"start": 4061.135, "end": 4064.099, "text": " Question one, inference capacity as a state variable.", "speaker": "SPEAKER_00"}, {"start": 4064.84, "end": 4078.016, "text": "You write that the scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time, and that that policy update is modeled as a soft Bayesian inference process.", "speaker": "SPEAKER_00"}, {"start": 4079.838, "end": 4084.324, "text": "In cognitive neuroscience, inference capacity is state dependent and collapses under load.", "speaker": "SPEAKER_00"}, {"start": 4084.878, "end": 4090.288, "text": " So where in your formulation is the agent's capacity to perform inference represented?", "speaker": "SPEAKER_00"}, {"start": 4091.069, "end": 4099.624, "text": "Is inference assumed to be always available, or is there a state variable that modulates whether the agent can actually execute the Bayesian update?", "speaker": "SPEAKER_00"}, {"start": 4100.526, "end": 4103.511, "text": "And I'm going to paste this into the Zoom chat as well.", "speaker": "SPEAKER_00"}, {"start": 4107.322, "end": 4115.953, "text": " Yeah, so I think it's the last part of the question that's really exciting.", "speaker": "Omar Hashash"}, {"start": 4116.414, "end": 4126.087, "text": "So I do agree with the question itself as being there are limitations for the architecture.", "speaker": "Omar Hashash"}, {"start": 4127.469, "end": 4133.717, "text": "But in our case, I think we're considering mostly kind of like optimal scenarios, optimal execution scenarios.", "speaker": "Omar Hashash"}, {"start": 4134.777, "end": 4140.367, "text": " So the thing is that I agree is that I think inference collapses.", "speaker": "Omar Hashash"}, {"start": 4140.847, "end": 4145.575, "text": "I think the word was collapses in certain scenarios.", "speaker": "Omar Hashash"}, {"start": 4145.596, "end": 4156.574, "text": "But in our case, we specifically highlighted the fact here that we are using our system two, sorry, our system one most of the time to take actions.", "speaker": "Omar Hashash"}, {"start": 4157.256, "end": 4160.481, "text": "So in that case, our resources for reasoning", "speaker": "Omar Hashash"}, {"start": 4160.731, "end": 4188.409, "text": " assuming that we have a lot of time possibly not instantaneous reflective reflexive like a reflex while driving that i have enough time to think about the different scenarios that i can take in that specific scenario so i'm not considering here a where like an example where an agent where a human for example needs to take a split second action i'm feeling that there needs to be enough time", "speaker": "Omar Hashash"}, {"start": 4188.389, "end": 4201.842, "text": " so that the agent thinks because assuming i'm on a highway and there's like roughly four five six seconds which i think is more than enough for a human to think in that case at least in the broad sense", "speaker": "Omar Hashash"}, {"start": 4202.21, "end": 4213.169, "text": " So these are the type of problems that we're considering over here, which in that case, I think the inference part doesn't crash specifically in that case or collapses in that case.", "speaker": "Omar Hashash"}, {"start": 4213.189, "end": 4226.091, "text": "But I do acknowledge the fact that there are situations where this kind of has to be all short circuited, I would say, and you just have to pass to just taking an instantaneous reflex.", "speaker": "Omar Hashash"}, {"start": 4228.045, "end": 4254.222, "text": " Yeah, one point I'll add there, and this is often brought up by the RxInfer and sort of message graph developers, is that with the message graph doing local free energy minimization, you can have different frequencies or different number of iterations of optimization between observations.", "speaker": "SPEAKER_00"}, {"start": 4254.202, "end": 4257.747, "text": " You could be doing an observation slower or at a faster frequency.", "speaker": "SPEAKER_00"}, {"start": 4258.188, "end": 4262.534, "text": "You could have intermittent or a fix or a variable compute budget.", "speaker": "SPEAKER_00"}, {"start": 4263.275, "end": 4277.796, "text": "And so that's like an advantage for online methods is that you're kind of just doing as good as you can in an online fashion given the amount of cognitive or computational resourcing available.", "speaker": "SPEAKER_00"}, {"start": 4279.244, "end": 4289.809, "text": " Modulo, anything like a catastrophic failure, that kind of breaks out of the box of the performance on the self-driving car itself breaking down mechanically.", "speaker": "SPEAKER_00"}, {"start": 4289.849, "end": 4291.994, "text": "It's sort of the mortal computing.", "speaker": "SPEAKER_00"}, {"start": 4292.054, "end": 4294.62, "text": "That's like the embodiment side.", "speaker": "SPEAKER_00"}, {"start": 4294.6, "end": 4319.376, "text": " and here is like really scoping out a lot of the formalisms and methods for the computational the cognitive that kind of presuppose the integrity of the functioning machine even though that becomes also a question when you're talking about these like network systems yeah i i i believe you said it in the best way possible and so uh it's uh", "speaker": "SPEAKER_00"}, {"start": 4319.71, "end": 4323.435, "text": " What you raised over here are the perfect times at that specific moment.", "speaker": "Omar Hashash"}, {"start": 4323.876, "end": 4327.781, "text": "So of course, we're working on a strict budget over here, but we're considering that.", "speaker": "Omar Hashash"}, {"start": 4329.303, "end": 4331.265, "text": "So I'm processing it in a different way.", "speaker": "Omar Hashash"}, {"start": 4331.666, "end": 4344.503, "text": "So I think our goal of this work was to lay down the foundation of how can an agent exist in the world in order to solve the problem that", "speaker": "Omar Hashash"}, {"start": 4345.023, "end": 4349.61, "text": " the problem that we were facing or possibly what we were missing before in that case.", "speaker": "Omar Hashash"}, {"start": 4350.692, "end": 4360.568, "text": "Of course, the more practical sense that you were talking about, which is different frequencies that you may be taking in samples in about the world, the different noise", "speaker": "Omar Hashash"}, {"start": 4360.835, "end": 4365.928, "text": " the noises that you may have, how much computationally feasible is that.", "speaker": "Omar Hashash"}, {"start": 4366.068, "end": 4368.515, "text": "I think these things are right on point.", "speaker": "Omar Hashash"}, {"start": 4369.056, "end": 4371.803, "text": "And they should follow exactly from this specific moment.", "speaker": "Omar Hashash"}, {"start": 4371.823, "end": 4375.232, "text": "So I think this is roughly a", "speaker": "Omar Hashash"}, {"start": 4375.887, "end": 4399.026, "text": " a i wouldn't say a sketch but i would say as a initial solution of a problem uh that we can start with and of course on top of that we can add a different uh feasibility type of thing i assume there's a lot of solution that needs to come up in this space more but the tricky part about it was uh as you were saying about the different um", "speaker": "Omar Hashash"}, {"start": 4399.006, "end": 4403.395, "text": " connecting it into the broader AI scenario, like what was really missing.", "speaker": "Omar Hashash"}, {"start": 4403.415, "end": 4407.644, "text": "Like we're seeing these different LLMs, chain of thought, reasoning, planning, agentic.", "speaker": "Omar Hashash"}, {"start": 4408.566, "end": 4412.034, "text": "But we already knew that, like we already had agents before.", "speaker": "Omar Hashash"}, {"start": 4412.054, "end": 4413.156, "text": "So what was actually missing?", "speaker": "Omar Hashash"}, {"start": 4413.176, "end": 4416.002, "text": "What did we really not see before?", "speaker": "Omar Hashash"}, {"start": 4416.269, "end": 4433.012, "text": " And it was actually that fact that I think that the new part in LLMs was mostly based on the test time part more than the regular vanilla kind of scaling the training part, the model size, the parameters.", "speaker": "Omar Hashash"}, {"start": 4432.992, "end": 4442.342, "text": " So we try to focus on that more, try to see why is it really that I have an empirical scaling law with language models, and what's the intersection over there?", "speaker": "Omar Hashash"}, {"start": 4443.183, "end": 4447.768, "text": "So yeah, it was putting down the theoretical part of the story, I think, at first.", "speaker": "Omar Hashash"}, {"start": 4448.229, "end": 4453.575, "text": "But definitely, the next steps should be based on making it more computationally tractable.", "speaker": "Omar Hashash"}, {"start": 4455.016, "end": 4455.317, "text": "Awesome.", "speaker": "SPEAKER_00"}, {"start": 4455.457, "end": 4457.219, "text": "All right, here's another question from Joel.", "speaker": "SPEAKER_00"}, {"start": 4457.86, "end": 4458.24, "text": "He wrote,", "speaker": "SPEAKER_00"}, {"start": 4460.06, "end": 4484.515, "text": " temporal depth and prediction window collapse you note that humans simulate counterfactual scenarios and plan future states of the world and this is central to resolving prediction error empirically temporal depth is not fixed it narrows under stress fatigue or overload is temporal depth treated as a dynamic state variable variable in your model if the prediction window collapse how does the scaling law behave", "speaker": "SPEAKER_00"}, {"start": 4488.19, "end": 4495.268, "text": " So if I got the answer correctly, you were asking about the temporal depth?", "speaker": "Omar Hashash"}, {"start": 4496.872, "end": 4499.078, "text": "Yeah, like the depth of planning.", "speaker": "SPEAKER_00"}, {"start": 4500.577, "end": 4502.921, "text": " OK, yeah.", "speaker": "SPEAKER_00"}, {"start": 4502.981, "end": 4511.875, "text": "So of course, the depth here is, of course, focused on we're having it as the least shallow layer, I would say.", "speaker": "Omar Hashash"}, {"start": 4511.935, "end": 4517.323, "text": "Of course, we're not considering that we have a hierarchical form of the world, of course, which we know about.", "speaker": "Omar Hashash"}, {"start": 4518.385, "end": 4523.513, "text": "And we're not considering the different temporal aspects of the world that can exist over there.", "speaker": "Omar Hashash"}, {"start": 4523.493, "end": 4528.862, "text": " Of course, the next steps should be that this world is composed in a hierarchical way.", "speaker": "Omar Hashash"}, {"start": 4529.262, "end": 4530.645, "text": "There's a lot of structure in there.", "speaker": "Omar Hashash"}, {"start": 4531.166, "end": 4540.982, "text": "At which level of planning should we actually limit ourselves in so that we don't fall into the kind of collapse that Joel's talking about over here?", "speaker": "Omar Hashash"}, {"start": 4541.002, "end": 4548.454, "text": "So I think even there are some interesting parts over here related to hierarchical active inference.", "speaker": "Omar Hashash"}, {"start": 4548.822, "end": 4557.631, "text": " So the hierarchical part of modeling the world, which I assume should be also our next objective over here.", "speaker": "Omar Hashash"}, {"start": 4557.932, "end": 4560.52, "text": "I'm not sure if Christo wants to add a certain part over here.", "speaker": "Daniel Friedman"}, {"start": 4564.972, "end": 4576.104, "text": " Yeah, so I think Christo, because he had some words on hierarchical abstractions and hierarchical representations.", "speaker": "Omar Hashash"}, {"start": 4576.785, "end": 4582.371, "text": "So this is a different part of the depth that we were talking about, about the world model.", "speaker": "Omar Hashash"}, {"start": 4582.391, "end": 4588.498, "text": "So we're talking about the depth of the world in terms of the aspects that it may have hierarchically.", "speaker": "Omar Hashash"}, {"start": 4588.478, "end": 4603.135, "text": " And at the same time, we're talking about the depth in terms of time going millisecond by millisecond, or going in terms of representations that can possibly last longer over time, that don't change in an instantaneous manner.", "speaker": "Omar Hashash"}, {"start": 4604.157, "end": 4610.284, "text": "Yeah, I'll just copy two more questions in from Joel.", "speaker": "SPEAKER_00"}, {"start": 4613.707, "end": 4630.551, "text": " on some related topics these are all great great comments so just so it's all um there the one one area was about generalization and again i think it speaks to this nexus of basically like", "speaker": "SPEAKER_00"}, {"start": 4630.531, "end": 4635.501, "text": " there's what is the computational, like structure learning?", "speaker": "SPEAKER_00"}, {"start": 4635.862, "end": 4639.65, "text": "There still could be novel situations that are outside your structure learning envelope.", "speaker": "SPEAKER_00"}, {"start": 4640.371, "end": 4647.105, "text": "There could be like unknown unknowns, or there could be things that were, or there could even be structural possibilities that aren't adjacencies.", "speaker": "SPEAKER_00"}, {"start": 4647.727, "end": 4650.85, "text": " Then there's the implementation of the computer resources.", "speaker": "SPEAKER_00"}, {"start": 4651.831, "end": 4655.194, "text": "The math is like the computer science is free.", "speaker": "SPEAKER_00"}, {"start": 4656.195, "end": 4666.664, "text": "And then there's these questions about the capacity and how the compute system's capacity re-enters into the test time.", "speaker": "SPEAKER_00"}, {"start": 4666.704, "end": 4675.472, "text": "And what if you develop the test time in one situation and then it has a poor drop-off", "speaker": "SPEAKER_00"}, {"start": 4675.452, "end": 4702.911, "text": " in a compute limited setting but i think that your that um point about this um training objective or constraint being an underappreciated or underutilized factor in large model training is very key and i hope that it comes through yeah so um", "speaker": "SPEAKER_00"}, {"start": 4703.498, "end": 4710.371, "text": " I'll try to hit on the part which you're talking about, the unknown unknowns, those parts.", "speaker": "Omar Hashash"}, {"start": 4710.711, "end": 4717.123, "text": "So of course, there are still room enough for a lot of things to pop up in the world.", "speaker": "Omar Hashash"}, {"start": 4717.143, "end": 4725.118, "text": "And the fact that what we were dealing with the world is we're trying to deal with the part of the world", "speaker": "Omar Hashash"}, {"start": 4726.094, "end": 4737.353, "text": " that we don't really know about so that's the whole premise of the whole story which is the fact that you want to escape from the training grounds and escape into the test time part", "speaker": "Omar Hashash"}, {"start": 4738.126, "end": 4740.21, "text": " Of course, at this time, there's always going to be capacity.", "speaker": "Omar Hashash"}, {"start": 4740.23, "end": 4753.715, "text": "I think we do have, we included some resources in our paper about like, for example, if you have a new scenario that that's actually not explained by none of our states, let's say if you're doing perception.", "speaker": "Omar Hashash"}, {"start": 4754.255, "end": 4757.061, "text": "So in that case, you would need to, let's say, for example, expand your model.", "speaker": "Omar Hashash"}, {"start": 4757.261, "end": 4762.21, "text": "So you need to go through from a Bayesian sense, you would need to look at Bayesian model expansion.", "speaker": "Omar Hashash"}, {"start": 4762.73, "end": 4767.641, "text": " But of course, these have other limitations, and you have to specify the criteria.", "speaker": "Omar Hashash"}, {"start": 4767.661, "end": 4777.363, "text": "Just like we said that we have a prediction error, and we have a certain epsilon, that if we cross that prediction error, this means that you're going to foresee the scenario.", "speaker": "Omar Hashash"}, {"start": 4777.597, "end": 4788.038, "text": " would probably be in a case where you're trying to fit in your certain state of the world into one of the states that you have, but probably none of those states are actually explaining your situation.", "speaker": "Omar Hashash"}, {"start": 4788.198, "end": 4791.925, "text": "So in that case, the most logical thing to say is that this doesn't fit anything.", "speaker": "Omar Hashash"}, {"start": 4791.966, "end": 4793.749, "text": "This is something very new.", "speaker": "Omar Hashash"}, {"start": 4793.729, "end": 4799.375, "text": " you might probably need to say that this is something like a composition of different states together.", "speaker": "Omar Hashash"}, {"start": 4799.395, "end": 4801.037, "text": "I haven't seen this before.", "speaker": "Omar Hashash"}, {"start": 4801.097, "end": 4810.668, "text": "But in our case, let's say we took the simplistic first step about this, which is the fact that just the probability distribution is changing.", "speaker": "Omar Hashash"}, {"start": 4811.509, "end": 4819.939, "text": "So in a very simple scenario that all of us, I think, see roughly in our lifetime, which is the fact that someone is jaywalking.", "speaker": "Omar Hashash"}, {"start": 4819.919, "end": 4821.822, "text": " So this is something very easy to grasp.", "speaker": "Omar Hashash"}, {"start": 4822.062, "end": 4829.594, "text": "And the fact that we already know what a human is, what a traffic light is, what a green light is, we already know the traffic rules.", "speaker": "Omar Hashash"}, {"start": 4830.115, "end": 4831.617, "text": "So this is something very well known.", "speaker": "Omar Hashash"}, {"start": 4832.258, "end": 4840.772, "text": "In that case, we just flip the rules a little bit so that rather than crossing at a red light, you're crossing at a green light to see what the behavior is.", "speaker": "Omar Hashash"}, {"start": 4840.852, "end": 4844.037, "text": "But of course, I'll keep your imagination", "speaker": "Omar Hashash"}, {"start": 4844.017, "end": 4865.805, "text": " to help you in that case to see how much still we have a lot more to deal with in the future but i think that should be the main focus of this whole thing is trying to figure out step by step what are those critical things how can they actually be posed at this time so that we found find a", "speaker": "Omar Hashash"}, {"start": 4866.477, "end": 4869.164, "text": " first principle kind of solution to these systems.", "speaker": "Omar Hashash"}, {"start": 4869.185, "end": 4874.259, "text": "So building on the fact that all of us solve that problem in that certain way.", "speaker": "Omar Hashash"}, {"start": 4875.823, "end": 4877.989, "text": "So I think that should hit directly on that point.", "speaker": "Daniel Friedman"}, {"start": 4880.585, "end": 4896.881, "text": " Yeah, one area that I see this being very relevant to is the distillation and the on-policy distillation, the teacher, the co-learner, all these kind of model transformation and training strategies.", "speaker": "SPEAKER_00"}, {"start": 4898.583, "end": 4910.355, "text": "All different policy training strategies and this as a comparable measure, even if one", "speaker": "SPEAKER_00"}, {"start": 4911.989, "end": 4930.825, "text": " um test time is well it's 30 hours for this long horizon coding agent it's eight hours for this camera image model even though they're totally different domains they're not going to have a domain benchmark that's comparable and then for the models that are multimodal", "speaker": "SPEAKER_00"}, {"start": 4930.805, "end": 4938.658, "text": " or they're used in a multimodal setting, there aren't benchmarks at the domain level that would apply.", "speaker": "SPEAKER_00"}, {"start": 4938.678, "end": 4947.952, "text": "So it ends up being a sort of weakest link question with these systems sometimes.", "speaker": "SPEAKER_00"}, {"start": 4949.956, "end": 4957.808, "text": "I think that connects to the model training strategies, which are upstream of how well these models do.", "speaker": "SPEAKER_00"}, {"start": 4958.851, "end": 4972.725, "text": " And the fact, this is actually the first point that I wanted to talk about, which is you don't really have a benchmark of what's good or bad, which is the fact that you need to figure things out.", "speaker": "Omar Hashash"}, {"start": 4973.165, "end": 4985.938, "text": "In order to do that, there needs to be a first principle kind of way to make sure at least that it's doing the minimum best possible thing that any human could do.", "speaker": "Omar Hashash"}, {"start": 4986.745, "end": 4994.597, "text": " which is the fact of why we're looking at this active inference part, which is putting it as a limiting factor, actually than putting it as an upper bound.", "speaker": "Omar Hashash"}, {"start": 4995.078, "end": 5002.069, "text": "So we're saying that this is the minimum amount of reason that you need to have in order to solve this type of problem.", "speaker": "Omar Hashash"}, {"start": 5002.45, "end": 5005.114, "text": "But this does not mean that it's the only reason.", "speaker": "Omar Hashash"}, {"start": 5005.154, "end": 5012.145, "text": "So you can add on top of that other things and you can just expand as much as possible, but at least,", "speaker": "Omar Hashash"}, {"start": 5012.463, "end": 5024.216, "text": " to exist in the world and to solve problems like the problems that we see with these autonomous vehicles, Teslas, Waymos, to think about those humanoid robots that just crash down and fall.", "speaker": "Omar Hashash"}, {"start": 5024.957, "end": 5029.081, "text": "So we're trying to at least address that type of problem.", "speaker": "Omar Hashash"}, {"start": 5029.982, "end": 5031.724, "text": "But of course, that's bare minimum.", "speaker": "Daniel Friedman"}, {"start": 5035.228, "end": 5035.588, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 5035.969, "end": 5037.07, "text": "Yeah, and the paper.", "speaker": "Daniel Friedman"}, {"start": 5038.248, "end": 5040.19, "text": " really lays out that direction.", "speaker": "SPEAKER_00"}, {"start": 5041.351, "end": 5042.592, "text": "Do you have any last comments?", "speaker": "SPEAKER_00"}, {"start": 5042.712, "end": 5045.774, "text": "Otherwise, I hope this has been a good entry point to the topic.", "speaker": "SPEAKER_00"}, {"start": 5047.436, "end": 5048.056, "text": "Yeah.", "speaker": "Omar Hashash"}, {"start": 5048.797, "end": 5050.498, "text": "It's actually been a pleasure.", "speaker": "Omar Hashash"}, {"start": 5050.518, "end": 5054.222, "text": "I think the questions were really great and exciting.", "speaker": "Omar Hashash"}, {"start": 5055.543, "end": 5057.424, "text": "And I really enjoyed presenting over here today.", "speaker": "Omar Hashash"}, {"start": 5058.345, "end": 5059.146, "text": "Thank you for having us.", "speaker": "Omar Hashash"}, {"start": 5060.367, "end": 5060.727, "text": "Thank you.", "speaker": "SPEAKER_00"}, {"start": 5061.007, "end": 5063.569, "text": "Thank you to authors and to Joel for the questions.", "speaker": "SPEAKER_00"}, {"start": 5064.21, "end": 5065.331, "text": "So, bye.", "speaker": "Daniel Friedman"}, {"start": 5066.932, "end": 5067.152, "text": "Bye-bye.", "speaker": "Daniel Friedman"}, {"start": 5067.172, "end": 5067.453, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 5067.533, "end": 5067.833, "text": "Thank you.", "speaker": "Daniel Friedman"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt new file mode 100644 index 000000000..43c5b9c98 --- /dev/null +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt @@ -0,0 +1,1310 @@ +Daniel Friedman: +Hello, welcome. + + +SPEAKER_00: +It's July 14th, 2026. + +We're in active guest stream 128.1 on active inference as the test time scaling law for physical AI agents. + +Thank you, Omar and Christo, for joining, and please take it away for the presentation. + + +Daniel Friedman: +Okay. + + +Omar Hashash: +Thank you, Daniel, for having us today in this presentation. + +Today, as you said, we'll be presenting one of our new works. + +So first of all, for the people that don't know me, I'm Omar Hashash. + +I'm a postdoc at Virginia Tech. + +And today I'll be also joined with my colleague, Christa Thomas, which is an assistant professor at WPI. + +And together we'll be presenting one of our recent works, which is active inference as a test time scaling law for physical AI. + +So it's a bit of an interesting work. + +It gives a bit of new concepts to the role of active inference and how it plays it in physical AI and why is it really necessary to be posed as a scaling law for physical AI agents. + +So just as a brief + +introduction uh so i recently finished my phd at the brand the department of electrical computer engineering at virginia tech back in december 2025 and where my focus was on wireless communications and ai including topics like world models and digital twins and edge intelligence and i'll pass it over here to crystal just a little bit to introduce himself + + +Christo Kurisummoottil Thomas: +uh yeah thanks for more uh i'm christopher thomas i am currently an assistant professor at uh wpi and i lead a research group uh named trustworthy resilient ai and networks um and i did my prayer to this i did my post postdoc associate at virginia tech ec department + +And my research mainly lies at the intersection of AI, native wireless networks, semantic communication, physical AI, and mathematical foundations of AI. + + +Omar Hashash: +Yeah, so both of us have this kind of hybrid background on AI and wireless communications at the same time. + +And we're going to see how that plays a role in extending us more into neuroscience concepts like active inference. + +So without any further notice, I think we can start first with the fact that AI has become the most transformational technology that is changing all our worlds. + +And of course, one of those worlds is actually the wireless communications world, and especially the 6G, the next generation of wireless communications that we're expecting to see roughly in a couple of years. + +So we already started to touch on improvements with the current generations of AI that we have. + +We call these networks as AI-native networks. + +We started to see improvements in terms of efficiency and latency. + +The improvements are massive, actually. + +But that's only the good part of the story. + +But there's also, as any other story, there's also the bad and the ugly. + +So how do the bad and the ugly look like? + +Well, they look something like this. + +So the fact that we're training what we call as an AI native interface, for example, we train it to send beams to users in the network. + +So these users can be humans, they can be physical agents. + +So a part of these + +uh beams actually work very well but the fact is that the world is always non-stationary and dynamic and it's always changing so these so these kind of uh tasks hold some sort of problems and even fail at test time um so even if we look at the level of the agents themselves that are connected to the network we can see that we have plenty of examples on failures as well so for example we have autonomous vehicles that are trained on millions and millions of driving trials + +They still continue to fail in ways that we just can't comprehend in the world. + +At the same time, we have these kind of robots that are trained to win Olympics, for example, like this robot over here, but still crash into a man on an athletics track, even though it's trained to win an Olympics. + +So there's a lot of things that don't make a lot of sense over here. + +And ironically, with the current generation of AI, we do have these big, large language models that can actually win in math and math. + +So I think here now we can pose a question, which is, why is it that today's AI is good enough to win math Olympiad, but it fails at this type of Olympiad, or more generally, in very simple green world scenarios that are very easy for us, especially with this massive amount of training? + +I'm not going to answer this now, but I'm going to use it as a motivation for the rest of the presentation and to say that we're actually missing something very big over here. + +so so if i want to look now at what do we want from these ai ages so the real the physical agents the robots the the vehicle the network so what we want from these ai agents is to become adaptive to the worlds that are happening we can start from the first one the fact that the world is always changing uh we need to train it on something but there's always going to be more data that we didn't train on so we need to have a um + +a principled way that we act in the states that we didn't actually see before. + +At the same time, we want those agents to be generalizable. + +So we want them to train them on something, but at the same time, to be able to use them on something else. + +At the same time, we want to be autonomous. + +We don't want to stay in the loop or probably minimize our presence in the loop with the AI system as much as possible. + +At the same time, we want them to have this continual learning capability. + +So we wanted to learn on top and top and top of other things. + +So we're starting from a very far away part over here with the current generation of AI, which is mostly based on large language models like the GPT models that we have, which are a massive success. + +So part of the success that started with GPT, at least the new part, was introducing these reasoning capabilities. + +So I think back with the old models, the old three models, the old one model, we started to see how this reasoning can pop up into these models and actually give you a better answer for the agents. + +So for the first time, we had this new kind of test time scaling law, which started from the fact that if you give it more computed test time, you give the agent more computed test time, you can have a better answer. + +So this was followed by another kind of capability, which is called planning. + +And this is what ushered in the agentic era or AI agents that we see today. + +And this is where GPT-5 and all the clouds that we have today. + +So basically, planning is the capability of breaking down a complex task into a sequence of actions in order to achieve a goal. + +So this is where we currently roughly are right now, probably more or less. + +But we can see here that we're moving on a cognitive path. + +so what we can say is that the current forms of reasoning and planning are good for language models are probably good enough for language models but are not enough for real world agents because the world works in a different way than language so the world is driven by laws that works apart from language how language works in terms of syntax and all those stuff so but that that route seems to be right + +so if we were to build on these cognitive capabilities and the fact that they were starting to work so this means that there should be other cognitive capabilities that we're searching for in order to reach the characteristics of the AI systems that we're searching for so if you want to dig deep here and try to search for this missing cognitive capability that's needed to move forward with the world + +we can see that it needs to be a cognitive capability that is in direct touch with the world and it's focused on the world. + +And that capability is nothing but something called common sense. + +So what is common sense? + +Common sense is, as the name implies, it's the common amount of knowledge of how the world works that we humans use in order to make sense of our world. + +It's our ability to understand the world and make use of it, which is something that is found in each human of us. + +So that's great. + +How can we plug in this common sense? + +So common sense is a broadly defined term, but we kind of know if you want to introduce this common sense into AI age, we have to search for the component of the cognitive system that drives it. + +So that component is actually what we call as a work model. + +So what is a world model? + +A world model of the physical world allows humans to understand the world in terms of its real-time state, in terms of the causal structures that exist between the elements of the world. + +And at the same time, it allows us to understand the dynamical evolution of how this world will evolve with time. + +so but the introduction of the world model itself so the fact that you're taking a state of the world and trying to understand it and trying to see how it's going to evolve in the future this requires other cognitive capabilities such as perception for example so perception is the capability to actually understand the world and try to grasp it and to make meaning of it but if we try to zoom out now at this specific moment + +What are we trying to say with all of these cognitive capabilities? + +So we said that we have that ladder that started with reasoning, we added planning, we added common sense for the world model, and now we're saying that we need perception. + +What are we trying to say with all these cognitive capabilities? + +What we're indirectly acknowledging over here is the fact that we need a cognitive solution that includes all these cognitive capabilities on top of today's AI. + +So, and this is not something new. + +We already know that because we already know that the AI that we have today, that's roughly based on neural networks, they're called system one architectures from psychology. + +They're called system one architectures because they're reactive. + +uh but we know that we're missing these system two type of architectures so we know already that the system one it composes around roughly 95 of how we take actions in the world and this is what we've based ai systems on since the beginning and now + +What we're trying to do is to look at these other system two rational type of thinking ways and trying to see how can these come up with a new architecture in order to solve the problems that we have. + +so how does this architecture look like it looks something roughly like this it looks something like these cognitive modules that we have that i talked about the perception module and planning the world model all connected together uh of course over a network and the network acts here as a bridge uh or additional computing resource for the ages that exist in the world + +so that it benefits itself for the decisions, intelligent decisions that it can take, just like the beamforming example I was giving at the beginning, and all the other intelligent actions that it can help the agents do in the world. + +So this was based on one of our previous works. + +And here, I'll shout out to one of our colleagues + +which mentioned our presentation in one of the theoretical neurobiology groups before and + +gave us credit for this nice work. + +So this is actually us, yours truly over here with Crystal. + +And what he was presenting is that our work is kind of an intersection with other leading works in the field, like, for example, Yann LeCun's vision of modular cognitive brain architecture and his famous model of the world. + +But what we did over here is that we actually + +We push things a little bit forward. + +So what we kind of did is that we saw the other intersection with what Carfis is actually also presenting with active inference and the free energy principle. + +And we tried to combine all of these things together into one kind of coherent story around world models that we're going to see here today. + +So going back to this architecture that I was talking about, the first step to actually use this architecture is to sense the world. + +So we already, in terms of wireless networks, already studied this before. + +We have the abilities now in next generation networks to develop sensing capabilities that allow us to divide this world over wireless networks. + +We said this before in previous works. + +And I'll try to highlight here a little bit on the fact that we have these little + +red dots that i have over here in red which are called digital twins so digital twins are the copies or the replicas or the different models of the ai agents of these physical ai agents so they're the models of the agents the physical agents that exist in the world and we have the world models + +that are basically the other things around it. + +And digital twins are a part of the world model itself, along with the different elements that exist in the world. + +So basically, this is just the high-level concept how digital twins work. + +So they're based on the fact that they're a part of the world model. + +They intersect with world models. + +Both of them, they allow + +Basically, with a digital twin, you can actually move that work model forward. + +And this intersects with the notion of AI, which is called action-conditioned work models. + +So digital twins allow you to move forward in time. + +At the same time, they allow you to send some configuration back to what we call the physical twin in the physical world. + +So these configurations, we still don't know what they are, basically, in 6G. + +And this is actually what we'll be uncovering in this work. + +So basically, as I was saying, digital twins, they take real-time updates from the world, and they return some real-time optimization feedback back to the agents in the world. + +And their role is roughly somewhere around what-if analysis and doing predictions and monitoring for the AI agent. + +And we've also shown in one of our previous work before that these agents can have continual learning capabilities, but we're going to show it more in depth over here today. + +So roughly going back to that architecture that I was talking about, now we can look at that part where we talk about the agent, the digital twin. + +of the agent as going forward in time as a process of reasoning, because we're seeing it from a different lens, not from a 6G wireless communications lens, but probably from a neuroscience or AI point of view, where reasoning is actually trying to optimize some certain action for the agent in the world. + +But at the same time, what we want to do over here is similar to what LLMs do. + +So we want to reason in order to take, just like LLMs reason to give you a better answer for some harder prompt, for example. + +What we want to do is that we actually want to use this reasoning capability that's now on the network in order + +to allow the agent to reason to take a better action in the world. + +But that will require something like test time scaling, the one that was shown empirically in large language models, but is basically missing today in these physical AI systems. + +so this is basically the goal of our work today we're going to present how these how this reasoning over the network can benefit the ai agents so that it reasons to make to take a better action and generalize in new unforeseen scenarios that it hasn't been trained on and to see how this actually results once built from first principles like active inference it results in a test time scaling law similar to the test time scanning law that we saw with large language models + +So this is our work that I'll be presenting today. + +And it has been recently released on archive for people that want to read more in depth about it. + +So let's start with test time scaling in the real world. + +So before we start, we have to look at the story from the beginning. + +So what we were doing before with our L agents, for example, is that we used to train them. + +And then after a certain time, we would reach a, for example, we're talking, of course, you were talking, for example, about some example of autonomous driving. + +what we used to do is that we used to train these autonomous agents and then we used to plug them in the world after some convergence of policy we just to plug them in the world and just we just hope that things work out perfectly fine but that's not always the case and ai fails once the world changes and of course we have a lot of examples about that + +so before before we go to solve the problem what we have to do is to actually go back to the origin so what are we actually doing over there how does the brain actually does this type of learning and to see what we are actually missing from the story + +So to start first, we have to acknowledge that we have two regions of the brain that are working together. + +Basically, we have the prefrontal cortex where all the learning, reasoning, planning, the world model, everything happens over there. + +And then after that, once we reach that convergence that I was talking about over here, what we do is that we take the policy and plug it back in the basal ganglia. + +And after that, roughly what we do is that we really don't need all the effort from the prefrontal cortex anymore because that's the smart part of our brain. + +We don't need to devote any big effort for that task. + +And it becomes somehow a subconscious task at that specific moment. + +And we start to rely more on our policy while working. + +And this is actually what we did before. + +But the question is now, if I were to ask you, how much do you think about driving while driving? + +the answer would probably be around zero. + +You don't think about driving while driving. + +And this is something we know, and this is something agrees with what the famous psychologist, Daniel Kahneman, was a Nobel Prize winner, actually identified before. + +So he roughly said that we use our intuition system one around 95% of the time in order to take actions. + +but at the same time there's this five percent that we also take actions with so when does this five percent pop up actually and why is it just five percent so one example of how we can use that is through surprise + +So for example, if you're here driving a vehicle, and let's say you're on a highway, and then you get a jaywalking pedestrian, for example, that suddenly wants to cross the road, you weren't expecting it. + +And actually, after that, you got a surprise that you didn't expect, right? + +Because if you didn't expect something, you wouldn't get a surprise at the beginning. + +And once that happens, what you do in that case is that you start to think. + +So if I move forward, I probably hit the person, but I don't want that to happen. + +So this is what you'd likely be saying, or probably saying that if I possibly go slower, the person behind me will probably bump into me and I'll head into a crash, another crash. + +So probably what I need to do is probably just lower my speed just a little bit to allow the person to move in front of me, cross whatever they want to go. + +And then after that, I can just continue my path. + +So this type of thinking over here is exactly necessary in order to pass that certain situation that I wasn't really trained on. + +I didn't really expect in front of me. + +So roughly, if I want to combine it now, so this is kind of how we do this type of decision making. + +So basically, when you're driving, if there's no prediction error in front of you, you did not expect anything surprising. + +Everything in the world is working precisely as you would want it to do. + +You basically use your Bayes of Ganglia and basically the policy that exists in your Bayes of Ganglia. + +But at the same time, if you do have a prediction error, + +this means that your policy is somehow sub-optimal and you need to go through your frontal cortex again you need to kick it in again until it handle this new situation for you so what you would do in that case is that you would reason before you take the right action at that specific moment + +So now I probably ask you, where did we see the system before? + +This kind of system that toggles between reasoning and not reasoning. + +This is exactly what we've seen in + +models and large language models like GPT-5. + +So that's smart routing. + +So this is, you could say probably this is from where it came from. + +I'm going to see how that connects afterwards, but basically what the agents, the soft kind of agents, digital agents, language agents that we see today, when they reason or don't reason, they follow that specific scheme that is in the brain. + +So basically the prediction error is the, um, + +is the natural trigger, rather than having a holistic router. + +The router is trying to approximate that trigger, natural trigger. + +But the question here becomes, how do we actually deal with the scaling at that specific moment? + +I said that we thought and took an action. + +how is the action actually resulting over here? + +How is it actually coming into place? + +We still don't know. + +And that's something still mysterious in the AI space. + +But you see here what I presented just now in terms of that + +prediction error and trying to reason or to move that prediction error. + +This is exactly what active inference is. + +So active inference is a first principle that describes how all living systems survive in the world. + +And returning back to the example that I presented in the beginning, that humanoid robot and that people that is crashing. + +So this is how that in other ways that + +the absence of this minimum amount of freezing on top of the AIs that we have today can actually explain why those systems are dying systems in the world. + +If this is how all systems survive in the world, all living systems like us, humans or animals, + +survival in the world, and the fact that this is absent in those autonomous agents, autonomous physical AI agents that exist in the world, you can possibly find an explanation to the fact why these systems are dying. + +So of course, uh, in contrast to what, uh, to what we have, uh, we can see here that what we need to do actually is to allow those systems to survive at this time, because as I said, it's the first principle of that explains how we survive in the world. + +So survival at that specific moment over here, when we reached the conversions, so we were reaching something in non equilibrium steady state in terms of physics. + +So from a genetic point of view, you need to maintain + +non-equilibrium steady state by preserving the policy and the world model so that you can survive in the world. + +So preserving that actually has two routes. + +The first route is what existing works do, which is that they assume stationary conditions about the environment. + +So in that case, the system works just in the states that you actually saw on the training set. + +And survival is limited to just those conditions. + +But + +There's an alternative route for the system to survive, which is by integrating the first principles of survival from the beginning. + +So you don't need to actually make your world frozen at this time. + +You can keep it dynamic and non-stationary. + +but give the agent the first principle of survival so that it survives in dynamic, non-stationary worlds. + +So in that case, survival is no longer limited just to what the agent saw in the training set. + +Now it actually survives even in new situations, and it allows it to generalize, which was our main goal from the beginning. + +What we want is an agent that generalizes in unforeseen scenarios that we haven't seen before. + +And in that case, by definition, + +Surviving in the world is the main facet of a system, of an AI agent, basically, that generalizes. + +So by guaranteeing survival, you're actually guaranteeing that the agent generalizes. + +So, roughly speaking, this is what active inferences and what Carl Friston presented it as a way to remove prediction errors from the brain. + +Of course, we always have these sensory information. + +We're seeing if we don't have a prediction error, it means that we're in the region that minimizes free energy or basically surprised and we don't need to take any action. + +However, if we do have a prediction error, we will try to act on the world so that we minimize that prediction error. + +uh so of course in this way we can see that reasoning to reduce surprise is the minimum amount of reasoning in any in any living system to survive in the world that's the minimum amount of intelligence that you need to have in order to exist in the world over a long time so what we kind of did over here is that we tried to integrate this into what we know in terms of agents that we just don't need them to just survive we want to make them usable so that they can actually perform tasks + +So we went back to integrate this with what we had in terms of reinforcement learning and the policies that we had. + +So what happens in this case is that what we saw is that if we take in states and if those states are somehow unforeseen, it means that we can go reason + +in order to remove the prediction error that results from this unforeseen scenario. + +And at the same time, if we can use that test time scaling, we can use it to alter the actions to fit the specific scenario in order to make it optimal in that specific scenario so that the agent can actually generalize in that specific moment. + +So roughly speaking, if you want to put it down graphically over here in the figure, what we were doing with reinforcement learning is that we were looking at those states that are highly probable and trying to focus on those states. + +finding a policy that fits those states. + +This is basically what reinforcement learning is doing. + +And after that, what we did is that we froze the world. + +We said that the world is not going to change. + +So we're always going to be stuck in those states over there and those policies. + +But we know that that fades in the real world. + +What's actually happening in the real world is that because it's non-stationary, it will probably push you to a state + +That is not a problem for you. + +It's not a part of the set that you've seen in your training distribution. + +What you need to do with that type of reasoning to minimize prediction errors is actually what pushes you back or scales you back to your characteristic set, which is the set of states where you can use your policy again. + +So this is basically what we mean by test time scaling. + +In this case, we can see, and this is a very nice viewpoint on the story, which is that every living system, in this case, shows a general objective to survive in the world. + +And specifically, I'm saying here something general, because this is something that all living systems perform in the world. + +So the fact that you are a living system, it means that you have an objective to survive in the world. + +or else you would die so in that case we are governed by general objectives like survival over here and at the same time we have narrow objectives which allow us to achieve our specific tasks and goals with subsumed within that specific region of survival that we have + +So I'll just give here an example, just like the jaywalking example of here, just to say how this system kind of works. + +So basically, what we have is the fact that I have an autonomous driving vehicle, for example. + +It's driving. + +And everything's fine. + +I have no prediction error. + +This is something that we call passive influence in that case. + +So we're just taking in sensory observations. + +Everything is working as expected. + +I have no problem. + +And then suddenly what I have is that an agent that decides to jaywalk. + +Sorry, a human that decides to jaywalk. + +In that case, the network, of course, the network is acting here as an additional computing resource in addition to the agent that we have. + +you don't need to focus a lot on what the network is doing. + +The network is similar to the agent in that case. + +They're kind of one system together. + +So in this case, the network that we have will detect that there's a prediction error, and it will switch from passive to active inference about the world. + +So what that means is that the network will engage in reasoning in order to minimize the surprise. + +And it's going to use those digital twins that I talked about at the beginning, which capture the model of the agent and the work model itself in order to reason. + +And what we want to do is that we want to reason, just like I was talking about, I was giving an example here, should I move forward faster? + +Should I move slower? + +So all that thinking together, we're going to use it in order to put down a value function. + +So the value function here, we call it as a delta v. So basically, I'm trying to measure which action allows me to resolve my prediction error in that specific state that I am in. + +And then what I want to do is that I'm going to take in that feedback from the network and send it back to the physical AI agent. + +The physical AI agent + +What it's going to do is that it's going to use this feedback. + +It's like the prefrontal cortex sending the signals down to the basal ganglia. + +So the policy is over here at the level of the agent of the vehicle here. + +And it's going to use it in order to reason and do inference. + +So it's going to scale its policy from pi node to pi node prime, for example, here. + +So in that specific case, the action that it's going to take is going to be different than its normal. + +It's going to be possibly to decrease its velocity in order to avoid an accident, for instance. + +And then what happens is that after this jaywalking pedestrian just passes, I go back to my regular cases where I have my prediction error resolved and the pedestrian passes. + +So I can go back just to my passive inventory. + +So what you can see here is that our world is basically a transition always between foreseen scenarios, unforeseen scenarios, resolving these unforeseen scenarios, and then going back to foreseen scenarios. + +This is how we survive in the world. + +We're always transitioning between those different states. + +So now I'm going to go into the system model. + +So as I said before, we have an agent that is in the world. + +It has a policy panel. + +And we have a world model that exists over the network in this architecture. + +One part of the world model is composed from these digital tools, these different alternative policies that the agent can have. + +or models of the agent. + +At the same time, we have other things that are called assets which capture the different things that exist in the world, just like the human, the houses, everything else besides the agents themselves. + +And what's going to happen is that we're going to take sensing observations. + +So it's not the agent itself that's doing the sensing, it's actually the network that's doing the sensing on behalf of the agent. + +And what we're going to do is that we're going to do perception. + +So we're basically going to map our observations to + +the set of states that we have trying to make sense of the states, trying to infer what's going to happen over here in terms of the states. + +And then once we have a surprise, so we have an unforeseen scenario, what's going to happen is that the network is going to engage in counterfactual reasoning to minimize surprise. + +So it's going to think about all these different alternative routes with the different policies that they may have over here, different alternative policies. + +And then, as I said, it's going to send back these digital twin configurations. + +back to the agent in order to scale the policy and generalize in the unforeseen scenario. + +So mathematically, we can look at perception, of course, as an inference process. + +But to do that, we have to first define our generative world model in terms of the states and observations and the policies that we have. + +Of course, before we do perception, we have to predict what the expected observation is going to be. + +And then we're going to get our sensing observations from the world. + +And then we're going to measure if there's a surprise or not. + +If we do have a surprise, of course, this is how we measure the surprise. + +It's just the easiest surprise. + +And after that, if we do have a surprise, it means that we're going to cross some threshold epsilon in order to signal that there's an unforeseen scenario in front of us. + +so the first thing that should happen in that case is that we should actually calculate the posterior because basically what we predicted about the world is not right we have an error so we have the first thing we have to do is to make sense of our world again by calculating this posterior after we saw our observations which is basically an inference process of it just we can see before we engage in plan and reason + +So, the planning part, it constitutes these digital twins and work models that engage in counterfactual reasoning. + +So, basically, the what-if analysis that I was giving an example about, in order to plan the plausible future world states. + +And the goal, of course, of this reasoning is to actually minimize future surprise. + +So our goal is to see how much surprise is going to result from each of these policies over here. + +And you can see that it's going to be nothing but a measure of the entropy at each of those states in the future. + +So from t plus 1 until the end of the time horizon that we have. + +And then after that, what we're going to do is that we're going to use this reasoning that we did in order to build a value function. + +We call it delta v. + +And where the reward is actually minimizing surprise. + +So as I said at the beginning, we have a general objective in the brain, which is to survive. + +So basically it has a value function over here. + +And then what we're going to do is that we're going to take this value and we're going to send it back to the agents in the world in order to scale the policy. + +how that looks like it's a little bit similar to what perception did so we have the perception as inference of course planning is also as inference over here because planning is an extension of minimizing surprises in the future so it's also planning as inference and now we have also we have to model action as inference but here it's a different type of conference + +the fact that when we did perception as influence the evidence that happened in front of us was 100 true it happened in front of us in the world so we're confident about it but when you want to do action as influence it's a little bit different + +So it's based on imaginary virtual data in our brain, right? + +When we were thinking, we were actually generating data. + +That thing didn't happen in the world. + +It happened inside our brain. + +But luckily, we have Judea Pearl, who came up with this method of virtual evidence. + +So he has a nice technique in order to update the beliefs in that case. + +So we can see here, if we write it down, this is going to be something called a soft Bayesian update. + +So it's soft because you don't really need to do a complete inference update over here. + +So if you write it down, you can see here at the last line, you can write it in terms of a feed forward and reasoning or inference term. + +which is something very nice. + +Because if you really look at it, if you put that theta, which is basically your normalized surprise, if the world is mostly predictable, so your theta or surprise is basically equal to 0. + +And if you put that exponential part over here equal to 0, you basically result in the feedforward policy that we had in reinforcement learning. + +The fact is that action is always an inference process. + +It's always modulated by that exponential term that we have over here. + +And this is what scales the policy once we have prediction errors and once we have an unforeseen scenario. + +So the policies in today's solution, they remain limited to the case where surprise is null due to the stationary assumptions about the world, just like we do in reinforcement learning. + +So reinforcement learning is not stationary by itself at this time. + +We made it stationary because we did not compensate for this factor over here in equation 5, which is clearly a special case of the equation. + +However, the inference at this part can be challenging if you want to really solve the problem. + +It can be challenging because marginalizing over all the possible states is computationally intractable in practice. + +So to find a better solution, we need to resort to a variational Bayesian inference solution. + +And I'll keep it here for Christo to present the solution. + + +Daniel Friedman: +I think I'll stop sharing for Christo to share. + +Thanks, Omar. + + +Christo Kurisummoottil Thomas: +So I will go over the solution roadmap first. + +As Omar mentioned, our solution is grounded in variational variation inference principle. + +So this evolves as a four-step cognitive loop, starting with perception, which is about perceiving the current state by minimizing a variational free energy. + +and then uh you know we go to the future time steps where we you know estimate the expected free energy static from the inferred state and then uh you know infer policy for actions that minimize the future surprise so that's the planning phase + +and then in the action phase uh so there is we introduced the concept of marco blanket which acts as a statistical boundary between agent and the environment and then policy scaling will be more modeled as a gradient descent on the free energy + +um and finally okay we incorporate this uh updated policy as well as uh the land posterior distributions about the states into you know updating the world model that so that again happens via vfe minimization so + +These four stages is grounded on the same or minimizing the same variation free energy concept. + +That's about the solution. + +Then first, let's look at the perception. + +So perception is modeled as inference, inferring the posterior distribution of the states S given the observations. + +But if you expand it in terms of the Bayes rule, you can see that this is interactable, computationally interactable because of the marginal distribution of the observations which appear in the denominator here. + +so here we move to a to find a tractable posterior distribution we actually uh formulate the kl divergence between the approximate posterior q of s comma pi zero + +and the KL divergence to the actual or the true posterior so that's that's the concept of variational inference so where if you actually expand this KL divergence we can see that it can be written as a variation free energy times variation of free energy as well as an additional yeah plus or so basically variation free energy can be written as a KL divergence plus a surprise time which is ln p of time here + +uh and so minimizing the scale divergence become equal to uh yeah minimizing variation and then okay once we actually once we actually find a posterior that minimizes the kl divergence along with that we all we are also computing or we are also estimating the surprise quantity which is here uh so that's the advantage of this variational variation inference formulation here + +um yeah so yeah so once we find an optimal or optimal posterior distribution that actually means that uh the resetting variation free energy is equal to the surprise quantity uh so which surprise the estimation of the surprise is actually needed for the policy scaling that uh homer mentioned in the previous slide + +So now the question is how we actually go about solving this. + +So this can be modeled as a sequential inference process, which can be formulated as a partially observed Markov decision process as shown here. + +And this leads to the following factor graph as is shown here. + +So where we can see that, okay, so we actually modeled the transition from the state's + +to the observations using the matrix A here. + +So we actually model all these transition probabilities as categorical distributions. + +and we use the matrix b to actually model the transition from the previous state to the current state st uh yeah so now optimizing the or minimizing the variation of free energy this leads to an expression as is shown here uh so okay this basically leads to variational message passing uh wherein we can see that okay so there is the message one which is basically uh accounting for the + +uh yeah basically a forward transition which basically depends upon the forward transition probability and message 2 which depends upon the backward transition probability which is the message 2 component here and then message 3 actually uh incorporates the uh observation likelihood distribution + +And further, in order to compute the posterior distribution, which is basically defined as the s variable here, we basically take a softmax over the logarithm tense here. + +So that's how we actually obtain the ln q of s or the posterior distribution here. + +and uh so this posterior distribution can be shown to be okay so as we mentioned we started with minimizing the variation free energy component so at converge at the equilibrium or yeah at the convergence this um optimal posterior distribution should um converge to that of the true posterior so it basically is minimizing the um + +variational free energy component so basically it's a gradient descent on the um free energy um component so that's what basically this uh yeah perception component basically shows and we can actually show that this uh variational free energy can be written to be uh okay the posterior times an error on the + +uh posterior posterior distribution so basically minimizing the variation of free energy is equivalent to minimizing the error on the estimated posterior or inferred posterior distributions + +and yeah so now uh when we go to planning so what we actually do is that uh so we look at future time steps and we we want to find the policy that minimizes the average free energy uh across time so that's basically so so for that we need expected free energy component + +and we show that this expected free energy component can be nicely decomposed into two components so one is the risk component which actually uh measures the kl divergence between uh the you know approximate uh oh basically the inferred um + +distribution of observations given the current policy how that so the inferred uh distribute inferred observational distribution how that is actually comparing to the uh expected or preferred observations or the preferences that we have uh for the agent so that's basically the risk component + +plus there is an ambiguity component that uh so basically basically it's a entropy of the probability distributions which is basically the likelihood probability distribution here which so basically if you want to minimize the expected free energy you should make sure that the inferred world model is actually close to that or inferred observations are basically close to that of the preferred observations as well as + +we should try to reduce the ambiguity in terms of the observations that the agent observe in future time steps. + +So that's basically the expected free energy component. + +uh yeah so now uh the the prior over the posterior the the prior over the policies can be written as s of max of uh minus gamma g so we should yeah the policy should ensure that it should always go in the direction of minimizing the expected free energy + +and uh yeah inferred policy uh inferred action policy can be written as um yeah summation over all the policies pi uh where yeah corresponding to action a times the the prior distribution of the particular policy pi here + +uh and we can yeah so that's that's the inferred policy and then finally as omar mentioned in the previous slides what we actually do is using this inferred policy so this inferred policy basically uh needs an estimate of the surprise so so basically inferred policy depends upon the expected free energy + +which basically is equivalent to the surprise quantity, as I mentioned. + +So basically, you cannot compute an exact value for the inference policy, but you can only estimate using the surprise quantity. + +So to estimate the surprise, we need to perform reasoning. + +So how much reasoning that you need to do depends upon this parameter called theta here. + +So that is similar to that of the test time computing LLMs like Omar mentioned. + +So that's how we perform the planning and do the test time scaling. + +Yeah, so now the question is whether all these computations are really stable or not. + +For that, we need to look at the concept of Markov blanket. + +So an AI agent that generalizes is a random dynamical system + +that exist over time so for the a for this ai agent to be able to generalize across unforeseen scenarios so there should exist a marco blanket which basically means that uh so marco blanket is consisting of those sensors okay so the agent interacts with the environment through the sensory state + +and so so basically environment influences the agent through the sensory state and the agent influences the environment through the active state so this active so this marco blanket compose which is composed of the active states as well as the sensory state uh that is something which you know separates the internal agent state from the external states + +yeah so that's how this the closed loop of planning um perception planning action and sensing happens and the key equation here is that so we can write the states x the flow for the state x just composed of this marker blanket states as shown in as is shown in here which is composed of two temps so one is the one is a great dissipative gradient time and another is the solenoid + +uh time which basically both of this time depends upon the log of the model evidence which is ln p of x here + +uh so yeah for the agent to exist over time so there should be a mod uh marco blanket which also means that it should actually maximize the this log p of x quantity or the model evidence quantity which is basically equal to minimizing a free and i mean minimizing uh my uh l minus ln p of x which is basically corresponding to uh free energy estimate + +uh yeah so that's what basically this um concept of marco blanket tells us and also this also gives us a principled criterion for principle stopping criterion which basically means that uh so when do when do the agent actually stop reasoning so that can be written like uh yeah that can be written as depending upon this quantity called theta that i mentioned + +So yeah, theta basically decides how much of reasoning that the agent should solve. + +So the stopping criteria is written as like one minus theta times the KL divergence between the feedforward policy as well as the inferred or the tested time scaled policy. + +So yeah, so that's basically about the stopping criteria. + +Yeah, so now we talked about perception planning and action policies. + +So now we have to incorporate this land policies or the land posterior distributions into the world model. + +For that, we can actually write the generative world model as is shown here, which is composed of the joint distribution of observation state policy as well as the A and B parameters here. + +We can actually write this joint distribution. + +And then, yeah, so how do we actually compute this? + +distributions a and b so if you do variation inference we can actually write the posterior distributions for these components a as is shown here so basically what we do is yeah in the variation free energy expression we can actually filter out all those stems that are independent of a then if you actually consider this prior distribution of a to be a Dirichlet distribution so we choose Dirichlet because + +so as i mentioned i mentioned that this your a a is actually a categorical distribution so in order for the posterior to be tractable uh we need to find a conjugate distribution so that's the reason why we basically choose this dirichlet distribution so traditionally distribution can actually be parameterized by this quantity called alpha ij here and we actually can find a closed form expression for this alpha ij + +um yeah a close to home expression for this posterior distribution as is shown here so as you can see here um so what basically this expression does is that it basically okay basically counts the number of joint occurrences of + +uh the state um j as well as the the particular observation auto equal to i so basically but yeah so basically as we get more evidence about the world observations so we actually basically update the observation model based on that so that's what basically this observation model posterior inference does similarly + +uh yeah i won't go into the details of these things yeah similarly we actually compute the posterior distribution for the transition model b again we use the same procedure of minimizing the variation inference and we also assume that the b also follows the dirichlet prior distribution which basically leads to a tractable posterior distribution as is shown here here again we can see that basically what we do is basically counting the number of co-occurrence of + +uh s store minus one and s store so that's what basically this distribution does uh so this basically leads to uh okay given that um yeah so and given that we also assume that the states um corresponding to different um physical assets are independent in the posterior distribution we can actually write the posterior distribution as a scaled factor as is shown here + +uh and yeah so so basically this b i j times basically compute the ij value of this b matrix uh and finally we also compute the posterior distribution for the policy again considering that the prior distribution for the policy follows a dirichlet distribution leading to a tractable posterior distribution as i mentioned uh previously so that's that's how we basically + +um sold the transition model observation and the policy uh for the world model updates so uh to conclude uh the entire these four stages uh perception planning action policy and finally world model update occurred by minimizing a single quantity called variational free energy uh yeah so that's that's that's about the solution so i will um + + +Daniel Friedman: +give it back to Omar now. + +Okay. + +OK, thank you, Christo, for the solution. + + +Omar Hashash: +So as Christo was saying, just to recap both parts of the story, the first part shows how we can actually model everything. + +So all the perception, planning, action, everything can be modeled as an inference process in this case, which is mathematically also equivalent to a gradient descent process, basically machine learning. + +At the same time, we can find tractable solutions through approximate Bayesian inference in order to do the perception, the planning, the action, and the learning at the end to enable this continual type of learning. + +So now what we're going to do is that we're going to see some simulation results around the jaywalking. + +example that we have been covering since the beginning of the presentation so we're going to simulate now at test time basically our solution which is a test time solution in comparison to other solutions like Q-learning for example and see how they're going to react in that specific scenario + +so what's going to happen is that we're going to train uh new solutions we're going to consider the base solution our solution to be a q learning on top of it we have this minimum reasoning active influence capability that we're talking about the reasoning to minimize free energy + +At the same time, we have a Q-learning solution that does not have that capability. + +So first of all, of course, when the world changes and now we're going to have a jaywalking pedestrian that the agent has not seen before. + +So for example, when you have a green traffic light, + +it always sees that there's no pedestrian over there. + +So it's kind of confident that it needs to move forward and just go towards its goal. + +This is what it was trained for. + +But at that specific moment, when the jaywalking pedestrian passes, it should have a prediction error. + +But the fact that it has no world model on top of its policy, it cannot do it. + +So this is exactly what happens here on the left-hand side. + +We can see here that the agent starts at distance 3. + +It starts to move towards distance 1. + +But then what's going to happen is that at step two, the agent, the jaywalking pedestrian appears. + +In that case, + +What's going to happen, as I said, because it relies just on its policy, it's going to move forward. + +And it's going to still accelerate, because this is what it was used to in the beginning. + +And it did not realize that its world changed in that case. + +So it starts to accelerate, and it crashes on the pedestrian, of course. + +This is typical with methods like Q-learning. + +But however, if we look at our solution, the scale policy one, we can see here that what happened is that the agent stopped at distance two. + +It did not move after that. + +So you can see here in the background, there's a change towards a yellow color, which means that the agent is now in a reasoning type of mode. + +So what happened is that unlike the Q-learning agent, it actually slowed down. + +So the Q-learning agent actually sped up. + +This one actually slowed down. + +and kept itself in its place until the jaywalking pedestrian actually left, after which the prediction error is resolved and it starts to accelerate again. + +Now, that's a scenario we haven't trained it on before. + +But the fact that it has a world model and can reason in those specific cases, it can actually figure out a way in order to keep itself surviving. + +And of course, after that, it continues after that specific moment from distance two until it reaches its destination at the other end. + +behind that intersection. + +So looking here at the rewards, at least, if we look at the Q-Learning rewards, of course, what happened over here is that when the agent crashed, it had a very big negative penalty over here, of course. + +However, in our case, what happened is that the agent was able + +uh to detect that there's something wrong okay there's a surprise so it detected the surprise and you can see it here from the purple line over here so the surprise went above the threshold which is this 0.2 this epsilon threshold and after that it started to reason so you can see here what happened is that the agent is still getting some kind of reward in the background for surviving + +And after that, what was happening in this region between the two lines over here, the agent was actually reasoning. + +So it was reasoning in order to adapt and scale its policy in order to avoid the crashing scenario that it may face over here, although it lost a bit of its rewards. + +So we're going to see this now in a different state. + +We're going to compare our solution with a Q-learning solution and a Bayesian RL solution. + +Of course, Q-learning is a + +is a model-free scenario, model-free solution, and Bayesian RL is a model-based solution. + +So in order just to draw contrast with them. + +So what happens here is that we're going to train the three scenarios. + +We can see here that roughly they have the same training reward. + +And then at test time, we're going to let them see the jaywalking pedestrian. + +What's going to happen is that the reward will drop massively for Q-learning and Bayesian RL, and also for our solution. + +It's going to drop, but not that severe as the Q-learning and Bayesian RL scenarios. + +Although it does, it really drops. + +However, this comes at the expense of success. + +So our agent, in order to scale, it had to sacrifice its narrow rewards in order to ensure the success of its long-term objective, right? + +Because this is what RL is about. + +It's about maximizing the long-term reward. + +uh in order to ensure success while efficiently utilizing inference so what it did over here is that it showed us that you don't actually need to do reasoning and inference all the time you just need to do it when needed so it can be adaptive not like what a lot of people think now that if you add always add reasoning on top of the models it's always better + +So the efficient human-like way is that reasoning is only included once you know this is human behavior. + +And so it showed around 36% improvement more than above the Bayesian RL solution in terms of utilizing the computing resources in that case. + +So it shows a trade-off between model-based and model-free solutions. + +And after that, what we did is that we trained the agent that we have with our solution on a jaywalking pedestrian scenario. + +But we allowed it to see that scenario again and again in order to see what's going to happen. + +So basically what happened in that case is that the first time it saw the scenario, it had a very high surprise. + +after that the surprise started to decrease over time because it started to update and started to make sense that i'm always going to see a jaywalking pedestrian at that specific scenario it got used to it basically and you can see it here from the fact that it starts to predict from the other graph you can start to see in the blue graph that it starts to predict that there's a pedestrian whenever i have a green light + +So it starts to update the probability of the agent having a pedestrian in that case. + +So basically, it becomes less surprising for the agent. + +And at the same time, at the level of the policy itself, it also updates the probability of taking that action in that specific state. + +So basically what's happening is that every time it's seeing the scenario, it's reinforcing that I should see a jaywalking pedestrian at that green light in the jaywalking scenario that I'm going to see. + +And I know very well what my action, that specific action that I was able to get through my reasoning, I know very well what that action should be and it reinforces it in the policy. + +So after 50 episodes, what happens is that the decision making moves from being + +controlled through inference through reasoning to become feed forward because basically it becomes a part of its work model and the policy after that in order to get it basically it gets reinforced over time at test time so you can see here that even with we have reinforcement learning now at test time and this is basically how the agent can become a continual learning agent + +So just to conclude this now, this is just a comprehensive roadmap of wireless communication and AI working together. + +Basically, our networks are moving forward very fast, just like AI. + +And what we want is that we have this ultimate intersection between the two so that we can reach what we call as AGI native networks in the future. + +So we're moving from AI native networks to AGI native networks. + +where these things build on the different technologies and systems that we talked about today, the world model, the digital twins, our presented test time scaling, and more broadly, a big cognitive system to architecture of the brain. + +And with that, I'll conclude the presentation, and I'll be happy to answer any questions you may have. + + +Daniel Friedman: +Crystal and I, of course. + +Awesome. + +Thank you. + + +SPEAKER_00: +Yeah, just one short personal comment. + +I think this really builds nicely from Christo's previous presentation, which was about the semantic information theory. + +And this takes it and really connects it quite well to the passive, active planning, online learning setting and how those features are being implemented in a functional sense by the different recent generations of LLM. + +and sort of associated systems like the skills, tools, sub-agent dispatch, chain of thought, all those kinds of things. + +All right. + + +Daniel Friedman: +Krista, do you want to add anything? + +And then I'll read some questions. + +Okay. + +All right. + + +SPEAKER_00: +If anyone in live chat asks a question, I'll read it, but I'll start with a few questions that were submitted by email from Joel Robinson. + +Okay, Jill wrote, I've been reading active inference as the test time scaling law for physical AI agents and had several questions that arise from my own work on regulatory architectures. + +I'm asking these because the empirical literature suggests certain variables must exist in any physical or biological agent, and I'm trying to understand how they map onto your formulation. + +So here we go. + +Question one, inference capacity as a state variable. + +You write that the scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time, and that that policy update is modeled as a soft Bayesian inference process. + +In cognitive neuroscience, inference capacity is state dependent and collapses under load. + +So where in your formulation is the agent's capacity to perform inference represented? + +Is inference assumed to be always available, or is there a state variable that modulates whether the agent can actually execute the Bayesian update? + +And I'm going to paste this into the Zoom chat as well. + + +Omar Hashash: +Yeah, so I think it's the last part of the question that's really exciting. + +So I do agree with the question itself as being there are limitations for the architecture. + +But in our case, I think we're considering mostly kind of like optimal scenarios, optimal execution scenarios. + +So the thing is that I agree is that I think inference collapses. + +I think the word was collapses in certain scenarios. + +But in our case, we specifically highlighted the fact here that we are using our system two, sorry, our system one most of the time to take actions. + +So in that case, our resources for reasoning + +assuming that we have a lot of time possibly not instantaneous reflective reflexive like a reflex while driving that i have enough time to think about the different scenarios that i can take in that specific scenario so i'm not considering here a where like an example where an agent where a human for example needs to take a split second action i'm feeling that there needs to be enough time + +so that the agent thinks because assuming i'm on a highway and there's like roughly four five six seconds which i think is more than enough for a human to think in that case at least in the broad sense + +So these are the type of problems that we're considering over here, which in that case, I think the inference part doesn't crash specifically in that case or collapses in that case. + +But I do acknowledge the fact that there are situations where this kind of has to be all short circuited, I would say, and you just have to pass to just taking an instantaneous reflex. + + +SPEAKER_00: +Yeah, one point I'll add there, and this is often brought up by the RxInfer and sort of message graph developers, is that with the message graph doing local free energy minimization, you can have different frequencies or different number of iterations of optimization between observations. + +You could be doing an observation slower or at a faster frequency. + +You could have intermittent or a fix or a variable compute budget. + +And so that's like an advantage for online methods is that you're kind of just doing as good as you can in an online fashion given the amount of cognitive or computational resourcing available. + +Modulo, anything like a catastrophic failure, that kind of breaks out of the box of the performance on the self-driving car itself breaking down mechanically. + +It's sort of the mortal computing. + +That's like the embodiment side. + +and here is like really scoping out a lot of the formalisms and methods for the computational the cognitive that kind of presuppose the integrity of the functioning machine even though that becomes also a question when you're talking about these like network systems yeah i i i believe you said it in the best way possible and so uh it's uh + + +Omar Hashash: +What you raised over here are the perfect times at that specific moment. + +So of course, we're working on a strict budget over here, but we're considering that. + +So I'm processing it in a different way. + +So I think our goal of this work was to lay down the foundation of how can an agent exist in the world in order to solve the problem that + +the problem that we were facing or possibly what we were missing before in that case. + +Of course, the more practical sense that you were talking about, which is different frequencies that you may be taking in samples in about the world, the different noise + +the noises that you may have, how much computationally feasible is that. + +I think these things are right on point. + +And they should follow exactly from this specific moment. + +So I think this is roughly a + +a i wouldn't say a sketch but i would say as a initial solution of a problem uh that we can start with and of course on top of that we can add a different uh feasibility type of thing i assume there's a lot of solution that needs to come up in this space more but the tricky part about it was uh as you were saying about the different um + +connecting it into the broader AI scenario, like what was really missing. + +Like we're seeing these different LLMs, chain of thought, reasoning, planning, agentic. + +But we already knew that, like we already had agents before. + +So what was actually missing? + +What did we really not see before? + +And it was actually that fact that I think that the new part in LLMs was mostly based on the test time part more than the regular vanilla kind of scaling the training part, the model size, the parameters. + +So we try to focus on that more, try to see why is it really that I have an empirical scaling law with language models, and what's the intersection over there? + +So yeah, it was putting down the theoretical part of the story, I think, at first. + +But definitely, the next steps should be based on making it more computationally tractable. + + +SPEAKER_00: +Awesome. + +All right, here's another question from Joel. + +He wrote, + +temporal depth and prediction window collapse you note that humans simulate counterfactual scenarios and plan future states of the world and this is central to resolving prediction error empirically temporal depth is not fixed it narrows under stress fatigue or overload is temporal depth treated as a dynamic state variable variable in your model if the prediction window collapse how does the scaling law behave + + +Omar Hashash: +So if I got the answer correctly, you were asking about the temporal depth? + + +SPEAKER_00: +Yeah, like the depth of planning. + +OK, yeah. + + +Omar Hashash: +So of course, the depth here is, of course, focused on we're having it as the least shallow layer, I would say. + +Of course, we're not considering that we have a hierarchical form of the world, of course, which we know about. + +And we're not considering the different temporal aspects of the world that can exist over there. + +Of course, the next steps should be that this world is composed in a hierarchical way. + +There's a lot of structure in there. + +At which level of planning should we actually limit ourselves in so that we don't fall into the kind of collapse that Joel's talking about over here? + +So I think even there are some interesting parts over here related to hierarchical active inference. + +So the hierarchical part of modeling the world, which I assume should be also our next objective over here. + + +Daniel Friedman: +I'm not sure if Christo wants to add a certain part over here. + + +Omar Hashash: +Yeah, so I think Christo, because he had some words on hierarchical abstractions and hierarchical representations. + +So this is a different part of the depth that we were talking about, about the world model. + +So we're talking about the depth of the world in terms of the aspects that it may have hierarchically. + +And at the same time, we're talking about the depth in terms of time going millisecond by millisecond, or going in terms of representations that can possibly last longer over time, that don't change in an instantaneous manner. + + +SPEAKER_00: +Yeah, I'll just copy two more questions in from Joel. + +on some related topics these are all great great comments so just so it's all um there the one one area was about generalization and again i think it speaks to this nexus of basically like + +there's what is the computational, like structure learning? + +There still could be novel situations that are outside your structure learning envelope. + +There could be like unknown unknowns, or there could be things that were, or there could even be structural possibilities that aren't adjacencies. + +Then there's the implementation of the computer resources. + +The math is like the computer science is free. + +And then there's these questions about the capacity and how the compute system's capacity re-enters into the test time. + +And what if you develop the test time in one situation and then it has a poor drop-off + +in a compute limited setting but i think that your that um point about this um training objective or constraint being an underappreciated or underutilized factor in large model training is very key and i hope that it comes through yeah so um + + +Omar Hashash: +I'll try to hit on the part which you're talking about, the unknown unknowns, those parts. + +So of course, there are still room enough for a lot of things to pop up in the world. + +And the fact that what we were dealing with the world is we're trying to deal with the part of the world + +that we don't really know about so that's the whole premise of the whole story which is the fact that you want to escape from the training grounds and escape into the test time part + +Of course, at this time, there's always going to be capacity. + +I think we do have, we included some resources in our paper about like, for example, if you have a new scenario that that's actually not explained by none of our states, let's say if you're doing perception. + +So in that case, you would need to, let's say, for example, expand your model. + +So you need to go through from a Bayesian sense, you would need to look at Bayesian model expansion. + +But of course, these have other limitations, and you have to specify the criteria. + +Just like we said that we have a prediction error, and we have a certain epsilon, that if we cross that prediction error, this means that you're going to foresee the scenario. + +would probably be in a case where you're trying to fit in your certain state of the world into one of the states that you have, but probably none of those states are actually explaining your situation. + +So in that case, the most logical thing to say is that this doesn't fit anything. + +This is something very new. + +you might probably need to say that this is something like a composition of different states together. + +I haven't seen this before. + +But in our case, let's say we took the simplistic first step about this, which is the fact that just the probability distribution is changing. + +So in a very simple scenario that all of us, I think, see roughly in our lifetime, which is the fact that someone is jaywalking. + +So this is something very easy to grasp. + +And the fact that we already know what a human is, what a traffic light is, what a green light is, we already know the traffic rules. + +So this is something very well known. + +In that case, we just flip the rules a little bit so that rather than crossing at a red light, you're crossing at a green light to see what the behavior is. + +But of course, I'll keep your imagination + +to help you in that case to see how much still we have a lot more to deal with in the future but i think that should be the main focus of this whole thing is trying to figure out step by step what are those critical things how can they actually be posed at this time so that we found find a + +first principle kind of solution to these systems. + +So building on the fact that all of us solve that problem in that certain way. + + +Daniel Friedman: +So I think that should hit directly on that point. + + +SPEAKER_00: +Yeah, one area that I see this being very relevant to is the distillation and the on-policy distillation, the teacher, the co-learner, all these kind of model transformation and training strategies. + +All different policy training strategies and this as a comparable measure, even if one + +um test time is well it's 30 hours for this long horizon coding agent it's eight hours for this camera image model even though they're totally different domains they're not going to have a domain benchmark that's comparable and then for the models that are multimodal + +or they're used in a multimodal setting, there aren't benchmarks at the domain level that would apply. + +So it ends up being a sort of weakest link question with these systems sometimes. + +I think that connects to the model training strategies, which are upstream of how well these models do. + + +Omar Hashash: +And the fact, this is actually the first point that I wanted to talk about, which is you don't really have a benchmark of what's good or bad, which is the fact that you need to figure things out. + +In order to do that, there needs to be a first principle kind of way to make sure at least that it's doing the minimum best possible thing that any human could do. + +which is the fact of why we're looking at this active inference part, which is putting it as a limiting factor, actually than putting it as an upper bound. + +So we're saying that this is the minimum amount of reason that you need to have in order to solve this type of problem. + +But this does not mean that it's the only reason. + +So you can add on top of that other things and you can just expand as much as possible, but at least, + +to exist in the world and to solve problems like the problems that we see with these autonomous vehicles, Teslas, Waymos, to think about those humanoid robots that just crash down and fall. + +So we're trying to at least address that type of problem. + + +Daniel Friedman: +But of course, that's bare minimum. + +Awesome. + +Yeah, and the paper. + + +SPEAKER_00: +really lays out that direction. + +Do you have any last comments? + +Otherwise, I hope this has been a good entry point to the topic. + + +Omar Hashash: +Yeah. + +It's actually been a pleasure. + +I think the questions were really great and exciting. + +And I really enjoyed presenting over here today. + +Thank you for having us. + + +SPEAKER_00: +Thank you. + +Thank you to authors and to Joel for the questions. + + +Daniel Friedman: +So, bye. + +Bye-bye. + +Thank you. + +Thank you. diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json new file mode 100644 index 000000000..bd6cacdf1 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json @@ -0,0 +1 @@ +[{"video_id": "MjeQeWeyYhE", "segments": [{"start": 3.727, "end": 29.498, "text": " all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense", "speaker": "SPEAKER_05"}, {"start": 29.798, "end": 40.317, "text": " And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two.", "speaker": "SPEAKER_05"}, {"start": 40.337, "end": 43.923, "text": "We're going to do active inputs proper, which is going to be very, very exciting going forward.", "speaker": "SPEAKER_05"}, {"start": 43.943, "end": 50.575, "text": "So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts.", "speaker": "SPEAKER_05"}, {"start": 51.027, "end": 56.833, "text": " that we're going to start doing actual, quote unquote, active inference in the second part.", "speaker": "SPEAKER_05"}, {"start": 56.853, "end": 67.444, "text": "But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection.", "speaker": "SPEAKER_05"}, {"start": 68.305, "end": 72.409, "text": "We're going to sort of slow down and think about part one in general, chapters one to five.", "speaker": "SPEAKER_05"}, {"start": 73.531, "end": 80.578, "text": "This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one.", "speaker": "SPEAKER_05"}, {"start": 80.812, "end": 85.079, "text": " you know, going so that we can be adequately grounded going forward.", "speaker": "SPEAKER_05"}, {"start": 85.981, "end": 94.696, "text": "I will say we have one one question is, you know, we've done each chapter across two weeks.", "speaker": "SPEAKER_05"}, {"start": 95.658, "end": 100.947, "text": "Historically, you know, chapter one, we've had two weeks going forward.", "speaker": "SPEAKER_05"}, {"start": 101.72, "end": 128.115, "text": " depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week", "speaker": "SPEAKER_05"}, {"start": 128.095, "end": 152.685, "text": " of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe", "speaker": "SPEAKER_05"}, {"start": 152.935, "end": 159.784, "text": " It would be a good idea to reach out through the blankets email.", "speaker": "SPEAKER_05"}, {"start": 160.425, "end": 166.993, "text": "Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place.", "speaker": "SPEAKER_05"}, {"start": 168.415, "end": 169.296, "text": "That would be amazing.", "speaker": "SPEAKER_05"}, {"start": 169.416, "end": 170.057, "text": "Good idea.", "speaker": "SPEAKER_05"}, {"start": 170.698, "end": 171.198, "text": "Yes, please.", "speaker": "SPEAKER_05"}, {"start": 171.238, "end": 176.445, "text": "I see a lot of people are saying yes.", "speaker": "SPEAKER_05"}, {"start": 176.695, "end": 182.461, "text": " In this chat here, it would be very helpful if you could give your yay or nay to that as well.", "speaker": "SPEAKER_05"}, {"start": 182.581, "end": 185.384, "text": "So we can actually look at that just directly through the chat here.", "speaker": "SPEAKER_05"}, {"start": 185.464, "end": 188.848, "text": "So if you want a second week review, yes.", "speaker": "SPEAKER_05"}, {"start": 189.108, "end": 190.089, "text": "If you don't, no.", "speaker": "SPEAKER_05"}, {"start": 190.71, "end": 192.252, "text": "And then we can go forward.", "speaker": "SPEAKER_05"}, {"start": 193.553, "end": 193.893, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 195.495, "end": 196.096, "text": "Enough of that.", "speaker": "SPEAKER_05"}, {"start": 196.616, "end": 197.898, "text": "I'll start sharing my screen here.", "speaker": "SPEAKER_05"}, {"start": 199.259, "end": 200.28, "text": "Probably my entire screen.", "speaker": "SPEAKER_04"}, {"start": 204.084, "end": 204.505, "text": "Okie dokie.", "speaker": "SPEAKER_05"}, {"start": 204.525, "end": 205.446, "text": "So people should be able to see.", "speaker": "SPEAKER_05"}, {"start": 205.486, "end": 206.567, "text": "Maybe just get rid of...", "speaker": "SPEAKER_05"}, {"start": 207.357, "end": 214.009, "text": " The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise.", "speaker": "SPEAKER_05"}, {"start": 215.011, "end": 227.172, "text": "about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here.", "speaker": "SPEAKER_05"}, {"start": 227.928, "end": 229.152, "text": " So you've got your overview.", "speaker": "SPEAKER_05"}, {"start": 229.935, "end": 232.945, "text": "I've added one or two additional resources.", "speaker": "SPEAKER_05"}, {"start": 232.965, "end": 233.827, "text": "So these actually came up.", "speaker": "SPEAKER_05"}, {"start": 233.908, "end": 235.593, "text": "Andrew shared the first of these.", "speaker": "SPEAKER_05"}, {"start": 235.613, "end": 236.436, "text": "These are all videos.", "speaker": "SPEAKER_05"}, {"start": 237.239, "end": 238.242, "text": "One of them is an active inference", "speaker": "SPEAKER_05"}, {"start": 240.668, "end": 241.791, "text": " coding and active inference.", "speaker": "SPEAKER_05"}, {"start": 243.094, "end": 245.139, "text": "So that's Ryan Smith and his team.", "speaker": "SPEAKER_05"}, {"start": 245.841, "end": 248.227, "text": "They've done excellent work on predictive coding.", "speaker": "SPEAKER_05"}, {"start": 248.708, "end": 249.45, "text": "I'll play that just now.", "speaker": "SPEAKER_05"}, {"start": 249.991, "end": 258.171, "text": "This is a great stream just about predictive coding more generally and its relation to active inference and its relation to", "speaker": "SPEAKER_05"}, {"start": 258.151, "end": 260.855, "text": " you know, its status in actual brains.", "speaker": "SPEAKER_05"}, {"start": 261.616, "end": 265.781, "text": "So if you're interested in more context, this would be excellent to go through.", "speaker": "SPEAKER_05"}, {"start": 265.801, "end": 267.003, "text": "And there's two more links there as well.", "speaker": "SPEAKER_05"}, {"start": 267.043, "end": 271.99, "text": "One by Jakob Howie, he's a philosopher over in Australia in the University of Monash.", "speaker": "SPEAKER_05"}, {"start": 272.01, "end": 274.013, "text": "What is predictive processing and what is it good for?", "speaker": "SPEAKER_05"}, {"start": 274.133, "end": 274.914, "text": "Another excellent talk.", "speaker": "SPEAKER_05"}, {"start": 275.495, "end": 280.041, "text": "And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke.", "speaker": "SPEAKER_05"}, {"start": 280.325, "end": 281.426, "text": " from the University of Toronto.", "speaker": "SPEAKER_05"}, {"start": 281.446, "end": 286.772, "text": "He's actually speaking about Neoplatonism and mystical experiences, believe it or not.", "speaker": "SPEAKER_05"}, {"start": 287.513, "end": 289.135, "text": "And I know it sounds weird.", "speaker": "SPEAKER_05"}, {"start": 289.575, "end": 292.799, "text": "Predictive processing actually shows up in here in a very significant way.", "speaker": "SPEAKER_05"}, {"start": 292.819, "end": 297.084, "text": "So that might be a nice mood setter, as it were.", "speaker": "SPEAKER_05"}, {"start": 297.164, "end": 298.365, "text": "So there's some additional resources there.", "speaker": "SPEAKER_05"}, {"start": 298.686, "end": 301.128, "text": "I've begun filling in the map for chapter five.", "speaker": "SPEAKER_05"}, {"start": 301.949, "end": 307.075, "text": "So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid.", "speaker": "SPEAKER_05"}, {"start": 307.73, "end": 334.446, "text": " the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that", "speaker": "SPEAKER_05"}, {"start": 334.814, "end": 338.84, "text": " A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation.", "speaker": "SPEAKER_05"}, {"start": 339.401, "end": 353.881, "text": "What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you.", "speaker": "SPEAKER_05"}, {"start": 354.322, "end": 359.269, "text": "You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you.", "speaker": "SPEAKER_05"}, {"start": 359.249, "end": 363.355, "text": " So it's not ideal, but the content is there at least.", "speaker": "SPEAKER_05"}, {"start": 363.736, "end": 367.081, "text": "And especially for those who don't have the book, I think this is quite useful.", "speaker": "SPEAKER_05"}, {"start": 367.101, "end": 374.753, "text": "So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter.", "speaker": "SPEAKER_05"}, {"start": 375.274, "end": 382.405, "text": "I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts", "speaker": "SPEAKER_05"}, {"start": 382.975, "end": 403.431, "text": " uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there", "speaker": "SPEAKER_05"}, {"start": 404.373, "end": 404.954, "text": " Any questions?", "speaker": "SPEAKER_05"}, {"start": 404.994, "end": 416.831, "text": "Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that.", "speaker": "SPEAKER_05"}, {"start": 418.033, "end": 419.795, "text": "I'll start my share temporarily.", "speaker": "SPEAKER_05"}, {"start": 420.756, "end": 421.858, "text": "Just coming back.", "speaker": "SPEAKER_05"}, {"start": 423.863, "end": 424.404, "text": " Very good.", "speaker": "SPEAKER_05"}, {"start": 425.745, "end": 426.386, "text": "Yes, but it's limited.", "speaker": "SPEAKER_05"}, {"start": 428.008, "end": 430.991, "text": "Formatting in LaTeX is a bit strange in Coda.", "speaker": "SPEAKER_05"}, {"start": 432.032, "end": 442.664, "text": "We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda.", "speaker": "SPEAKER_05"}, {"start": 444.026, "end": 445.928, "text": "OK, so I'll share, come back.", "speaker": "SPEAKER_05"}, {"start": 447.89, "end": 452.976, "text": "Let's go all the way back, hopefully people can see my screen again, to the introduction.", "speaker": "SPEAKER_05"}, {"start": 454.441, "end": 456.885, "text": " So everyone's seen this figure here.", "speaker": "SPEAKER_05"}, {"start": 456.905, "end": 458.408, "text": "This is essentially the first figure in the book.", "speaker": "SPEAKER_05"}, {"start": 458.849, "end": 461.814, "text": "This is the breakdown of the partitions of the book, so part one, part two, part three.", "speaker": "SPEAKER_05"}, {"start": 462.396, "end": 468.046, "text": "We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding.", "speaker": "SPEAKER_05"}, {"start": 469.649, "end": 471.232, "text": "Part two is active inference core.", "speaker": "SPEAKER_05"}, {"start": 471.252, "end": 473.215, "text": "So this is the heart of the book, really.", "speaker": "SPEAKER_05"}, {"start": 473.255, "end": 476.922, "text": "This is the fundamentals of active inference per se.", "speaker": "SPEAKER_05"}, {"start": 476.902, "end": 498.64, "text": " what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end", "speaker": "SPEAKER_05"}, {"start": 499.328, "end": 504.875, "text": " But as has been the case, everything has been just perception only.", "speaker": "SPEAKER_05"}, {"start": 504.895, "end": 506.057, "text": "So we haven't actually dealt with actions.", "speaker": "SPEAKER_05"}, {"start": 506.437, "end": 508.24, "text": "And we're going to be moving into that with part two.", "speaker": "SPEAKER_05"}, {"start": 509.221, "end": 511.564, "text": "But even more fundamentally, everything has been static.", "speaker": "SPEAKER_05"}, {"start": 511.824, "end": 516.651, "text": "So all of the models we've been looking at, I don't know if I can get a nice example.", "speaker": "SPEAKER_05"}, {"start": 517.592, "end": 519.014, "text": "All of the models, so this is chapter one.", "speaker": "SPEAKER_05"}, {"start": 519.995, "end": 523.159, "text": "Let's maybe go down to some of the equations.", "speaker": "SPEAKER_05"}, {"start": 523.443, "end": 524.244, "text": " They've all been static.", "speaker": "SPEAKER_05"}, {"start": 524.264, "end": 527.387, "text": "So the environment hasn't really been changing at all.", "speaker": "SPEAKER_05"}, {"start": 527.407, "end": 530.371, "text": "And this is very simple.", "speaker": "SPEAKER_05"}, {"start": 531.372, "end": 534.235, "text": "So part one, really going to chapter two, I think.", "speaker": "SPEAKER_05"}, {"start": 536.858, "end": 538.88, "text": "So we have an environment.", "speaker": "SPEAKER_05"}, {"start": 539.501, "end": 541.503, "text": "This is one we've seen this many times now.", "speaker": "SPEAKER_05"}, {"start": 541.523, "end": 548.691, "text": "This is a representation of the generative process, the real process in and of itself out there.", "speaker": "SPEAKER_05"}, {"start": 549.262, "end": 553.407, "text": " This is fairly static, so nothing is really changing.", "speaker": "SPEAKER_05"}, {"start": 553.467, "end": 555.609, "text": "The hidden states aren't really changing from moment to moment.", "speaker": "SPEAKER_05"}, {"start": 555.629, "end": 562.397, "text": "We're just presented with a situation and we have to update our beliefs about what hidden states might be.", "speaker": "SPEAKER_05"}, {"start": 562.437, "end": 564.199, "text": "That's been constant throughout all of Chapter 1.", "speaker": "SPEAKER_05"}, {"start": 564.879, "end": 566.721, "text": "Things have gotten progressively more complicated.", "speaker": "SPEAKER_05"}, {"start": 566.741, "end": 570.906, "text": "We've looked at univariate hidden states where there's literally just one hidden state.", "speaker": "SPEAKER_05"}, {"start": 571.567, "end": 577.193, "text": "We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things.", "speaker": "SPEAKER_05"}, {"start": 577.662, "end": 579.304, "text": " But in every case, it's been static.", "speaker": "SPEAKER_05"}, {"start": 579.884, "end": 585.331, "text": "The dynamics of the hidden state have been non-existent.", "speaker": "SPEAKER_05"}, {"start": 585.351, "end": 586.992, "text": "That is going to change going into part two.", "speaker": "SPEAKER_05"}, {"start": 587.052, "end": 591.918, "text": "We're going to start looking at generative processes which change over time.", "speaker": "SPEAKER_05"}, {"start": 592.579, "end": 593.94, "text": "And they're the interesting ones.", "speaker": "SPEAKER_05"}, {"start": 593.98, "end": 596.283, "text": "They're the ones that are actually useful.", "speaker": "SPEAKER_05"}, {"start": 596.343, "end": 602.55, "text": "And they're the ones that allow us to actually begin to have a reason to act in the environment.", "speaker": "SPEAKER_05"}, {"start": 602.61, "end": 605.593, "text": "We haven't really had any reason to perform actions yet.", "speaker": "SPEAKER_05"}, {"start": 606.805, "end": 611.032, "text": " So that's a big move that's going to take place and that we're going to have to deal with.", "speaker": "SPEAKER_05"}, {"start": 612.074, "end": 613.517, "text": "So maybe just coming back.", "speaker": "SPEAKER_05"}, {"start": 613.557, "end": 618.325, "text": "So chapter one was about, what was this about?", "speaker": "SPEAKER_05"}, {"start": 618.345, "end": 619.187, "text": "It was about perception.", "speaker": "SPEAKER_05"}, {"start": 619.427, "end": 619.588, "text": "Okay.", "speaker": "SPEAKER_05"}, {"start": 619.628, "end": 624.176, "text": "So like in terms of how we're going to think about perception in active inference.", "speaker": "SPEAKER_05"}, {"start": 624.757, "end": 627.261, "text": "So in chapter one, we cast or we framed", "speaker": "SPEAKER_05"}, {"start": 630.565, "end": 654.058, "text": " of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at", "speaker": "SPEAKER_05"}, {"start": 655.118, "end": 661.187, "text": " you know, how to do like exact Bayesian inference, okay, if we were able to do everything.", "speaker": "SPEAKER_05"}, {"start": 663.11, "end": 665.213, "text": "We saw how to use Bayes' rule basically.", "speaker": "SPEAKER_05"}, {"start": 665.233, "end": 675.187, "text": "So we cemented the distinction personally before that around the difference between the generative process and the generative model.", "speaker": "SPEAKER_05"}, {"start": 675.989, "end": 679.354, "text": "So I come all the way down to the figures here.", "speaker": "SPEAKER_05"}, {"start": 679.374, "end": 681.076, "text": "I don't have a particularly good figure for that.", "speaker": "SPEAKER_05"}, {"start": 681.096, "end": 682.358, "text": "Here we go, 2.4.", "speaker": "SPEAKER_05"}, {"start": 682.793, "end": 691.766, "text": " The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world.", "speaker": "SPEAKER_05"}, {"start": 692.307, "end": 695.191, "text": "This was the crucial thing, I think, in chapter two, basically.", "speaker": "SPEAKER_05"}, {"start": 695.992, "end": 699.978, "text": "And we have a way of notating this convention.", "speaker": "SPEAKER_05"}, {"start": 699.998, "end": 708.531, "text": "We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah.", "speaker": "SPEAKER_05"}, {"start": 708.971, "end": 712.176, "text": "These belong to the generative process, the real world out there.", "speaker": "SPEAKER_05"}, {"start": 712.797, "end": 716.822, "text": " Okay, and then modeled variables are without a star.", "speaker": "SPEAKER_05"}, {"start": 717.142, "end": 719.184, "text": "So they correspond to our model.", "speaker": "SPEAKER_05"}, {"start": 720.746, "end": 721.467, "text": "And what is our model?", "speaker": "SPEAKER_05"}, {"start": 722.668, "end": 730.037, "text": "It is quite literally a joint probability distribution over hidden states, latent states, and observations.", "speaker": "SPEAKER_05"}, {"start": 730.978, "end": 733.421, "text": "We use x for latent states and y for observations.", "speaker": "SPEAKER_05"}, {"start": 734.722, "end": 742.551, "text": "But you can see, of course, the x in here is our model of the real latent state outside of here, x star.", "speaker": "SPEAKER_05"}, {"start": 743.172, "end": 744.755, "text": " And they need not agree with each other.", "speaker": "SPEAKER_05"}, {"start": 744.795, "end": 745.737, "text": "They need not be the same.", "speaker": "SPEAKER_05"}, {"start": 746.839, "end": 752.69, "text": "But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be.", "speaker": "SPEAKER_05"}, {"start": 752.731, "end": 754.795, "text": "That is p of x and y.", "speaker": "SPEAKER_05"}, {"start": 755.817, "end": 760.185, "text": "And then if we have the model evidence, the probability of observing anything,", "speaker": "SPEAKER_05"}, {"start": 760.722, "end": 769.453, "text": " or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation.", "speaker": "SPEAKER_05"}, {"start": 770.174, "end": 772.037, "text": "And that was kind of chapter two.", "speaker": "SPEAKER_05"}, {"start": 772.057, "end": 779.526, "text": "And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions.", "speaker": "SPEAKER_05"}, {"start": 779.546, "end": 780.648, "text": "And these are very ubiquitous.", "speaker": "SPEAKER_05"}, {"start": 780.668, "end": 781.369, "text": "They show up everywhere.", "speaker": "SPEAKER_05"}, {"start": 782.43, "end": 783.331, "text": "We're going to continue that.", "speaker": "SPEAKER_05"}, {"start": 783.351, "end": 790.24, "text": "In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference.", "speaker": "SPEAKER_05"}, {"start": 790.642, "end": 791.784, "text": " There are other distributions as well.", "speaker": "SPEAKER_05"}, {"start": 791.804, "end": 792.725, "text": "We haven't really looked at those yet.", "speaker": "SPEAKER_05"}, {"start": 793.967, "end": 799.554, "text": "But the real sort of, there's a problem with this, which is that, yes, we can do inference.", "speaker": "SPEAKER_05"}, {"start": 799.574, "end": 803.079, "text": "Yes, we might even be able to do exact inference for really, really simple problems.", "speaker": "SPEAKER_05"}, {"start": 804.14, "end": 812.952, "text": "But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model.", "speaker": "SPEAKER_05"}, {"start": 813.793, "end": 815.075, "text": "And then we're going to do inference.", "speaker": "SPEAKER_05"}, {"start": 815.095, "end": 820.142, "text": "And inference is like, I set the dials, and then my machine does some inference stuff, right?", "speaker": "SPEAKER_05"}, {"start": 820.983, "end": 847.716, "text": " that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation", "speaker": "SPEAKER_05"}, {"start": 848.303, "end": 850.966, "text": " And we're doing good old fashioned Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 851.346, "end": 854.99, "text": "But now we're having to do parameter learning and estimation as well.", "speaker": "SPEAKER_05"}, {"start": 856.692, "end": 867.903, "text": "And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference.", "speaker": "SPEAKER_05"}, {"start": 867.923, "end": 872.448, "text": "So we have these two processes, inference and learning parameter estimation.", "speaker": "SPEAKER_05"}, {"start": 873.109, "end": 877.213, "text": "And they're both able to be done by means of Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 878.56, "end": 879.421, "text": " That's the crucial thing.", "speaker": "SPEAKER_05"}, {"start": 879.441, "end": 889.438, "text": "So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah.", "speaker": "SPEAKER_05"}, {"start": 889.458, "end": 904.163, "text": "And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent.", "speaker": "SPEAKER_05"}, {"start": 904.868, "end": 908.953, "text": " And we saw many versions of this across the chapter and indeed in other chapters.", "speaker": "SPEAKER_05"}, {"start": 908.993, "end": 920.085, "text": "It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters.", "speaker": "SPEAKER_05"}, {"start": 920.185, "end": 926.172, "text": "So let's imagine we have our two parameters, I don't know, beta 1, beta 0.", "speaker": "SPEAKER_05"}, {"start": 926.232, "end": 934.521, "text": "And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space.", "speaker": "SPEAKER_05"}, {"start": 934.872, "end": 935.733, "text": " That's this surface.", "speaker": "SPEAKER_05"}, {"start": 936.754, "end": 944.063, "text": "And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point.", "speaker": "SPEAKER_05"}, {"start": 945.064, "end": 950.271, "text": "And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters.", "speaker": "SPEAKER_05"}, {"start": 950.291, "end": 952.654, "text": "Very general way of doing optimization.", "speaker": "SPEAKER_05"}, {"start": 954.015, "end": 964.288, "text": "However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be.", "speaker": "SPEAKER_05"}, {"start": 965.72, "end": 973.231, "text": " And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization.", "speaker": "SPEAKER_05"}, {"start": 975.574, "end": 985.308, "text": "So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now.", "speaker": "SPEAKER_05"}, {"start": 986.25, "end": 993.921, "text": "A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding.", "speaker": "SPEAKER_05"}, {"start": 994.357, "end": 1021.09, "text": " um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model", "speaker": "SPEAKER_05"}, {"start": 1021.525, "end": 1025.652, "text": " And also, how should I use those settings to do good inference?", "speaker": "SPEAKER_05"}, {"start": 1026.934, "end": 1040.558, "text": "It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters.", "speaker": "SPEAKER_05"}, {"start": 1040.578, "end": 1042.442, "text": "So imagine you've got learning up here.", "speaker": "SPEAKER_05"}, {"start": 1042.582, "end": 1043.744, "text": "What should the model parameters be?", "speaker": "SPEAKER_05"}, {"start": 1043.764, "end": 1046.048, "text": "And you can do inference on model parameters.", "speaker": "SPEAKER_05"}, {"start": 1046.45, "end": 1050.538, "text": " And then that can inform how you do inference at the low-level hidden states.", "speaker": "SPEAKER_05"}, {"start": 1050.558, "end": 1053.805, "text": "So there's kind of these two processes that are happening in two different timescales.", "speaker": "SPEAKER_05"}, {"start": 1053.825, "end": 1062.062, "text": "That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally.", "speaker": "SPEAKER_05"}, {"start": 1063.122, "end": 1064.023, "text": " So that's kind of chapter three.", "speaker": "SPEAKER_05"}, {"start": 1064.143, "end": 1065.264, "text": "I'm going to race through this.", "speaker": "SPEAKER_05"}, {"start": 1065.925, "end": 1068.828, "text": "And we'll get to chapter five and then to some actual questions from you guys.", "speaker": "SPEAKER_05"}, {"start": 1069.749, "end": 1081.062, "text": "Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference.", "speaker": "SPEAKER_05"}, {"start": 1081.082, "end": 1084.966, "text": "So chapter three, we're still doing exact inference.", "speaker": "SPEAKER_05"}, {"start": 1084.986, "end": 1090.632, "text": "But we saw that really we culminated in an algorithm", "speaker": "SPEAKER_05"}, {"start": 1091.658, "end": 1113.632, "text": " allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be", "speaker": "SPEAKER_05"}, {"start": 1113.95, "end": 1115.672, "text": " Those two problems are related to one another.", "speaker": "SPEAKER_05"}, {"start": 1115.752, "end": 1118.615, "text": "To know the hidden states, you need to know what the good parameters are.", "speaker": "SPEAKER_05"}, {"start": 1118.935, "end": 1121.378, "text": "But to know what the good parameters are, you have to know how to infer hidden states well.", "speaker": "SPEAKER_05"}, {"start": 1121.958, "end": 1131.588, "text": "So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm.", "speaker": "SPEAKER_05"}, {"start": 1131.608, "end": 1132.689, "text": "I won't go over that again here.", "speaker": "SPEAKER_05"}, {"start": 1132.73, "end": 1134.131, "text": "Maybe we will if people want me to.", "speaker": "SPEAKER_05"}, {"start": 1135.072, "end": 1139.957, "text": "But that was the solution to the problem of chapter 3, that chicken and egg problem.", "speaker": "SPEAKER_05"}, {"start": 1140.781, "end": 1151.908, "text": " The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations.", "speaker": "SPEAKER_05"}, {"start": 1151.928, "end": 1155.938, "text": "And that, unfortunately, is usually not something we can do.", "speaker": "SPEAKER_05"}, {"start": 1156.542, "end": 1161.508, "text": " Being able to find the exact posterior is usually impossible, for reasons that we saw.", "speaker": "SPEAKER_05"}, {"start": 1161.708, "end": 1162.75, "text": "We can review again if you want.", "speaker": "SPEAKER_05"}, {"start": 1163.33, "end": 1166.895, "text": "So then chapter 4 says, all right, well, we can't do exact Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 1167.536, "end": 1168.096, "text": "What are we going to do?", "speaker": "SPEAKER_05"}, {"start": 1168.116, "end": 1177.107, "text": "We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like.", "speaker": "SPEAKER_05"}, {"start": 1178.029, "end": 1180.812, "text": "So the idea with chapter 4 is, ah, OK, we're going to now", "speaker": "SPEAKER_05"}, {"start": 1181.804, "end": 1193.779, "text": " have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior.", "speaker": "SPEAKER_05"}, {"start": 1194.7, "end": 1198.304, "text": "We're going to try and find a posterior that's close enough to the true posterior.", "speaker": "SPEAKER_05"}, {"start": 1198.705, "end": 1208.537, "text": "And this then motivated the idea of variational free energy as a quantity that is a measurement of how well", "speaker": "SPEAKER_05"}, {"start": 1208.838, "end": 1213.784, "text": " an approximate posterior fits or how close it is to the true posterior.", "speaker": "SPEAKER_05"}, {"start": 1214.885, "end": 1222.794, "text": "And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal.", "speaker": "SPEAKER_05"}, {"start": 1223.055, "end": 1225.858, "text": "We saw from chapter three that surprisal is something we want to make very small.", "speaker": "SPEAKER_05"}, {"start": 1226.719, "end": 1237.191, "text": "So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal,", "speaker": "SPEAKER_05"}, {"start": 1237.778, "end": 1242.586, "text": " which is something we can't do directly, and therefore we can do approximate Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 1242.606, "end": 1248.976, "text": "So in a lot of ways, that really kind of spiritually is the end of part one, I would say.", "speaker": "SPEAKER_05"}, {"start": 1249.958, "end": 1265.102, "text": "The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE.", "speaker": "SPEAKER_05"}, {"start": 1265.842, "end": 1269.509, "text": " That's, in my mind, really kind of the end of part one.", "speaker": "SPEAKER_05"}, {"start": 1270.311, "end": 1278.908, "text": "Chapter five was a nice kind of detour into a different way of thinking about what the VFE is.", "speaker": "SPEAKER_05"}, {"start": 1280.451, "end": 1287.445, "text": "So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 1288.387, "end": 1314.799, "text": " we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical", "speaker": "SPEAKER_05"}, {"start": 1314.965, "end": 1334.813, "text": " know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was", "speaker": "SPEAKER_05"}, {"start": 1335.535, "end": 1349.242, "text": " lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning.", "speaker": "SPEAKER_05"}, {"start": 1350.083, "end": 1357.257, "text": "It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization", "speaker": "SPEAKER_05"}, {"start": 1357.574, "end": 1369.356, "text": " in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure", "speaker": "SPEAKER_05"}, {"start": 1370.163, "end": 1387.641, "text": " And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow.", "speaker": "SPEAKER_05"}, {"start": 1388.962, "end": 1391.865, "text": "And that's kind of where we left off with part one.", "speaker": "SPEAKER_05"}, {"start": 1391.885, "end": 1396.39, "text": "So that is, in effect, what we've done, where we've gone.", "speaker": "SPEAKER_05"}, {"start": 1396.943, "end": 1400.268, "text": " That's a lot of stuff to cover, but again, we haven't looked at actions yet.", "speaker": "SPEAKER_05"}, {"start": 1400.829, "end": 1407.12, "text": "We're gonna be doing that in part two, and that will bring us full circle to active inference.", "speaker": "SPEAKER_05"}, {"start": 1407.14, "end": 1420.021, "text": "And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception.", "speaker": "SPEAKER_05"}, {"start": 1420.922, "end": 1423.086, "text": "Okay, we're gonna see that these are deeply related to each other.", "speaker": "SPEAKER_05"}, {"start": 1423.623, "end": 1426.045, "text": " But that's the ground that we've covered.", "speaker": "SPEAKER_05"}, {"start": 1426.846, "end": 1429.829, "text": "I'd be interested if people have questions related to that.", "speaker": "SPEAKER_05"}, {"start": 1429.849, "end": 1431.29, "text": "I see that there's lots of questions.", "speaker": "SPEAKER_05"}, {"start": 1432.371, "end": 1432.912, "text": "Stop sharing.", "speaker": "SPEAKER_05"}, {"start": 1434.533, "end": 1437.476, "text": "And I see Mark has a question.", "speaker": "SPEAKER_05"}, {"start": 1437.496, "end": 1438.397, "text": "Please fire away, Mark.", "speaker": "SPEAKER_05"}, {"start": 1440.039, "end": 1440.519, "text": "As usual.", "speaker": "SPEAKER_02"}, {"start": 1441.16, "end": 1443.041, "text": "Thank you so much for this review.", "speaker": "SPEAKER_02"}, {"start": 1443.061, "end": 1443.602, "text": "This is great.", "speaker": "SPEAKER_02"}, {"start": 1445.724, "end": 1451.389, "text": "I was wondering if you could pull up that slide that showed the generative process and the generative model.", "speaker": "SPEAKER_02"}, {"start": 1451.429, "end": 1453.191, "text": "That might be helpful as a reference.", "speaker": "SPEAKER_02"}, {"start": 1453.964, "end": 1455.166, "text": " Oh, yeah, okay.", "speaker": "SPEAKER_05"}, {"start": 1455.186, "end": 1456.187, "text": "So let me come back here.", "speaker": "SPEAKER_05"}, {"start": 1458.43, "end": 1462.415, "text": "That's the one in chapter two, I assume you're referring to?", "speaker": "SPEAKER_05"}, {"start": 1462.435, "end": 1462.836, "text": "I think so.", "speaker": "SPEAKER_02"}, {"start": 1463.697, "end": 1464.138, "text": "This one here?", "speaker": "SPEAKER_05"}, {"start": 1464.618, "end": 1465.479, "text": "Yes, thank you.", "speaker": "SPEAKER_02"}, {"start": 1466.1, "end": 1470.126, "text": "Yeah, I really just kind of appreciate what this textbook is trying to do.", "speaker": "SPEAKER_02"}, {"start": 1470.146, "end": 1473.57, "text": "And I think it took me a while to appreciate that.", "speaker": "SPEAKER_02"}, {"start": 1473.69, "end": 1482.402, "text": "But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening.", "speaker": "SPEAKER_02"}, {"start": 1482.382, "end": 1505.689, "text": " in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise", "speaker": "SPEAKER_02"}, {"start": 1505.905, "end": 1532.082, "text": " in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of", "speaker": "SPEAKER_05"}, {"start": 1532.973, "end": 1538.437, "text": " It seems like chapter four kind of serves as a background where the story gets started for the most part.", "speaker": "SPEAKER_02"}, {"start": 1538.458, "end": 1542.375, "text": "And this is kind of more of a ground level", "speaker": "SPEAKER_02"}, {"start": 1542.457, "end": 1544.78, "text": " view, which is nice and helpful.", "speaker": "SPEAKER_02"}, {"start": 1546.603, "end": 1554.433, "text": "What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function.", "speaker": "SPEAKER_02"}, {"start": 1554.453, "end": 1559.68, "text": "Because in my mind, I'm kind of conditioned to seeing these Gaussians, right?", "speaker": "SPEAKER_02"}, {"start": 1559.7, "end": 1566.089, "text": "You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior.", "speaker": "SPEAKER_02"}, {"start": 1566.75, "end": 1571.316, "text": "However, here we have this linear function, which I guess is", "speaker": "SPEAKER_02"}, {"start": 1572.207, "end": 1575.451, "text": " then becomes part of, say, a likelihood.", "speaker": "SPEAKER_02"}, {"start": 1575.491, "end": 1588.409, "text": "And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense.", "speaker": "SPEAKER_02"}, {"start": 1588.429, "end": 1588.83, "text": "Yeah, yeah.", "speaker": "SPEAKER_05"}, {"start": 1589.651, "end": 1590.212, "text": "No, absolutely.", "speaker": "SPEAKER_05"}, {"start": 1590.252, "end": 1600.746, "text": "I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1601.131, "end": 1628.91, "text": " um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so", "speaker": "SPEAKER_05"}, {"start": 1629.464, "end": 1638.597, "text": " This is, in some sense, still our assumption about what the relationship is between observations and the hidden states.", "speaker": "SPEAKER_05"}, {"start": 1639.066, "end": 1646.86, "text": " For our purpose, we've assumed that the real world, we're sort of creating a world here.", "speaker": "SPEAKER_05"}, {"start": 1646.94, "end": 1651.348, "text": "This is the example of the food size and light intensity world.", "speaker": "SPEAKER_05"}, {"start": 1651.468, "end": 1653.572, "text": "This is the really simple world that we've seen for all of part one.", "speaker": "SPEAKER_05"}, {"start": 1654.153, "end": 1655.475, "text": "So what are the hidden states?", "speaker": "SPEAKER_05"}, {"start": 1656.076, "end": 1658.04, "text": "They are the sizes of the food.", "speaker": "SPEAKER_05"}, {"start": 1658.46, "end": 1661.666, "text": "And we can see that there's five different sizes, one to five.", "speaker": "SPEAKER_05"}, {"start": 1661.907, "end": 1662.728, "text": "That's the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1663.552, "end": 1665.214, "text": " And the observations are light intensities.", "speaker": "SPEAKER_05"}, {"start": 1665.354, "end": 1668.318, "text": "So light comes in, hits the food, and you get some sort of light intensity.", "speaker": "SPEAKER_05"}, {"start": 1668.758, "end": 1672.903, "text": "And depending on the size of the food, you get some specific light intensity.", "speaker": "SPEAKER_05"}, {"start": 1672.924, "end": 1685.619, "text": "So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size.", "speaker": "SPEAKER_05"}, {"start": 1685.639, "end": 1688.783, "text": "So we're saying that the relationship between the observation you get in", "speaker": "SPEAKER_05"}, {"start": 1689.235, "end": 1709.573, "text": " is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters", "speaker": "SPEAKER_05"}, {"start": 1710.548, "end": 1714.279, "text": " So again, this is the environment regenerative process.", "speaker": "SPEAKER_05"}, {"start": 1714.941, "end": 1721.942, "text": "The corresponding model is this one, 2.11, equation 2.11, I believe.", "speaker": "SPEAKER_05"}, {"start": 1722.513, "end": 1751.187, "text": " now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs", "speaker": "SPEAKER_05"}, {"start": 1751.555, "end": 1762.486, "text": " are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution.", "speaker": "SPEAKER_05"}, {"start": 1762.526, "end": 1780.684, "text": "So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations.", "speaker": "SPEAKER_05"}, {"start": 1781.204, "end": 1783.246, "text": " And the thing is, we're uncertain.", "speaker": "SPEAKER_05"}, {"start": 1784.268, "end": 1789.194, "text": "So the agent doesn't get to see what the real relationship is between food size and light intensity.", "speaker": "SPEAKER_05"}, {"start": 1789.334, "end": 1789.995, "text": "It has no idea.", "speaker": "SPEAKER_05"}, {"start": 1791.136, "end": 1793.679, "text": "It has a model about what it thinks the relationship is.", "speaker": "SPEAKER_05"}, {"start": 1794.24, "end": 1796.443, "text": "Now, we know that the relationship is linear.", "speaker": "SPEAKER_05"}, {"start": 1796.483, "end": 1805.994, "text": "We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity.", "speaker": "SPEAKER_05"}, {"start": 1806.014, "end": 1809.118, "text": "But in the real world, we have no idea what the relationship is.", "speaker": "SPEAKER_05"}, {"start": 1809.385, "end": 1813.791, "text": " It turns out that in this example, we've specified exactly the same relationship.", "speaker": "SPEAKER_05"}, {"start": 1814.752, "end": 1816.975, "text": "So that's quite a nice scenario to be in.", "speaker": "SPEAKER_05"}, {"start": 1817.676, "end": 1826.968, "text": "You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size.", "speaker": "SPEAKER_05"}, {"start": 1828.429, "end": 1834.397, "text": "Now that's wrong, but it's OK, maybe, for some settings.", "speaker": "SPEAKER_05"}, {"start": 1834.985, "end": 1843.699, "text": " That would be an example of a likelihood mapping that would be incorrect with respect to its generating function.", "speaker": "SPEAKER_05"}, {"start": 1844.721, "end": 1861.448, "text": "So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount.", "speaker": "SPEAKER_05"}, {"start": 1862.103, "end": 1869.693, "text": " And it's centered about the relationship we think is the case between the hidden states and the observations.", "speaker": "SPEAKER_05"}, {"start": 1870.234, "end": 1873.658, "text": "So the reason why we have this probability distribution is because we don't know.", "speaker": "SPEAKER_05"}, {"start": 1873.678, "end": 1877.924, "text": "We do not actually ever get to know what the real relationship is between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1877.944, "end": 1881.208, "text": "So does that answer your question, Mark, about why", "speaker": "SPEAKER_05"}, {"start": 1881.458, "end": 1887.967, "text": " why probability distributions, maybe why normal distributions, or did that help at all?", "speaker": "SPEAKER_05"}, {"start": 1888.007, "end": 1889.469, "text": "Yeah, that helped a lot.", "speaker": "SPEAKER_02"}, {"start": 1890.511, "end": 1890.951, "text": "Thank you.", "speaker": "SPEAKER_02"}, {"start": 1891.011, "end": 1900.164, "text": "Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right?", "speaker": "SPEAKER_02"}, {"start": 1900.464, "end": 1908.175, "text": "And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right?", "speaker": "SPEAKER_02"}, {"start": 1909.235, "end": 1934.061, "text": " uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all", "speaker": "SPEAKER_05"}, {"start": 1934.648, "end": 1938.613, "text": " We could plug, we could say, okay, the relationship is just equal to the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1938.813, "end": 1942.738, "text": "I think that the observation is exactly the hidden state, right?", "speaker": "SPEAKER_05"}, {"start": 1942.758, "end": 1943.438, "text": "That's an assumption.", "speaker": "SPEAKER_05"}, {"start": 1944.86, "end": 1952.529, "text": "Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1953.49, "end": 1955.653, "text": "Or we can have a quadratic relationship.", "speaker": "SPEAKER_05"}, {"start": 1955.673, "end": 1957.635, "text": "We can have anything we want in there at all, actually.", "speaker": "SPEAKER_05"}, {"start": 1957.856, "end": 1963.843, "text": "And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1964.083, "end": 1964.183, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 1966.036, "end": 1983.758, "text": " very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B,", "speaker": "SPEAKER_04"}, {"start": 1986.978, "end": 2006.903, "text": " Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from.", "speaker": "SPEAKER_05"}, {"start": 2006.923, "end": 2009.526, "text": "Michael Morehead , Other questions other questions Magdalena.", "speaker": "SPEAKER_05"}, {"start": 2011.582, "end": 2015.987, "text": " Okay, my question is this.", "speaker": "SPEAKER_01"}, {"start": 2016.608, "end": 2019.672, "text": "So could we do a thought experiment for a minute?", "speaker": "SPEAKER_01"}, {"start": 2019.712, "end": 2025.319, "text": "Because I think if I set it up as a thought experiment, I'll get the answer.", "speaker": "SPEAKER_01"}, {"start": 2026.821, "end": 2027.121, "text": "Okay.", "speaker": "SPEAKER_01"}, {"start": 2027.221, "end": 2027.942, "text": "Oh, so I'm inside.", "speaker": "SPEAKER_01"}, {"start": 2027.962, "end": 2030.966, "text": "Okay, so let's imagine the following.", "speaker": "SPEAKER_01"}, {"start": 2031.907, "end": 2040.357, "text": "In present-day societies, we have this societal", "speaker": "SPEAKER_01"}, {"start": 2040.775, "end": 2050.769, "text": " back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists.", "speaker": "SPEAKER_01"}, {"start": 2052.391, "end": 2060.623, "text": "And there's a tension in societies to update that belief by some group to science.", "speaker": "SPEAKER_01"}, {"start": 2060.89, "end": 2063.737, "text": " Exists a science so God exists.", "speaker": "SPEAKER_01"}, {"start": 2063.837, "end": 2071.114, "text": "So science tells us that God may not exist or something to that effect or quantum physics explains reality.", "speaker": "SPEAKER_01"}, {"start": 2071.214, "end": 2071.735, "text": "Not God.", "speaker": "SPEAKER_01"}, {"start": 2072.016, "end": 2072.597, "text": "Okay.", "speaker": "SPEAKER_01"}, {"start": 2072.617, "end": 2074.682, "text": "So there's a process, right?", "speaker": "SPEAKER_01"}, {"start": 2074.782, "end": 2077.228, "text": "That's in in the environment.", "speaker": "SPEAKER_01"}, {"start": 2077.208, "end": 2096.28, "text": " that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed", "speaker": "SPEAKER_01"}, {"start": 2096.733, "end": 2118.994, "text": " And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications.", "speaker": "SPEAKER_01"}, {"start": 2118.974, "end": 2148.88, "text": " for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is", "speaker": "SPEAKER_01"}, {"start": 2148.86, "end": 2163.055, "text": " Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis?", "speaker": "SPEAKER_01"}, {"start": 2165.296, "end": 2191.017, "text": " are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds", "speaker": "SPEAKER_01"}, {"start": 2191.25, "end": 2195.815, "text": " Okay, as a third level of information processing.", "speaker": "SPEAKER_01"}, {"start": 2196.296, "end": 2217.46, "text": "And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis?", "speaker": "SPEAKER_01"}, {"start": 2217.44, "end": 2219.723, "text": " Yeah, no, excellent, excellent question, Magdalena.", "speaker": "SPEAKER_05"}, {"start": 2219.964, "end": 2242.857, "text": "This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents.", "speaker": "SPEAKER_05"}, {"start": 2243.748, "end": 2250.721, "text": " Under what conditions do they interact so as to form a larger composite active inference agent?", "speaker": "SPEAKER_05"}, {"start": 2251.462, "end": 2257.513, "text": "So it's kind of a question like, how many active inference agents are there really in any given picture?", "speaker": "SPEAKER_05"}, {"start": 2260.018, "end": 2266.87, "text": "So your question, I think, insofar as I understand it, is, look, we have some notion about agency.", "speaker": "SPEAKER_05"}, {"start": 2267.423, "end": 2268.825, "text": " inactive inference.", "speaker": "SPEAKER_05"}, {"start": 2268.845, "end": 2274.433, "text": "We have some notion about something interacting with its environment, whatever the environment is, right?", "speaker": "SPEAKER_05"}, {"start": 2274.453, "end": 2278.098, "text": "So on the right hand side, we've got an agent, and on the left hand side, we have an environment.", "speaker": "SPEAKER_05"}, {"start": 2279.36, "end": 2293.699, "text": "But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent.", "speaker": "SPEAKER_05"}, {"start": 2293.84, "end": 2295.642, "text": "Okay, we haven't been able to tell that story yet.", "speaker": "SPEAKER_05"}, {"start": 2296.837, "end": 2311.04, "text": " Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers.", "speaker": "SPEAKER_05"}, {"start": 2311.12, "end": 2314.385, "text": "Let me see if I can find that figure.", "speaker": "SPEAKER_05"}, {"start": 2315.206, "end": 2317.83, "text": "Various layers within the predictive coding hierarchy.", "speaker": "SPEAKER_05"}, {"start": 2318.672, "end": 2319.433, "text": "Which one is it here?", "speaker": "SPEAKER_05"}, {"start": 2319.473, "end": 2320.074, "text": "This one?", "speaker": "SPEAKER_05"}, {"start": 2320.354, "end": 2320.735, "text": "No, this one.", "speaker": "SPEAKER_05"}, {"start": 2322.487, "end": 2329.997, "text": " You might interpret the various layers here as individual active implementations that are just doing local free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 2330.017, "end": 2333.682, "text": "So maybe this guy is an active implementation in some sense.", "speaker": "SPEAKER_05"}, {"start": 2333.702, "end": 2335.825, "text": "And then he's passing messages, blah, blah, blah.", "speaker": "SPEAKER_05"}, {"start": 2336.445, "end": 2338.989, "text": "And then the overall thing can be regarded as an active implementation.", "speaker": "SPEAKER_05"}, {"start": 2339.029, "end": 2342.393, "text": "That's the only inkling that we've seen thus far of this idea yet.", "speaker": "SPEAKER_05"}, {"start": 2343.234, "end": 2347.6, "text": "We're also not really going to see a lot of it in the rest of the book.", "speaker": "SPEAKER_05"}, {"start": 2348.052, "end": 2349.956, "text": " because it is a very open question.", "speaker": "SPEAKER_05"}, {"start": 2350.017, "end": 2351.38, "text": "It's a very difficult question.", "speaker": "SPEAKER_05"}, {"start": 2352.302, "end": 2360.36, "text": "And it gets at the heart of what agency even is at all, which is a question over and above active inference per se.", "speaker": "SPEAKER_05"}, {"start": 2360.781, "end": 2367.276, "text": "However, my personal interpretation or my personal thoughts about this issue", "speaker": "SPEAKER_05"}, {"start": 2367.678, "end": 2374.214, "text": " is that the thing that is currently missing from active inference is exactly this idea of compositionality.", "speaker": "SPEAKER_05"}, {"start": 2374.234, "end": 2375.938, "text": "So I've got agent here, agent here.", "speaker": "SPEAKER_05"}, {"start": 2376.219, "end": 2376.841, "text": "They interact.", "speaker": "SPEAKER_05"}, {"start": 2377.482, "end": 2381.873, "text": "Under what conditions is that whole thing one active inference agent?", "speaker": "SPEAKER_05"}, {"start": 2382.275, "end": 2388.688, "text": " That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research.", "speaker": "SPEAKER_05"}, {"start": 2388.709, "end": 2390.352, "text": "So I hope that answers your question in some respect.", "speaker": "SPEAKER_05"}, {"start": 2390.532, "end": 2391.494, "text": "Very, very important question.", "speaker": "SPEAKER_05"}, {"start": 2392.015, "end": 2397.828, "text": "But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding.", "speaker": "SPEAKER_05"}, {"start": 2398.81, "end": 2399.912, "text": "OK.", "speaker": "SPEAKER_01"}, {"start": 2400.23, "end": 2425.562, "text": " really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups", "speaker": "SPEAKER_01"}, {"start": 2425.542, "end": 2430.369, "text": " What I see and I'm going to use this language, I'm not a mathematician.", "speaker": "SPEAKER_01"}, {"start": 2430.429, "end": 2435.577, "text": "OK, so forgive me, but but I kind of I kind of get some things in math.", "speaker": "SPEAKER_01"}, {"start": 2435.597, "end": 2442.167, "text": "OK, so topologically generative models really work as huge attractors.", "speaker": "SPEAKER_01"}, {"start": 2442.147, "end": 2444.451, "text": " in informational systems in humans.", "speaker": "SPEAKER_01"}, {"start": 2444.912, "end": 2458.638, "text": "So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right?", "speaker": "SPEAKER_01"}, {"start": 2458.718, "end": 2462.265, "text": "So your language is a technology in human systems.", "speaker": "SPEAKER_01"}, {"start": 2462.245, "end": 2490.784, "text": " so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model", "speaker": "SPEAKER_01"}, {"start": 2491.692, "end": 2504.479, "text": " And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math.", "speaker": "SPEAKER_01"}, {"start": 2505.702, "end": 2509.55, "text": "If I just kind of sorry, is that okay?", "speaker": "SPEAKER_09"}, {"start": 2509.63, "end": 2509.951, "text": "All right.", "speaker": "SPEAKER_09"}, {"start": 2510.091, "end": 2510.632, "text": "Thanks.", "speaker": "SPEAKER_09"}, {"start": 2510.652, "end": 2512.536, "text": "Just ask someone who has a.", "speaker": "SPEAKER_09"}, {"start": 2512.516, "end": 2515.059, "text": " More immediate background in this social sciences.", "speaker": "SPEAKER_09"}, {"start": 2515.72, "end": 2516.001, "text": "Yeah.", "speaker": "SPEAKER_09"}, {"start": 2516.141, "end": 2516.822, "text": "So, so.", "speaker": "SPEAKER_09"}, {"start": 2517.323, "end": 2521.729, "text": "And I like how the question that Mark had prior to this had to do with, like.", "speaker": "SPEAKER_09"}, {"start": 2522.33, "end": 2524.332, "text": "You know, what is it to make linear assumptions?", "speaker": "SPEAKER_09"}, {"start": 2524.853, "end": 2527.777, "text": "So, in the context of the textbook, what we've seen thus far, like.", "speaker": "SPEAKER_09"}, {"start": 2528.278, "end": 2535.908, "text": "This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise.", "speaker": "SPEAKER_09"}, {"start": 2535.948, "end": 2539.774, "text": "And it's just to show how a model can be composed.", "speaker": "SPEAKER_09"}, {"start": 2540.755, "end": 2542.337, "text": "Like, how can.", "speaker": "SPEAKER_09"}, {"start": 2542.57, "end": 2551.986, "text": " In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right?", "speaker": "SPEAKER_09"}, {"start": 2552.407, "end": 2560.16, "text": "It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right?", "speaker": "SPEAKER_09"}, {"start": 2560.18, "end": 2561.422, "text": "So that's the action part.", "speaker": "SPEAKER_09"}, {"start": 2561.783, "end": 2563.245, "text": "We haven't reached that yet.", "speaker": "SPEAKER_09"}, {"start": 2563.225, "end": 2572.756, "text": " But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function?", "speaker": "SPEAKER_09"}, {"start": 2573.317, "end": 2590.597, "text": "So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right?", "speaker": "SPEAKER_09"}, {"start": 2591.198, "end": 2593.08, "text": "So that linear model we were looking at", "speaker": "SPEAKER_09"}, {"start": 2593.06, "end": 2616.688, "text": " we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people", "speaker": "SPEAKER_09"}, {"start": 2616.938, "end": 2619.842, "text": " what are the kinds of sensory information that I receive?", "speaker": "SPEAKER_09"}, {"start": 2620.522, "end": 2622.965, "text": "And what are the kinds of actions that I can do?", "speaker": "SPEAKER_09"}, {"start": 2623.005, "end": 2625.308, "text": "And what do I have beliefs about?", "speaker": "SPEAKER_09"}, {"start": 2625.408, "end": 2629.173, "text": "Or what variables do I think explains everything, right?", "speaker": "SPEAKER_09"}, {"start": 2629.774, "end": 2640.887, "text": "So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human", "speaker": "SPEAKER_09"}, {"start": 2640.867, "end": 2655.518, "text": " uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like", "speaker": "SPEAKER_09"}, {"start": 2656.021, "end": 2658.344, "text": " There's, there's a lot to track.", "speaker": "SPEAKER_09"}, {"start": 2658.364, "end": 2662.21, "text": "So there are some people who have tried to model this sort of thing.", "speaker": "SPEAKER_09"}, {"start": 2662.23, "end": 2663.572, "text": "And so I've shared this paper.", "speaker": "SPEAKER_09"}, {"start": 2664.313, "end": 2672.665, "text": "It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference.", "speaker": "SPEAKER_09"}, {"start": 2673.226, "end": 2676.19, "text": "And this kind of has to do with.", "speaker": "SPEAKER_09"}, {"start": 2676.507, "end": 2693.875, "text": " Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere.", "speaker": "SPEAKER_09"}, {"start": 2693.855, "end": 2711.429, "text": " And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing.", "speaker": "SPEAKER_09"}, {"start": 2711.489, "end": 2714.715, "text": "And that's one way of trying to look at a problem.", "speaker": "SPEAKER_09"}, {"start": 2714.695, "end": 2734.948, "text": " like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one", "speaker": "SPEAKER_09"}, {"start": 2735.333, "end": 2737.86, "text": " And a lot of agents who rally around idea two.", "speaker": "SPEAKER_09"}, {"start": 2737.9, "end": 2743.315, "text": "And as they do that, the two groups start to communicate directly with each other less and less.", "speaker": "SPEAKER_09"}, {"start": 2743.997, "end": 2749.613, "text": "The agent's actual actions that are simulated has to do with them communicating with one another and sharing their", "speaker": "SPEAKER_09"}, {"start": 2750.015, "end": 2753.178, "text": " They're kind of more verbal beliefs with one another.", "speaker": "SPEAKER_09"}, {"start": 2753.199, "end": 2764.771, "text": "And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing.", "speaker": "SPEAKER_09"}, {"start": 2764.791, "end": 2769.256, "text": "And somehow that gets all blended in with our assumptions of how science works and things like that.", "speaker": "SPEAKER_09"}, {"start": 2769.757, "end": 2771.338, "text": "It's much more complex.", "speaker": "SPEAKER_09"}, {"start": 2771.479, "end": 2779.988, "text": "And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy", "speaker": "SPEAKER_09"}, {"start": 2779.968, "end": 2783.633, "text": " But remember that free energy, based on how we've discussed it so far,", "speaker": "SPEAKER_09"}, {"start": 2783.967, "end": 2787.932, "text": " It's not some magical quantity of energy or something like that.", "speaker": "SPEAKER_09"}, {"start": 2788.914, "end": 2790.736, "text": "It's really just a proxy.", "speaker": "SPEAKER_09"}, {"start": 2790.796, "end": 2797.485, "text": "It's a word that is a proxy for something called surprisal, which is more or less uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2797.625, "end": 2799.608, "text": "So we want to minimize our uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2800.169, "end": 2813.687, "text": "So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another,", "speaker": "SPEAKER_09"}, {"start": 2813.667, "end": 2817.235, "text": " about is it idea one or is it idea two?", "speaker": "SPEAKER_09"}, {"start": 2817.716, "end": 2819.781, "text": "Is it science or is it something else?", "speaker": "SPEAKER_09"}, {"start": 2820.443, "end": 2824.993, "text": "One way to minimize uncertainty is to repeatedly tell each other", "speaker": "SPEAKER_09"}, {"start": 2825.328, "end": 2828.993, "text": " what the truth is and you all converge on what you all think the truth is.", "speaker": "SPEAKER_09"}, {"start": 2829.073, "end": 2829.293, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 2829.834, "end": 2838.545, "text": "And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much.", "speaker": "SPEAKER_09"}, {"start": 2839.085, "end": 2841.929, "text": "Because the other group is adding to your uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2842.129, "end": 2849.338, "text": "Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe.", "speaker": "SPEAKER_09"}, {"start": 2849.418, "end": 2854.845, "text": "So, so you see that there, there is a kind of like logic where the free energy principle is involved.", "speaker": "SPEAKER_09"}, {"start": 2855.129, "end": 2864.026, "text": " where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise.", "speaker": "SPEAKER_09"}, {"start": 2864.467, "end": 2870.679, "text": "And that has nothing to do with the true validity of science, right?", "speaker": "SPEAKER_09"}, {"start": 2871.641, "end": 2875.348, "text": "Like I'm attempting to be a scientist explaining all this stuff,", "speaker": "SPEAKER_09"}, {"start": 2875.683, "end": 2878.748, "text": " But the people who, you see what I'm saying?", "speaker": "SPEAKER_09"}, {"start": 2878.788, "end": 2884.538, "text": "Like to take in sensory information is what we do.", "speaker": "SPEAKER_09"}, {"start": 2884.578, "end": 2888.544, "text": "It's not about like following proper logic and stuff.", "speaker": "SPEAKER_09"}, {"start": 2888.564, "end": 2892.19, "text": "It's people learn to do that, right?", "speaker": "SPEAKER_09"}, {"start": 2892.17, "end": 2894.737, "text": " Can I interject something real quick?", "speaker": "SPEAKER_01"}, {"start": 2894.757, "end": 2906.128, "text": "I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're", "speaker": "SPEAKER_01"}, {"start": 2906.581, "end": 2923.723, "text": " Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality.", "speaker": "SPEAKER_01"}, {"start": 2924.364, "end": 2931.593, "text": "So what happens in humans groups is that the active inferencing that's happening in a lot of individuals,", "speaker": "SPEAKER_01"}, {"start": 2931.573, "end": 2937.991, "text": " is to analyze and update whether or not they believe the person who's giving the message.", "speaker": "SPEAKER_01"}, {"start": 2938.552, "end": 2942.864, "text": "It's not about evaluating the validity of the premises.", "speaker": "SPEAKER_01"}, {"start": 2943.486, "end": 2947.236, "text": "It's not about the... That gets at a very...", "speaker": "SPEAKER_01"}, {"start": 2947.216, "end": 2957.913, "text": " That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer.", "speaker": "SPEAKER_05"}, {"start": 2958.734, "end": 2961.178, "text": "We might imagine an active inference agent doing inference about this.", "speaker": "SPEAKER_05"}, {"start": 2961.759, "end": 2967.688, "text": "But above that is a question, how reliable is this thing that I'm hearing at all?", "speaker": "SPEAKER_05"}, {"start": 2967.668, "end": 2992.039, "text": " and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant", "speaker": "SPEAKER_05"}, {"start": 2992.019, "end": 3018.583, "text": " uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation", "speaker": "SPEAKER_05"}, {"start": 3018.563, "end": 3022.371, "text": " for what it means for one entity to be an active inference agent.", "speaker": "SPEAKER_05"}, {"start": 3022.391, "end": 3031.651, "text": "Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents.", "speaker": "SPEAKER_05"}, {"start": 3031.671, "end": 3033.595, "text": "But that's not going to be the focus of the book.", "speaker": "SPEAKER_05"}, {"start": 3034.396, "end": 3038.445, "text": "A lot of that is at the cutting edge of active inference research now.", "speaker": "SPEAKER_05"}, {"start": 3039.961, "end": 3044.874, "text": " Thanks for keeping it closer to the textbook by bringing up precision.", "speaker": "SPEAKER_09"}, {"start": 3044.934, "end": 3046.117, "text": "I very much agree with that.", "speaker": "SPEAKER_09"}, {"start": 3046.157, "end": 3052.915, "text": "And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but.", "speaker": "SPEAKER_09"}, {"start": 3054.459, "end": 3055.422, "text": "This.", "speaker": "SPEAKER_09"}, {"start": 3055.655, "end": 3065.723, "text": " Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference.", "speaker": "SPEAKER_09"}, {"start": 3065.743, "end": 3070.095, "text": "So active inferences is generally setting up these principles.", "speaker": "SPEAKER_09"}, {"start": 3070.115, "end": 3070.997, "text": "It.", "speaker": "SPEAKER_09"}, {"start": 3070.977, "end": 3092.985, "text": " definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different", "speaker": "SPEAKER_09"}, {"start": 3092.965, "end": 3093.386, "text": " Right?", "speaker": "SPEAKER_09"}, {"start": 3093.806, "end": 3097.411, "text": "Like, our variables that we're including in our respective models are different.", "speaker": "SPEAKER_09"}, {"start": 3097.772, "end": 3099.694, "text": "The precisions can definitely differ.", "speaker": "SPEAKER_09"}, {"start": 3099.714, "end": 3106.303, "text": "If I get information from someone in my group, I might have a higher precision and trust what they say more.", "speaker": "SPEAKER_09"}, {"start": 3106.764, "end": 3114.654, "text": "I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right?", "speaker": "SPEAKER_09"}, {"start": 3115.155, "end": 3118.72, "text": "So a lot of the, you know, a lot of the", "speaker": "SPEAKER_09"}, {"start": 3119.054, "end": 3131.018, "text": " political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right?", "speaker": "SPEAKER_09"}, {"start": 3131.159, "end": 3134.245, "text": "It's like, oh, I just lump them all in with political group", "speaker": "SPEAKER_09"}, {"start": 3134.225, "end": 3156.351, "text": " be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available", "speaker": "SPEAKER_09"}, {"start": 3156.331, "end": 3172.112, "text": " then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone", "speaker": "SPEAKER_09"}, {"start": 3172.092, "end": 3191.299, "text": " ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it", "speaker": "SPEAKER_09"}, {"start": 3192.308, "end": 3195.736, "text": " Do make sure, I've been absent for a little bit.", "speaker": "SPEAKER_05"}, {"start": 3196.237, "end": 3200.687, "text": "I'm gonna be much more active on the questions side of things on the CODA.", "speaker": "SPEAKER_05"}, {"start": 3200.748, "end": 3205.418, "text": "So do make sure if you've got questions, please do put them down in the questions tab here.", "speaker": "SPEAKER_05"}, {"start": 3205.458, "end": 3209.608, "text": "I've gone through and I've already answered quite a few of them for chapter five.", "speaker": "SPEAKER_05"}, {"start": 3209.588, "end": 3224.188, "text": " uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see", "speaker": "SPEAKER_05"}, {"start": 3224.573, "end": 3226.135, "text": " Are there any more live questions?", "speaker": "SPEAKER_05"}, {"start": 3227.777, "end": 3228.718, "text": "That would be nice.", "speaker": "SPEAKER_05"}, {"start": 3228.738, "end": 3233.584, "text": "We can stay around for maybe five or so minutes after the deadline if people so choose.", "speaker": "SPEAKER_05"}, {"start": 3234.645, "end": 3236.267, "text": "But we are coming up to the top of the hour.", "speaker": "SPEAKER_05"}, {"start": 3236.287, "end": 3241.994, "text": "So yeah, all of the questions in chapter five now are answered, or I've given my attempt.", "speaker": "SPEAKER_05"}, {"start": 3242.655, "end": 3246.379, "text": "As I say, I'll go through and attempt to answer all the previous ones as well.", "speaker": "SPEAKER_05"}, {"start": 3249.937, "end": 3276.737, "text": " think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry", "speaker": "SPEAKER_05"}, {"start": 3277.595, "end": 3301.122, "text": " okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian", "speaker": "SPEAKER_08"}, {"start": 3301.102, "end": 3308.391, "text": " inference to engage in parameter learning so that it can update beliefs about the current state.", "speaker": "SPEAKER_08"}, {"start": 3309.172, "end": 3319.545, "text": "If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself.", "speaker": "SPEAKER_08"}, {"start": 3319.985, "end": 3323.97, "text": "And I think this is what we call plasticity.", "speaker": "SPEAKER_08"}, {"start": 3323.95, "end": 3331.683, "text": " The ultimate goal of all this is to minimize expected free energy.", "speaker": "SPEAKER_08"}, {"start": 3332.404, "end": 3333.085, "text": "Is that right?", "speaker": "SPEAKER_08"}, {"start": 3334.548, "end": 3335.088, "text": "Wow, okay.", "speaker": "SPEAKER_05"}, {"start": 3335.129, "end": 3338.174, "text": "I mean, basically, yeah, that's substantively correct.", "speaker": "SPEAKER_05"}, {"start": 3339.115, "end": 3342.06, "text": "I will note, however, we haven't yet talked about expected free energy.", "speaker": "SPEAKER_05"}, {"start": 3342.395, "end": 3344.237, "text": " and planning and action selection.", "speaker": "SPEAKER_05"}, {"start": 3344.257, "end": 3345.358, "text": "So that is going to come later.", "speaker": "SPEAKER_05"}, {"start": 3346.379, "end": 3346.959, "text": "But yeah, you're right.", "speaker": "SPEAKER_05"}, {"start": 3346.979, "end": 3352.825, "text": "In terms of the general flavor of things, we have beliefs about hidden states in the world.", "speaker": "SPEAKER_05"}, {"start": 3353.686, "end": 3365.076, "text": "And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 3365.997, "end": 3369.921, "text": "Or we've seen that we can also recast this in terms of predictive processing.", "speaker": "SPEAKER_05"}, {"start": 3370.66, "end": 3372.943, "text": " minimizing the precision way to prediction errors.", "speaker": "SPEAKER_05"}, {"start": 3372.963, "end": 3373.924, "text": "It's the same kind of story.", "speaker": "SPEAKER_05"}, {"start": 3374.926, "end": 3386.221, "text": "And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states.", "speaker": "SPEAKER_05"}, {"start": 3387.063, "end": 3387.283, "text": "Yes.", "speaker": "SPEAKER_05"}, {"start": 3388.725, "end": 3391.328, "text": "The other thing, you mentioned parameter learning.", "speaker": "SPEAKER_05"}, {"start": 3391.388, "end": 3392.77, "text": "So yes, exactly.", "speaker": "SPEAKER_05"}, {"start": 3392.79, "end": 3393.932, "text": "There's these two problems.", "speaker": "SPEAKER_05"}, {"start": 3394.032, "end": 3397.317, "text": "We have the problem of inferring the hidden states.", "speaker": "SPEAKER_05"}, {"start": 3397.357, "end": 3398.338, "text": "What should the hidden states be?", "speaker": "SPEAKER_05"}, {"start": 3398.977, "end": 3407.788, "text": " But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference.", "speaker": "SPEAKER_05"}, {"start": 3408.529, "end": 3415.017, "text": "So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like.", "speaker": "SPEAKER_05"}, {"start": 3415.738, "end": 3419.183, "text": "But then above that, we need to deal with the problem of what should the setting of the parameters be?", "speaker": "SPEAKER_05"}, {"start": 3420.204, "end": 3426.853, "text": "And we've only really just begun to look at this in terms of prediction errors and hierarchical models.", "speaker": "SPEAKER_05"}, {"start": 3428.034, "end": 3428.815, "text": "But we saw", "speaker": "SPEAKER_05"}, {"start": 3429.149, "end": 3432.913, "text": " how to deal with that in terms of expectation maximization and earlier chapters.", "speaker": "SPEAKER_05"}, {"start": 3433.634, "end": 3443.644, "text": "We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture.", "speaker": "SPEAKER_05"}, {"start": 3443.964, "end": 3450.251, "text": "Because we generally can't do expectation maximization for the reason that we don't know the exact posterior.", "speaker": "SPEAKER_05"}, {"start": 3450.291, "end": 3458.299, "text": "So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah.", "speaker": "SPEAKER_05"}, {"start": 3458.339, "end": 3458.599, "text": "Thank you.", "speaker": "SPEAKER_08"}, {"start": 3459.49, "end": 3461.011, "text": " No problem.", "speaker": "SPEAKER_05"}, {"start": 3461.032, "end": 3462.473, "text": "Do we have more questions?", "speaker": "SPEAKER_05"}, {"start": 3462.633, "end": 3463.254, "text": "More questions?", "speaker": "SPEAKER_05"}, {"start": 3464.655, "end": 3466.857, "text": "Maybe one more question, if there is one more.", "speaker": "SPEAKER_05"}, {"start": 3467.117, "end": 3469.52, "text": "And then I'll stop the recording.", "speaker": "SPEAKER_05"}, {"start": 3475.566, "end": 3475.866, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 3475.886, "end": 3476.887, "text": "I might stop the recording here.", "speaker": "SPEAKER_05"}, {"start": 3478.068, "end": 3478.549, "text": "Well, Mark.", "speaker": "SPEAKER_05"}, {"start": 3478.569, "end": 3479.71, "text": "Hello, everyone, and welcome.", "speaker": "SPEAKER_05"}, {"start": 3479.91, "end": 3481.732, "text": "I've got my esteemed friend.", "speaker": "SPEAKER_05"}, {"start": 3482.753, "end": 3484.975, "text": "Oh, hang on.", "speaker": "SPEAKER_05"}, {"start": 3486.423, "end": 3489.528, "text": " That was interesting audio feedback.", "speaker": "SPEAKER_02"}, {"start": 3489.548, "end": 3492.053, "text": "I thought I would jump in since nobody else asked us.", "speaker": "SPEAKER_02"}, {"start": 3493.135, "end": 3499.105, "text": "But earlier you mentioned using gradient descent to learn parameters, right?", "speaker": "SPEAKER_02"}, {"start": 3499.245, "end": 3501.289, "text": "Is that also used in perception?", "speaker": "SPEAKER_02"}, {"start": 3501.69, "end": 3503.893, "text": "Is it the same approach?", "speaker": "SPEAKER_02"}, {"start": 3503.994, "end": 3508.581, "text": "Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things.", "speaker": "SPEAKER_05"}, {"start": 3508.602, "end": 3510.625, "text": "So it's a very, very general", "speaker": "SPEAKER_05"}, {"start": 3511.06, "end": 3512.522, "text": " optimization techniques.", "speaker": "SPEAKER_05"}, {"start": 3512.602, "end": 3520.19, "text": "So yes, it certainly can be used for learning, well, for perception.", "speaker": "SPEAKER_05"}, {"start": 3520.711, "end": 3522.513, "text": "I'm trying to find a nice picture here.", "speaker": "SPEAKER_05"}, {"start": 3523.173, "end": 3531.983, "text": "Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent.", "speaker": "SPEAKER_05"}, {"start": 3532.524, "end": 3533.445, "text": "Let me just make sure of that.", "speaker": "SPEAKER_05"}, {"start": 3533.565, "end": 3535.107, "text": "Yes, in other words, is the answer.", "speaker": "SPEAKER_05"}, {"start": 3535.547, "end": 3537.049, "text": "It's used for a lot of things.", "speaker": "SPEAKER_05"}, {"start": 3538.39, "end": 3539.852, "text": "Very, very general technique.", "speaker": "SPEAKER_05"}, {"start": 3540.237, "end": 3549.91, "text": " Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six.", "speaker": "SPEAKER_05"}, {"start": 3549.93, "end": 3550.39, "text": "Absolutely.", "speaker": "SPEAKER_05"}, {"start": 3550.51, "end": 3550.891, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3550.911, "end": 3552.593, "text": "We're gonna talk about phase planes and so on.", "speaker": "SPEAKER_05"}, {"start": 3553.394, "end": 3553.635, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3553.955, "end": 3556.258, "text": "There are problems with it, but yes, we're gonna see it going forward.", "speaker": "SPEAKER_05"}, {"start": 3556.278, "end": 3559.242, "text": "It's a very, it's a very foundational technique.", "speaker": "SPEAKER_05"}, {"start": 3559.382, "end": 3562.807, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3562.827, "end": 3563.027, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 3563.047, "end": 3566.211, "text": "Any, any last minute questions for the YouTube recording before we, uh,", "speaker": "SPEAKER_05"}, {"start": 3569.093, "end": 3570.975, "text": " If not, I think I'm sorry.", "speaker": "SPEAKER_05"}, {"start": 3571.236, "end": 3572.377, "text": "Can I ask a quick question?", "speaker": "SPEAKER_07"}, {"start": 3572.437, "end": 3574.2, "text": "Yeah, sure.", "speaker": "SPEAKER_05"}, {"start": 3574.74, "end": 3575.461, "text": "I do apologize.", "speaker": "SPEAKER_07"}, {"start": 3575.702, "end": 3577.424, "text": "I have joined the group very late.", "speaker": "SPEAKER_07"}, {"start": 3577.484, "end": 3579.767, "text": "So maybe this was addressed in like the past weeks.", "speaker": "SPEAKER_07"}, {"start": 3579.827, "end": 3593.245, "text": "But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning?", "speaker": "SPEAKER_07"}, {"start": 3593.478, "end": 3619.553, "text": " well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms", "speaker": "SPEAKER_05"}, {"start": 3619.955, "end": 3629.547, "text": " If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control.", "speaker": "SPEAKER_05"}, {"start": 3630.147, "end": 3643.784, "text": "There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference.", "speaker": "SPEAKER_05"}, {"start": 3643.804, "end": 3646.988, "text": "So the answer is yes, there is a profound relationship.", "speaker": "SPEAKER_05"}, {"start": 3647.008, "end": 3649.371, "text": "We probably don't have time to get into it here.", "speaker": "SPEAKER_05"}, {"start": 3649.84, "end": 3659.489, "text": " One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain.", "speaker": "SPEAKER_05"}, {"start": 3660.41, "end": 3665.735, "text": "Because we've seen with, well, we haven't yet seen with expected free energy how that works.", "speaker": "SPEAKER_05"}, {"start": 3665.755, "end": 3672.461, "text": "But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning.", "speaker": "SPEAKER_05"}, {"start": 3673.062, "end": 3679.628, "text": "And we don't just have this scale of reward, this signal in reinforcement learning.", "speaker": "SPEAKER_05"}, {"start": 3680.114, "end": 3708.67, "text": " we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um", "speaker": "SPEAKER_05"}, {"start": 3708.97, "end": 3717.107, "text": " a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning.", "speaker": "SPEAKER_09"}, {"start": 3717.147, "end": 3727.688, "text": "Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning,", "speaker": "SPEAKER_09"}, {"start": 3727.668, "end": 3741.341, "text": " Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two.", "speaker": "SPEAKER_09"}, {"start": 3742.282, "end": 3752.331, "text": "But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven.", "speaker": "SPEAKER_09"}, {"start": 3753.192, "end": 3756.595, "text": "So there can be like a KL control agent or there can be a variety.", "speaker": "SPEAKER_09"}, {"start": 3756.575, "end": 3779.157, "text": " other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that", "speaker": "SPEAKER_09"}, {"start": 3779.137, "end": 3782.622, "text": " that things like curiosity exist.", "speaker": "SPEAKER_09"}, {"start": 3782.742, "end": 3787.429, "text": "Why does curiosity exist if we're actually always just driven towards a reward?", "speaker": "SPEAKER_09"}, {"start": 3787.509, "end": 3790.173, "text": "Once you know what the reward is, you should just go for that, right?", "speaker": "SPEAKER_09"}, {"start": 3790.213, "end": 3796.262, "text": "And you would have no reason to find some other strategy for going for it.", "speaker": "SPEAKER_09"}, {"start": 3796.463, "end": 3798.466, "text": "You would have no reason to go out of your way.", "speaker": "SPEAKER_09"}, {"start": 3798.606, "end": 3803.032, "text": "From what you already know, you just keep going for the reward as much as possible.", "speaker": "SPEAKER_09"}, {"start": 3803.393, "end": 3808.941, "text": "Perhaps you learn other things through happenstance along the way.", "speaker": "SPEAKER_09"}, {"start": 3809.258, "end": 3811.041, "text": " That's a big difference.", "speaker": "SPEAKER_05"}, {"start": 3811.722, "end": 3814.948, "text": "In reinforcement, you can just sort of start maximizing a reward signal.", "speaker": "SPEAKER_05"}, {"start": 3815.569, "end": 3822.601, "text": "But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards.", "speaker": "SPEAKER_05"}, {"start": 3822.621, "end": 3828.972, "text": "You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards.", "speaker": "SPEAKER_05"}, {"start": 3828.952, "end": 3856.731, "text": " yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite", "speaker": "SPEAKER_09"}, {"start": 3857.234, "end": 3868.402, "text": " how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning.", "speaker": "SPEAKER_09"}, {"start": 3868.743, "end": 3872.492, "text": "Meanwhile, active inference has a much more principled way that relates to this notion of", "speaker": "SPEAKER_09"}, {"start": 3872.472, "end": 3892.793, "text": " free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually", "speaker": "SPEAKER_09"}, {"start": 3893.06, "end": 3896.992, "text": " Um, when I play a sport, I throw the ball this way.", "speaker": "SPEAKER_09"}, {"start": 3897.012, "end": 3906.681, "text": "Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person.", "speaker": "SPEAKER_09"}, {"start": 3906.897, "end": 3909.963, "text": " but maybe I will kind of curve it or change it, right?", "speaker": "SPEAKER_09"}, {"start": 3910.343, "end": 3914.03, "text": "You're not randomly throwing it in the sky or in the opposite direction.", "speaker": "SPEAKER_09"}, {"start": 3914.431, "end": 3919.481, "text": "You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right?", "speaker": "SPEAKER_09"}, {"start": 3919.541, "end": 3927.235, "text": "There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better.", "speaker": "SPEAKER_09"}, {"start": 3927.552, "end": 3953.037, "text": " based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about", "speaker": "SPEAKER_09"}, {"start": 3953.523, "end": 3956.968, "text": " How do I perform the best whenever it comes to active inference?", "speaker": "SPEAKER_09"}, {"start": 3957.129, "end": 3965.422, "text": "It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them.", "speaker": "SPEAKER_09"}, {"start": 3966.223, "end": 3975.918, "text": "But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear.", "speaker": "SPEAKER_09"}, {"start": 3976.038, "end": 3977.32, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3978.194, "end": 3983.299, "text": " I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys.", "speaker": "SPEAKER_05"}, {"start": 3983.319, "end": 3986.502, "text": "So Giancuomo, fire away.", "speaker": "SPEAKER_05"}, {"start": 3986.522, "end": 3991.827, "text": "Very quickly, I think it ties in with what was just spoken and talked about now.", "speaker": "SPEAKER_00"}, {"start": 3992.648, "end": 4007.982, "text": "Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is?", "speaker": "SPEAKER_00"}, {"start": 4008.265, "end": 4033.468, "text": " Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty.", "speaker": "SPEAKER_00"}, {"start": 4033.508, "end": 4035.029, "text": "They say, this is the answer.", "speaker": "SPEAKER_00"}, {"start": 4035.109, "end": 4037.091, "text": "And you're like, no, no, no, it's not.", "speaker": "SPEAKER_00"}, {"start": 4037.594, "end": 4057.919, "text": " And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware.", "speaker": "SPEAKER_00"}, {"start": 4057.959, "end": 4064.007, "text": "Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind?", "speaker": "SPEAKER_00"}, {"start": 4065.455, "end": 4093.251, "text": " well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions", "speaker": "SPEAKER_05"}, {"start": 4093.687, "end": 4094.989, "text": " This is built in from the start.", "speaker": "SPEAKER_05"}, {"start": 4095.029, "end": 4103.639, "text": "So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right?", "speaker": "SPEAKER_05"}, {"start": 4104.22, "end": 4105.982, "text": "And okay, we can't do Bayes' rule exactly.", "speaker": "SPEAKER_05"}, {"start": 4106.022, "end": 4107.023, "text": "We have to do it approximately.", "speaker": "SPEAKER_05"}, {"start": 4107.083, "end": 4108.365, "text": "We do variational inference.", "speaker": "SPEAKER_05"}, {"start": 4108.405, "end": 4110.407, "text": "We do prediction error and minimization, that kind of thing.", "speaker": "SPEAKER_05"}, {"start": 4110.828, "end": 4114.052, "text": "But we're always, always, always reasoning about probability distribution.", "speaker": "SPEAKER_05"}, {"start": 4114.072, "end": 4118.597, "text": "So yes, every single active implementation ever is always", "speaker": "SPEAKER_05"}, {"start": 4120.535, "end": 4147.757, "text": " predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet", "speaker": "SPEAKER_05"}, {"start": 4147.737, "end": 4149.84, "text": " And there's issues around that and how to do that.", "speaker": "SPEAKER_05"}, {"start": 4150.821, "end": 4151.722, "text": "But we haven't seen that just yet.", "speaker": "SPEAKER_05"}, {"start": 4151.742, "end": 4153.024, "text": "So, yes, absolutely.", "speaker": "SPEAKER_05"}, {"start": 4153.084, "end": 4155.808, "text": "It's part and parcel of what it is to be an active infatuation.", "speaker": "SPEAKER_05"}, {"start": 4155.828, "end": 4157.15, "text": "It's the reason about uncertainty now.", "speaker": "SPEAKER_05"}, {"start": 4158.411, "end": 4158.932, "text": "Okay, thank you.", "speaker": "SPEAKER_05"}, {"start": 4159.613, "end": 4159.913, "text": "No problem.", "speaker": "SPEAKER_05"}, {"start": 4159.933, "end": 4162.036, "text": "And then, Mark, and then I think we can do... Oh, hang on.", "speaker": "SPEAKER_05"}, {"start": 4162.337, "end": 4167.163, "text": "Maybe just really quickly, did you have a comment, Andrew?", "speaker": "SPEAKER_05"}, {"start": 4167.413, "end": 4171.819, "text": " I also really have to go, but yeah, just briefly, you basically answered it.", "speaker": "SPEAKER_09"}, {"start": 4171.839, "end": 4173.622, "text": "Sorry, I was looking at the chat.", "speaker": "SPEAKER_09"}, {"start": 4173.642, "end": 4177.928, "text": "It's just, yeah, it depends on how the question is being asked.", "speaker": "SPEAKER_09"}, {"start": 4177.948, "end": 4185.339, "text": "Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are?", "speaker": "SPEAKER_09"}, {"start": 4185.459, "end": 4191.348, "text": "Like that's going to take a little bit more than just like only looking at the simplified models.", "speaker": "SPEAKER_09"}, {"start": 4191.368, "end": 4193.05, "text": "We're looking at the textbook here.", "speaker": "SPEAKER_09"}, {"start": 4193.03, "end": 4216.928, "text": " um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization", "speaker": "SPEAKER_09"}, {"start": 4216.908, "end": 4230.176, "text": " of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots.", "speaker": "SPEAKER_09"}, {"start": 4230.256, "end": 4232.18, "text": "There are 2 possible things.", "speaker": "SPEAKER_09"}, {"start": 4232.22, "end": 4235.507, "text": "It could be is completely 50 50 is fully uncertain.", "speaker": "SPEAKER_09"}, {"start": 4235.487, "end": 4246.867, "text": " Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision.", "speaker": "SPEAKER_09"}, {"start": 4247.388, "end": 4247.609, "text": "Right?", "speaker": "SPEAKER_09"}, {"start": 4247.709, "end": 4255.423, "text": "Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob.", "speaker": "SPEAKER_09"}, {"start": 4255.803, "end": 4256.044, "text": "Right?", "speaker": "SPEAKER_09"}, {"start": 4256.104, "end": 4258.508, "text": "The more you turn up precision, the more you're.", "speaker": "SPEAKER_09"}, {"start": 4258.488, "end": 4263.134, "text": " you're going to take into account the prediction errors you're receiving and vice versa.", "speaker": "SPEAKER_09"}, {"start": 4263.214, "end": 4275.609, "text": "So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving.", "speaker": "SPEAKER_09"}, {"start": 4275.669, "end": 4275.89, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 4275.91, "end": 4279.294, "text": "So so uncertainty is very much like throughout.", "speaker": "SPEAKER_09"}, {"start": 4279.474, "end": 4284.18, "text": "I mean, it's a very ubiquitous term through many parts of active inference.", "speaker": "SPEAKER_09"}, {"start": 4284.2, "end": 4284.34, "text": "Yeah.", "speaker": "SPEAKER_09"}, {"start": 4284.48, "end": 4285.742, "text": "So it's essential.", "speaker": "SPEAKER_09"}, {"start": 4287.224, "end": 4288.225, "text": "So all right.", "speaker": "SPEAKER_03"}, {"start": 4288.677, "end": 4311.58, "text": " so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see", "speaker": "SPEAKER_03"}, {"start": 4312.673, "end": 4337.179, "text": " right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh", "speaker": "SPEAKER_09"}, {"start": 4337.547, "end": 4338.789, "text": " who cares about priors.", "speaker": "SPEAKER_09"}, {"start": 4338.829, "end": 4342.634, "text": "And that's what we see with maximum likelihood estimation in chapter two there.", "speaker": "SPEAKER_09"}, {"start": 4343.655, "end": 4354.31, "text": "But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something.", "speaker": "SPEAKER_09"}, {"start": 4354.35, "end": 4366.086, "text": "It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like,", "speaker": "SPEAKER_09"}, {"start": 4366.403, "end": 4392.141, "text": " thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um", "speaker": "SPEAKER_09"}, {"start": 4392.661, "end": 4394.263, "text": " you know, how those connect up with each other.", "speaker": "SPEAKER_09"}, {"start": 4394.704, "end": 4403.396, "text": "So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there.", "speaker": "SPEAKER_09"}, {"start": 4403.476, "end": 4413.409, "text": "And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe.", "speaker": "SPEAKER_09"}, {"start": 4413.55, "end": 4416.193, "text": "And so I fully trust it without question.", "speaker": "SPEAKER_09"}, {"start": 4416.654, "end": 4416.854, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 4416.874, "end": 4419.057, "text": "So I say all these things because, uh,", "speaker": "SPEAKER_09"}, {"start": 4419.037, "end": 4426.085, "text": " Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe.", "speaker": "SPEAKER_09"}, {"start": 4426.666, "end": 4442.303, "text": "And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things.", "speaker": "SPEAKER_09"}, {"start": 4442.343, "end": 4444.065, "text": "Yeah, it's very interesting to think about.", "speaker": "SPEAKER_09"}, {"start": 4445.294, "end": 4449.623, "text": " Yeah, do take a look at, I think I'm sharing, but figure 2.11.", "speaker": "SPEAKER_05"}, {"start": 4450.525, "end": 4459.965, "text": "This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief.", "speaker": "SPEAKER_05"}, {"start": 4459.985, "end": 4462.29, "text": "So if the precision is really tight around the prior,", "speaker": "SPEAKER_05"}, {"start": 4462.692, "end": 4472.808, "text": " the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model.", "speaker": "SPEAKER_05"}, {"start": 4472.828, "end": 4474.511, "text": "So yeah, do take a look at that.", "speaker": "SPEAKER_05"}, {"start": 4474.531, "end": 4482.584, "text": "That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have.", "speaker": "SPEAKER_05"}, {"start": 4482.684, "end": 4483.786, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 4483.806, "end": 4484.146, "text": "Excellent.", "speaker": "SPEAKER_03"}, {"start": 4484.206, "end": 4484.667, "text": "Thank you.", "speaker": "SPEAKER_03"}, {"start": 4485.372, "end": 4485.512, "text": " Cool.", "speaker": "SPEAKER_05"}, {"start": 4485.532, "end": 4487.255, "text": "All right, guys, we're going to have to call it there.", "speaker": "SPEAKER_05"}, {"start": 4487.756, "end": 4491.963, "text": "Do put questions in the chats on the page, the code page.", "speaker": "SPEAKER_05"}, {"start": 4492.003, "end": 4493.606, "text": "I'll be a bit more attentive going forward.", "speaker": "SPEAKER_05"}, {"start": 4494.267, "end": 4495.489, "text": "I look forward to next week.", "speaker": "SPEAKER_05"}, {"start": 4495.509, "end": 4499.796, "text": "I'll be doing another session of this kind on Friday for the people who usually attend that session.", "speaker": "SPEAKER_05"}, {"start": 4499.816, "end": 4500.557, "text": "So thank you very much.", "speaker": "SPEAKER_05"}, {"start": 4500.918, "end": 4502.22, "text": "Stop sharing and stop the recording.", "speaker": "SPEAKER_05"}, {"start": 4503.522, "end": 4504.704, "text": "Have we got a recording here?", "speaker": "SPEAKER_05"}, {"start": 4506.247, "end": 4507.269, "text": "Okay, stop the recording.", "speaker": "SPEAKER_05"}, {"start": 4507.749, "end": 4508.551, "text": "Goodbye, YouTube people.", "speaker": "SPEAKER_05"}, {"start": 4508.751, "end": 4509.372, "text": "Until next time.", "speaker": "SPEAKER_05"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt new file mode 100644 index 000000000..e44125349 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt @@ -0,0 +1,1508 @@ +SPEAKER_05: +all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense + +And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two. + +We're going to do active inputs proper, which is going to be very, very exciting going forward. + +So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts. + +that we're going to start doing actual, quote unquote, active inference in the second part. + +But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection. + +We're going to sort of slow down and think about part one in general, chapters one to five. + +This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one. + +you know, going so that we can be adequately grounded going forward. + +I will say we have one one question is, you know, we've done each chapter across two weeks. + +Historically, you know, chapter one, we've had two weeks going forward. + +depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week + +of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe + +It would be a good idea to reach out through the blankets email. + +Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place. + +That would be amazing. + +Good idea. + +Yes, please. + +I see a lot of people are saying yes. + +In this chat here, it would be very helpful if you could give your yay or nay to that as well. + +So we can actually look at that just directly through the chat here. + +So if you want a second week review, yes. + +If you don't, no. + +And then we can go forward. + +All right. + +Enough of that. + +I'll start sharing my screen here. + + +SPEAKER_04: +Probably my entire screen. + + +SPEAKER_05: +Okie dokie. + +So people should be able to see. + +Maybe just get rid of... + +The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise. + +about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here. + +So you've got your overview. + +I've added one or two additional resources. + +So these actually came up. + +Andrew shared the first of these. + +These are all videos. + +One of them is an active inference + +coding and active inference. + +So that's Ryan Smith and his team. + +They've done excellent work on predictive coding. + +I'll play that just now. + +This is a great stream just about predictive coding more generally and its relation to active inference and its relation to + +you know, its status in actual brains. + +So if you're interested in more context, this would be excellent to go through. + +And there's two more links there as well. + +One by Jakob Howie, he's a philosopher over in Australia in the University of Monash. + +What is predictive processing and what is it good for? + +Another excellent talk. + +And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke. + +from the University of Toronto. + +He's actually speaking about Neoplatonism and mystical experiences, believe it or not. + +And I know it sounds weird. + +Predictive processing actually shows up in here in a very significant way. + +So that might be a nice mood setter, as it were. + +So there's some additional resources there. + +I've begun filling in the map for chapter five. + +So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid. + +the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that + +A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation. + +What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you. + +You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you. + +So it's not ideal, but the content is there at least. + +And especially for those who don't have the book, I think this is quite useful. + +So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter. + +I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts + +uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there + +Any questions? + +Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that. + +I'll start my share temporarily. + +Just coming back. + +Very good. + +Yes, but it's limited. + +Formatting in LaTeX is a bit strange in Coda. + +We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda. + +OK, so I'll share, come back. + +Let's go all the way back, hopefully people can see my screen again, to the introduction. + +So everyone's seen this figure here. + +This is essentially the first figure in the book. + +This is the breakdown of the partitions of the book, so part one, part two, part three. + +We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding. + +Part two is active inference core. + +So this is the heart of the book, really. + +This is the fundamentals of active inference per se. + +what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end + +But as has been the case, everything has been just perception only. + +So we haven't actually dealt with actions. + +And we're going to be moving into that with part two. + +But even more fundamentally, everything has been static. + +So all of the models we've been looking at, I don't know if I can get a nice example. + +All of the models, so this is chapter one. + +Let's maybe go down to some of the equations. + +They've all been static. + +So the environment hasn't really been changing at all. + +And this is very simple. + +So part one, really going to chapter two, I think. + +So we have an environment. + +This is one we've seen this many times now. + +This is a representation of the generative process, the real process in and of itself out there. + +This is fairly static, so nothing is really changing. + +The hidden states aren't really changing from moment to moment. + +We're just presented with a situation and we have to update our beliefs about what hidden states might be. + +That's been constant throughout all of Chapter 1. + +Things have gotten progressively more complicated. + +We've looked at univariate hidden states where there's literally just one hidden state. + +We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things. + +But in every case, it's been static. + +The dynamics of the hidden state have been non-existent. + +That is going to change going into part two. + +We're going to start looking at generative processes which change over time. + +And they're the interesting ones. + +They're the ones that are actually useful. + +And they're the ones that allow us to actually begin to have a reason to act in the environment. + +We haven't really had any reason to perform actions yet. + +So that's a big move that's going to take place and that we're going to have to deal with. + +So maybe just coming back. + +So chapter one was about, what was this about? + +It was about perception. + +Okay. + +So like in terms of how we're going to think about perception in active inference. + +So in chapter one, we cast or we framed + +of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at + +you know, how to do like exact Bayesian inference, okay, if we were able to do everything. + +We saw how to use Bayes' rule basically. + +So we cemented the distinction personally before that around the difference between the generative process and the generative model. + +So I come all the way down to the figures here. + +I don't have a particularly good figure for that. + +Here we go, 2.4. + +The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world. + +This was the crucial thing, I think, in chapter two, basically. + +And we have a way of notating this convention. + +We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah. + +These belong to the generative process, the real world out there. + +Okay, and then modeled variables are without a star. + +So they correspond to our model. + +And what is our model? + +It is quite literally a joint probability distribution over hidden states, latent states, and observations. + +We use x for latent states and y for observations. + +But you can see, of course, the x in here is our model of the real latent state outside of here, x star. + +And they need not agree with each other. + +They need not be the same. + +But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be. + +That is p of x and y. + +And then if we have the model evidence, the probability of observing anything, + +or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation. + +And that was kind of chapter two. + +And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions. + +And these are very ubiquitous. + +They show up everywhere. + +We're going to continue that. + +In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference. + +There are other distributions as well. + +We haven't really looked at those yet. + +But the real sort of, there's a problem with this, which is that, yes, we can do inference. + +Yes, we might even be able to do exact inference for really, really simple problems. + +But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model. + +And then we're going to do inference. + +And inference is like, I set the dials, and then my machine does some inference stuff, right? + +that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation + +And we're doing good old fashioned Bayesian inference. + +But now we're having to do parameter learning and estimation as well. + +And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference. + +So we have these two processes, inference and learning parameter estimation. + +And they're both able to be done by means of Bayesian inference. + +That's the crucial thing. + +So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah. + +And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent. + +And we saw many versions of this across the chapter and indeed in other chapters. + +It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters. + +So let's imagine we have our two parameters, I don't know, beta 1, beta 0. + +And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space. + +That's this surface. + +And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point. + +And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters. + +Very general way of doing optimization. + +However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be. + +And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization. + +So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now. + +A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding. + +um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model + +And also, how should I use those settings to do good inference? + +It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters. + +So imagine you've got learning up here. + +What should the model parameters be? + +And you can do inference on model parameters. + +And then that can inform how you do inference at the low-level hidden states. + +So there's kind of these two processes that are happening in two different timescales. + +That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally. + +So that's kind of chapter three. + +I'm going to race through this. + +And we'll get to chapter five and then to some actual questions from you guys. + +Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference. + +So chapter three, we're still doing exact inference. + +But we saw that really we culminated in an algorithm + +allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be + +Those two problems are related to one another. + +To know the hidden states, you need to know what the good parameters are. + +But to know what the good parameters are, you have to know how to infer hidden states well. + +So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm. + +I won't go over that again here. + +Maybe we will if people want me to. + +But that was the solution to the problem of chapter 3, that chicken and egg problem. + +The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations. + +And that, unfortunately, is usually not something we can do. + +Being able to find the exact posterior is usually impossible, for reasons that we saw. + +We can review again if you want. + +So then chapter 4 says, all right, well, we can't do exact Bayesian inference. + +What are we going to do? + +We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like. + +So the idea with chapter 4 is, ah, OK, we're going to now + +have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior. + +We're going to try and find a posterior that's close enough to the true posterior. + +And this then motivated the idea of variational free energy as a quantity that is a measurement of how well + +an approximate posterior fits or how close it is to the true posterior. + +And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal. + +We saw from chapter three that surprisal is something we want to make very small. + +So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal, + +which is something we can't do directly, and therefore we can do approximate Bayesian inference. + +So in a lot of ways, that really kind of spiritually is the end of part one, I would say. + +The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE. + +That's, in my mind, really kind of the end of part one. + +Chapter five was a nice kind of detour into a different way of thinking about what the VFE is. + +So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization. + +we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical + +know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was + +lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning. + +It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization + +in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure + +And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow. + +And that's kind of where we left off with part one. + +So that is, in effect, what we've done, where we've gone. + +That's a lot of stuff to cover, but again, we haven't looked at actions yet. + +We're gonna be doing that in part two, and that will bring us full circle to active inference. + +And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception. + +Okay, we're gonna see that these are deeply related to each other. + +But that's the ground that we've covered. + +I'd be interested if people have questions related to that. + +I see that there's lots of questions. + +Stop sharing. + +And I see Mark has a question. + +Please fire away, Mark. + + +SPEAKER_02: +As usual. + +Thank you so much for this review. + +This is great. + +I was wondering if you could pull up that slide that showed the generative process and the generative model. + +That might be helpful as a reference. + + +SPEAKER_05: +Oh, yeah, okay. + +So let me come back here. + +That's the one in chapter two, I assume you're referring to? + + +SPEAKER_02: +I think so. + + +SPEAKER_05: +This one here? + + +SPEAKER_02: +Yes, thank you. + +Yeah, I really just kind of appreciate what this textbook is trying to do. + +And I think it took me a while to appreciate that. + +But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening. + +in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise + + +SPEAKER_05: +in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of + + +SPEAKER_02: +It seems like chapter four kind of serves as a background where the story gets started for the most part. + +And this is kind of more of a ground level + +view, which is nice and helpful. + +What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function. + +Because in my mind, I'm kind of conditioned to seeing these Gaussians, right? + +You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior. + +However, here we have this linear function, which I guess is + +then becomes part of, say, a likelihood. + +And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense. + + +SPEAKER_05: +Yeah, yeah. + +No, absolutely. + +I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations. + +um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so + +This is, in some sense, still our assumption about what the relationship is between observations and the hidden states. + +For our purpose, we've assumed that the real world, we're sort of creating a world here. + +This is the example of the food size and light intensity world. + +This is the really simple world that we've seen for all of part one. + +So what are the hidden states? + +They are the sizes of the food. + +And we can see that there's five different sizes, one to five. + +That's the hidden state. + +And the observations are light intensities. + +So light comes in, hits the food, and you get some sort of light intensity. + +And depending on the size of the food, you get some specific light intensity. + +So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size. + +So we're saying that the relationship between the observation you get in + +is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters + +So again, this is the environment regenerative process. + +The corresponding model is this one, 2.11, equation 2.11, I believe. + +now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs + +are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution. + +So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations. + +And the thing is, we're uncertain. + +So the agent doesn't get to see what the real relationship is between food size and light intensity. + +It has no idea. + +It has a model about what it thinks the relationship is. + +Now, we know that the relationship is linear. + +We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity. + +But in the real world, we have no idea what the relationship is. + +It turns out that in this example, we've specified exactly the same relationship. + +So that's quite a nice scenario to be in. + +You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size. + +Now that's wrong, but it's OK, maybe, for some settings. + +That would be an example of a likelihood mapping that would be incorrect with respect to its generating function. + +So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount. + +And it's centered about the relationship we think is the case between the hidden states and the observations. + +So the reason why we have this probability distribution is because we don't know. + +We do not actually ever get to know what the real relationship is between hidden states and observations. + +So does that answer your question, Mark, about why + +why probability distributions, maybe why normal distributions, or did that help at all? + + +SPEAKER_02: +Yeah, that helped a lot. + +Thank you. + +Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right? + +And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right? + + +SPEAKER_05: +uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all + +We could plug, we could say, okay, the relationship is just equal to the hidden state. + +I think that the observation is exactly the hidden state, right? + +That's an assumption. + +Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state. + +Or we can have a quadratic relationship. + +We can have anything we want in there at all, actually. + +And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations. + +Yeah. + + +SPEAKER_04: +very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B, + + +SPEAKER_05: +Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from. + +Michael Morehead , Other questions other questions Magdalena. + + +SPEAKER_01: +Okay, my question is this. + +So could we do a thought experiment for a minute? + +Because I think if I set it up as a thought experiment, I'll get the answer. + +Okay. + +Oh, so I'm inside. + +Okay, so let's imagine the following. + +In present-day societies, we have this societal + +back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists. + +And there's a tension in societies to update that belief by some group to science. + +Exists a science so God exists. + +So science tells us that God may not exist or something to that effect or quantum physics explains reality. + +Not God. + +Okay. + +So there's a process, right? + +That's in in the environment. + +that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed + +And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications. + +for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is + +Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis? + +are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds + +Okay, as a third level of information processing. + +And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis? + + +SPEAKER_05: +Yeah, no, excellent, excellent question, Magdalena. + +This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents. + +Under what conditions do they interact so as to form a larger composite active inference agent? + +So it's kind of a question like, how many active inference agents are there really in any given picture? + +So your question, I think, insofar as I understand it, is, look, we have some notion about agency. + +inactive inference. + +We have some notion about something interacting with its environment, whatever the environment is, right? + +So on the right hand side, we've got an agent, and on the left hand side, we have an environment. + +But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent. + +Okay, we haven't been able to tell that story yet. + +Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers. + +Let me see if I can find that figure. + +Various layers within the predictive coding hierarchy. + +Which one is it here? + +This one? + +No, this one. + +You might interpret the various layers here as individual active implementations that are just doing local free energy minimization. + +So maybe this guy is an active implementation in some sense. + +And then he's passing messages, blah, blah, blah. + +And then the overall thing can be regarded as an active implementation. + +That's the only inkling that we've seen thus far of this idea yet. + +We're also not really going to see a lot of it in the rest of the book. + +because it is a very open question. + +It's a very difficult question. + +And it gets at the heart of what agency even is at all, which is a question over and above active inference per se. + +However, my personal interpretation or my personal thoughts about this issue + +is that the thing that is currently missing from active inference is exactly this idea of compositionality. + +So I've got agent here, agent here. + +They interact. + +Under what conditions is that whole thing one active inference agent? + +That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research. + +So I hope that answers your question in some respect. + +Very, very important question. + +But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding. + + +SPEAKER_01: +OK. + +really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups + +What I see and I'm going to use this language, I'm not a mathematician. + +OK, so forgive me, but but I kind of I kind of get some things in math. + +OK, so topologically generative models really work as huge attractors. + +in informational systems in humans. + +So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right? + +So your language is a technology in human systems. + +so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model + +And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math. + + +SPEAKER_09: +If I just kind of sorry, is that okay? + +All right. + +Thanks. + +Just ask someone who has a. + +More immediate background in this social sciences. + +Yeah. + +So, so. + +And I like how the question that Mark had prior to this had to do with, like. + +You know, what is it to make linear assumptions? + +So, in the context of the textbook, what we've seen thus far, like. + +This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise. + +And it's just to show how a model can be composed. + +Like, how can. + +In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right? + +It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right? + +So that's the action part. + +We haven't reached that yet. + +But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function? + +So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right? + +So that linear model we were looking at + +we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people + +what are the kinds of sensory information that I receive? + +And what are the kinds of actions that I can do? + +And what do I have beliefs about? + +Or what variables do I think explains everything, right? + +So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human + +uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like + +There's, there's a lot to track. + +So there are some people who have tried to model this sort of thing. + +And so I've shared this paper. + +It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference. + +And this kind of has to do with. + +Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere. + +And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing. + +And that's one way of trying to look at a problem. + +like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one + +And a lot of agents who rally around idea two. + +And as they do that, the two groups start to communicate directly with each other less and less. + +The agent's actual actions that are simulated has to do with them communicating with one another and sharing their + +They're kind of more verbal beliefs with one another. + +And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing. + +And somehow that gets all blended in with our assumptions of how science works and things like that. + +It's much more complex. + +And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy + +But remember that free energy, based on how we've discussed it so far, + +It's not some magical quantity of energy or something like that. + +It's really just a proxy. + +It's a word that is a proxy for something called surprisal, which is more or less uncertainty. + +So we want to minimize our uncertainty. + +So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another, + +about is it idea one or is it idea two? + +Is it science or is it something else? + +One way to minimize uncertainty is to repeatedly tell each other + +what the truth is and you all converge on what you all think the truth is. + +Right. + +And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much. + +Because the other group is adding to your uncertainty. + +Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe. + +So, so you see that there, there is a kind of like logic where the free energy principle is involved. + +where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise. + +And that has nothing to do with the true validity of science, right? + +Like I'm attempting to be a scientist explaining all this stuff, + +But the people who, you see what I'm saying? + +Like to take in sensory information is what we do. + +It's not about like following proper logic and stuff. + +It's people learn to do that, right? + + +SPEAKER_01: +Can I interject something real quick? + +I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're + +Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality. + +So what happens in humans groups is that the active inferencing that's happening in a lot of individuals, + +is to analyze and update whether or not they believe the person who's giving the message. + +It's not about evaluating the validity of the premises. + +It's not about the... That gets at a very... + + +SPEAKER_05: +That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer. + +We might imagine an active inference agent doing inference about this. + +But above that is a question, how reliable is this thing that I'm hearing at all? + +and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant + +uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation + +for what it means for one entity to be an active inference agent. + +Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents. + +But that's not going to be the focus of the book. + +A lot of that is at the cutting edge of active inference research now. + + +SPEAKER_09: +Thanks for keeping it closer to the textbook by bringing up precision. + +I very much agree with that. + +And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but. + +This. + +Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference. + +So active inferences is generally setting up these principles. + +It. + +definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different + +Right? + +Like, our variables that we're including in our respective models are different. + +The precisions can definitely differ. + +If I get information from someone in my group, I might have a higher precision and trust what they say more. + +I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right? + +So a lot of the, you know, a lot of the + +political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right? + +It's like, oh, I just lump them all in with political group + +be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available + +then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone + +ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it + + +SPEAKER_05: +Do make sure, I've been absent for a little bit. + +I'm gonna be much more active on the questions side of things on the CODA. + +So do make sure if you've got questions, please do put them down in the questions tab here. + +I've gone through and I've already answered quite a few of them for chapter five. + +uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see + +Are there any more live questions? + +That would be nice. + +We can stay around for maybe five or so minutes after the deadline if people so choose. + +But we are coming up to the top of the hour. + +So yeah, all of the questions in chapter five now are answered, or I've given my attempt. + +As I say, I'll go through and attempt to answer all the previous ones as well. + +think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry + + +SPEAKER_08: +okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian + +inference to engage in parameter learning so that it can update beliefs about the current state. + +If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself. + +And I think this is what we call plasticity. + +The ultimate goal of all this is to minimize expected free energy. + +Is that right? + + +SPEAKER_05: +Wow, okay. + +I mean, basically, yeah, that's substantively correct. + +I will note, however, we haven't yet talked about expected free energy. + +and planning and action selection. + +So that is going to come later. + +But yeah, you're right. + +In terms of the general flavor of things, we have beliefs about hidden states in the world. + +And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization. + +Or we've seen that we can also recast this in terms of predictive processing. + +minimizing the precision way to prediction errors. + +It's the same kind of story. + +And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states. + +Yes. + +The other thing, you mentioned parameter learning. + +So yes, exactly. + +There's these two problems. + +We have the problem of inferring the hidden states. + +What should the hidden states be? + +But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference. + +So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like. + +But then above that, we need to deal with the problem of what should the setting of the parameters be? + +And we've only really just begun to look at this in terms of prediction errors and hierarchical models. + +But we saw + +how to deal with that in terms of expectation maximization and earlier chapters. + +We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture. + +Because we generally can't do expectation maximization for the reason that we don't know the exact posterior. + +So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah. + + +SPEAKER_08: +Thank you. + + +SPEAKER_05: +No problem. + +Do we have more questions? + +More questions? + +Maybe one more question, if there is one more. + +And then I'll stop the recording. + +All right. + +I might stop the recording here. + +Well, Mark. + +Hello, everyone, and welcome. + +I've got my esteemed friend. + +Oh, hang on. + + +SPEAKER_02: +That was interesting audio feedback. + +I thought I would jump in since nobody else asked us. + +But earlier you mentioned using gradient descent to learn parameters, right? + +Is that also used in perception? + +Is it the same approach? + + +SPEAKER_05: +Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things. + +So it's a very, very general + +optimization techniques. + +So yes, it certainly can be used for learning, well, for perception. + +I'm trying to find a nice picture here. + +Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent. + +Let me just make sure of that. + +Yes, in other words, is the answer. + +It's used for a lot of things. + +Very, very general technique. + +Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six. + +Absolutely. + +Yeah. + +We're gonna talk about phase planes and so on. + +Yeah. + +There are problems with it, but yes, we're gonna see it going forward. + +It's a very, it's a very foundational technique. + +Yeah. + +All right. + +Any, any last minute questions for the YouTube recording before we, uh, + +If not, I think I'm sorry. + + +SPEAKER_07: +Can I ask a quick question? + + +SPEAKER_05: +Yeah, sure. + + +SPEAKER_07: +I do apologize. + +I have joined the group very late. + +So maybe this was addressed in like the past weeks. + +But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning? + + +SPEAKER_05: +well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms + +If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control. + +There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference. + +So the answer is yes, there is a profound relationship. + +We probably don't have time to get into it here. + +One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain. + +Because we've seen with, well, we haven't yet seen with expected free energy how that works. + +But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning. + +And we don't just have this scale of reward, this signal in reinforcement learning. + +we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um + + +SPEAKER_09: +a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning. + +Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning, + +Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two. + +But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven. + +So there can be like a KL control agent or there can be a variety. + +other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that + +that things like curiosity exist. + +Why does curiosity exist if we're actually always just driven towards a reward? + +Once you know what the reward is, you should just go for that, right? + +And you would have no reason to find some other strategy for going for it. + +You would have no reason to go out of your way. + +From what you already know, you just keep going for the reward as much as possible. + +Perhaps you learn other things through happenstance along the way. + + +SPEAKER_05: +That's a big difference. + +In reinforcement, you can just sort of start maximizing a reward signal. + +But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards. + +You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards. + + +SPEAKER_09: +yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite + +how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning. + +Meanwhile, active inference has a much more principled way that relates to this notion of + +free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually + +Um, when I play a sport, I throw the ball this way. + +Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person. + +but maybe I will kind of curve it or change it, right? + +You're not randomly throwing it in the sky or in the opposite direction. + +You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right? + +There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better. + +based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about + +How do I perform the best whenever it comes to active inference? + +It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them. + +But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear. + + +SPEAKER_05: +Yeah. + +I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys. + +So Giancuomo, fire away. + + +SPEAKER_00: +Very quickly, I think it ties in with what was just spoken and talked about now. + +Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is? + +Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty. + +They say, this is the answer. + +And you're like, no, no, no, it's not. + +And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware. + +Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind? + + +SPEAKER_05: +well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions + +This is built in from the start. + +So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right? + +And okay, we can't do Bayes' rule exactly. + +We have to do it approximately. + +We do variational inference. + +We do prediction error and minimization, that kind of thing. + +But we're always, always, always reasoning about probability distribution. + +So yes, every single active implementation ever is always + +predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet + +And there's issues around that and how to do that. + +But we haven't seen that just yet. + +So, yes, absolutely. + +It's part and parcel of what it is to be an active infatuation. + +It's the reason about uncertainty now. + +Okay, thank you. + +No problem. + +And then, Mark, and then I think we can do... Oh, hang on. + +Maybe just really quickly, did you have a comment, Andrew? + + +SPEAKER_09: +I also really have to go, but yeah, just briefly, you basically answered it. + +Sorry, I was looking at the chat. + +It's just, yeah, it depends on how the question is being asked. + +Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are? + +Like that's going to take a little bit more than just like only looking at the simplified models. + +We're looking at the textbook here. + +um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization + +of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots. + +There are 2 possible things. + +It could be is completely 50 50 is fully uncertain. + +Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision. + +Right? + +Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob. + +Right? + +The more you turn up precision, the more you're. + +you're going to take into account the prediction errors you're receiving and vice versa. + +So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving. + +Right. + +So so uncertainty is very much like throughout. + +I mean, it's a very ubiquitous term through many parts of active inference. + +Yeah. + +So it's essential. + + +SPEAKER_03: +So all right. + +so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see + + +SPEAKER_09: +right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh + +who cares about priors. + +And that's what we see with maximum likelihood estimation in chapter two there. + +But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something. + +It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like, + +thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um + +you know, how those connect up with each other. + +So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there. + +And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe. + +And so I fully trust it without question. + +Right. + +So I say all these things because, uh, + +Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe. + +And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things. + +Yeah, it's very interesting to think about. + + +SPEAKER_05: +Yeah, do take a look at, I think I'm sharing, but figure 2.11. + +This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief. + +So if the precision is really tight around the prior, + +the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model. + +So yeah, do take a look at that. + +That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have. + + +SPEAKER_03: +Yeah. + +Excellent. + +Thank you. + + +SPEAKER_05: +Cool. + +All right, guys, we're going to have to call it there. + +Do put questions in the chats on the page, the code page. + +I'll be a bit more attentive going forward. + +I look forward to next week. + +I'll be doing another session of this kind on Friday for the people who usually attend that session. + +So thank you very much. + +Stop sharing and stop the recording. + +Have we got a recording here? + +Okay, stop the recording. + +Goodbye, YouTube people. + +Until next time. diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.json new file mode 100644 index 000000000..5412d0868 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.json @@ -0,0 +1 @@ +[{"video_id": "FtGnkRJwk6U", "segments": [{"start": 3.187, "end": 28.722, "text": " all right hello everyone so we're here it says july 3rd 2026 we're doing the um part one review basically the fundamentals of active inference uh so we've gone through chapters one through five that uh constitute the first half of the book part one there's three parts uh so today's really kind of a session opportunity to ask questions about the entirety of part one anything that anyone would like cleared up or elaborated upon", "speaker": "SPEAKER_02"}, {"start": 29.849, "end": 31.371, "text": " before we move into part two.", "speaker": "SPEAKER_02"}, {"start": 31.432, "end": 40.827, "text": "And that's going to be, part two is going to be active inference proper, where we actually begin to look at action, how action is selected, how we plan, how we do, quote unquote, active inference.", "speaker": "SPEAKER_02"}, {"start": 41.808, "end": 42.97, "text": "So that's kind of what we're doing today.", "speaker": "SPEAKER_02"}, {"start": 43.771, "end": 45.234, "text": "I shall share my screen.", "speaker": "SPEAKER_02"}, {"start": 45.274, "end": 53.888, "text": "I'll give, I think, maybe just a little bit of review on what we've done, where we've gone, where we are.", "speaker": "SPEAKER_02"}, {"start": 54.108, "end": 56.532, "text": "And then I'll open up the floor for questions.", "speaker": "SPEAKER_02"}, {"start": 56.714, "end": 59.237, "text": " from you guys for anything at all related to part one.", "speaker": "SPEAKER_02"}, {"start": 60.118, "end": 67.466, "text": "And hopefully that will kind of set us up to being able to begin to see where we're going to have to go for part two.", "speaker": "SPEAKER_02"}, {"start": 68.808, "end": 73.413, "text": "So I guess very basically, we've seen this picture here.", "speaker": "SPEAKER_02"}, {"start": 73.433, "end": 76.717, "text": "This is, I think, the first figure in the book from the introduction.", "speaker": "SPEAKER_02"}, {"start": 77.177, "end": 78.619, "text": "This is the outline of the entire book.", "speaker": "SPEAKER_02"}, {"start": 78.799, "end": 81.342, "text": "And again, hopefully everyone can see my screen.", "speaker": "SPEAKER_02"}, {"start": 81.362, "end": 84.105, "text": "Please do make a lot of noise if you can't.", "speaker": "SPEAKER_02"}, {"start": 84.372, "end": 85.853, "text": " So part one, chapters one to five.", "speaker": "SPEAKER_02"}, {"start": 87.035, "end": 91.219, "text": "This is, you know, the book itself is called The Fundamentals of Active Inference.", "speaker": "SPEAKER_02"}, {"start": 91.239, "end": 93.381, "text": "But part one is really the fundamentals of the fundamentals.", "speaker": "SPEAKER_02"}, {"start": 93.401, "end": 105.613, "text": "What do you even need to understand in terms of the constituent mathematics and formalisms and ideas that surround active inference and where it comes from in order to be able to understand active inference itself?", "speaker": "SPEAKER_02"}, {"start": 105.673, "end": 106.634, "text": "So that's part one.", "speaker": "SPEAKER_02"}, {"start": 107.795, "end": 110.398, "text": "I'll give a brief overview of each chapter and the core ideas.", "speaker": "SPEAKER_02"}, {"start": 110.458, "end": 114.362, "text": "So what we had to do initially,", "speaker": "SPEAKER_02"}, {"start": 114.612, "end": 117.795, "text": " is we had to come to part one here.", "speaker": "SPEAKER_02"}, {"start": 117.835, "end": 125.401, "text": "The entirety of part one is completely concerned with the problem of perception only.", "speaker": "SPEAKER_02"}, {"start": 125.441, "end": 128.944, "text": "So with active infants, we have this fundamental notion of perception and action.", "speaker": "SPEAKER_02"}, {"start": 128.964, "end": 132.948, "text": "But we're just going to completely assume actions, and we're going to look at perception only.", "speaker": "SPEAKER_02"}, {"start": 133.088, "end": 137.472, "text": "So we're going to have to try and formalize what it means to be doing perception.", "speaker": "SPEAKER_02"}, {"start": 137.632, "end": 139.774, "text": "And that's the entirety of part one, really.", "speaker": "SPEAKER_02"}, {"start": 140.254, "end": 143.337, "text": "And that's going to carry through to the remainder of the book as well.", "speaker": "SPEAKER_02"}, {"start": 143.772, "end": 171.433, "text": " so chapter one this was all about uh perception as inference okay so the whole point of chapter one was that we have this question what is perception how do we do perception and how are we going to think about perception in the context of the modeling paradigm that is active inference and the idea fundamentally is that perception is inference we have beliefs about hidden states of the world uh we do not directly observe hidden states we only observe", "speaker": "SPEAKER_02"}, {"start": 171.413, "end": 173.496, "text": " through our sensory apparatus.", "speaker": "SPEAKER_02"}, {"start": 173.997, "end": 180.446, "text": "And on the basis of those operations, we can update our beliefs about hidden sensory, well, hidden states of the world.", "speaker": "SPEAKER_02"}, {"start": 180.886, "end": 181.908, "text": "And that process is inference.", "speaker": "SPEAKER_02"}, {"start": 181.928, "end": 184.652, "text": "So inference is just the process of changing your beliefs about something, right?", "speaker": "SPEAKER_02"}, {"start": 185.633, "end": 188.718, "text": "And that's what we're going to mean by perception.", "speaker": "SPEAKER_02"}, {"start": 189.559, "end": 190.941, "text": "And there's a whole story to be woven here.", "speaker": "SPEAKER_02"}, {"start": 191.181, "end": 195.587, "text": "Well, before we get to that story, the most fundamental distinction is this distinction here.", "speaker": "SPEAKER_02"}, {"start": 196.929, "end": 197.25, "text": "Where is it?", "speaker": "SPEAKER_02"}, {"start": 199.893, "end": 200.374, "text": "Yeah, this one.", "speaker": "SPEAKER_02"}, {"start": 201.518, "end": 204.882, "text": " This, again and again and again, is the most fundamental thing.", "speaker": "SPEAKER_02"}, {"start": 204.902, "end": 208.747, "text": "And we're going to see something like this diagram repeat throughout the book.", "speaker": "SPEAKER_02"}, {"start": 209.367, "end": 213.212, "text": "We've already seen it repeat throughout the chapters that we've gone through so far.", "speaker": "SPEAKER_02"}, {"start": 213.272, "end": 227.008, "text": "So the fundamental separation, the really bedrock starting point, is this idea that we have the environment, which is in some sense the real world, and we have the agent, which is somehow contained in the real world but not identical to it.", "speaker": "SPEAKER_02"}, {"start": 227.809, "end": 229.551, "text": "And they are separated from one another.", "speaker": "SPEAKER_02"}, {"start": 230.054, "end": 237.193, "text": " Okay, so there's some process out there, there's some dynamics, there's some physics, whatever that's quote unquote out there.", "speaker": "SPEAKER_02"}, {"start": 237.233, "end": 243.45, "text": "And the agent is trying to predict what is going on in the environment.", "speaker": "SPEAKER_02"}, {"start": 243.47, "end": 246.358, "text": "And it can only do that on the basis of observations.", "speaker": "SPEAKER_02"}, {"start": 246.338, "end": 268.043, "text": " so observations come in and never the agent never gets access to the real world the real environment but on the basis of those observations the agent has a model of the real world and using that model it's able to update predictions or beliefs about the real environment okay and those predictions and beliefs we see in chapter two", "speaker": "SPEAKER_02"}, {"start": 269.744, "end": 275.82, "text": " we formulate this idea of perception as hidden state inference using Bayes' rule.", "speaker": "SPEAKER_02"}, {"start": 276.321, "end": 278.386, "text": "So chapter one was all about what is perception?", "speaker": "SPEAKER_02"}, {"start": 278.427, "end": 279.77, "text": "It's this process of inference.", "speaker": "SPEAKER_02"}, {"start": 280.632, "end": 284.783, "text": "And we're doing inference because we're separated from our environment in the sense that", "speaker": "SPEAKER_02"}, {"start": 284.999, "end": 314.348, "text": " there's a boundary between the agent and the environment and chapter two that makes this precise mathematically okay so we actually start doing some mathematical operations here in chapter two so we we know we're doing inference but specifically chapter two says we're going to be doing Bayesian inference okay so and the the recurring the recurring you know running example throughout this book is this really really really simple example where we have some notion of an agent that's farmed for food", "speaker": "SPEAKER_02"}, {"start": 315.475, "end": 320.462, "text": " So the idea is that some food out there in the environment has a certain size.", "speaker": "SPEAKER_02"}, {"start": 321.303, "end": 322.965, "text": "That's the state of the real world.", "speaker": "SPEAKER_02"}, {"start": 323.806, "end": 324.647, "text": "I'll get this picture up here.", "speaker": "SPEAKER_02"}, {"start": 326.911, "end": 330.736, "text": "So the state of the real world is out here in the environment, there's some state.", "speaker": "SPEAKER_02"}, {"start": 330.916, "end": 334.982, "text": "It's food sizes and arbitrary food sizes from 1 to 5, let's say.", "speaker": "SPEAKER_02"}, {"start": 335.843, "end": 337.465, "text": "And those generate observations.", "speaker": "SPEAKER_02"}, {"start": 337.845, "end": 339.908, "text": "And the observations are light intensities.", "speaker": "SPEAKER_02"}, {"start": 341.29, "end": 344.875, "text": "So the observations are related to the hidden states in some way.", "speaker": "SPEAKER_02"}, {"start": 345.918, "end": 351.365, "text": " We don't actually know really how the observations are related to the hidden states.", "speaker": "SPEAKER_02"}, {"start": 352.987, "end": 356.011, "text": "But what we say is we're going to assume that there is some kind of relationship.", "speaker": "SPEAKER_02"}, {"start": 356.892, "end": 367.526, "text": "And our assumptions about the relationship between the hidden states and the observations, between the food sizes and the light intensities, those assumptions are going to be our generative model.", "speaker": "SPEAKER_02"}, {"start": 369.328, "end": 372.352, "text": "And we're explicitly reasoning probabilistically.", "speaker": "SPEAKER_02"}, {"start": 372.472, "end": 374.895, "text": "So the entire North Star", "speaker": "SPEAKER_02"}, {"start": 375.432, "end": 389.928, "text": " the whole sort of guiding apparatus of what we're doing is we're trying to, as best we can, implement Bayes' rule, which is the rule that tells us how to update our beliefs about hidden states on the basis of sensory evidence.", "speaker": "SPEAKER_02"}, {"start": 389.948, "end": 392.211, "text": "And Bayes' rule tells us this precisely and exactly.", "speaker": "SPEAKER_02"}, {"start": 393.432, "end": 393.993, "text": "So what do you need?", "speaker": "SPEAKER_02"}, {"start": 394.073, "end": 401.541, "text": "You need a generative model that encodes our assumptions about the relationships between hidden states X and observations Y.", "speaker": "SPEAKER_02"}, {"start": 402.348, "end": 408.518, "text": " And we've seen that we can factorize this generative model in terms of our likelihood.", "speaker": "SPEAKER_02"}, {"start": 409.479, "end": 414.687, "text": "So we have a probability distribution of our observations given hidden states and prior beliefs about hidden states.", "speaker": "SPEAKER_02"}, {"start": 414.727, "end": 417.492, "text": "Those two things together constitute our generative model.", "speaker": "SPEAKER_02"}, {"start": 418.573, "end": 421.538, "text": "And then the last piece, the missing ingredient that we need", "speaker": "SPEAKER_02"}, {"start": 421.838, "end": 423.802, "text": " is our model evidence.", "speaker": "SPEAKER_02"}, {"start": 424.343, "end": 428.851, "text": "This is the probability of observing something independent of any state.", "speaker": "SPEAKER_02"}, {"start": 429.552, "end": 432.137, "text": "And those three things together, they compose Bayes' rule.", "speaker": "SPEAKER_02"}, {"start": 432.858, "end": 436.685, "text": "And then we can get our posterior belief about hidden states on the basis of sensory evidence.", "speaker": "SPEAKER_02"}, {"start": 436.785, "end": 439.43, "text": "And then bingo, bango, bongo, we're doing inference.", "speaker": "SPEAKER_02"}, {"start": 439.731, "end": 440.131, "text": "Very nice.", "speaker": "SPEAKER_02"}, {"start": 441.314, "end": 444.319, "text": "So that fills in the picture of what's going on", "speaker": "SPEAKER_02"}, {"start": 444.721, "end": 449.467, "text": " at least abstractly, mathematically, ideally, inside the agent.", "speaker": "SPEAKER_02"}, {"start": 449.487, "end": 452.651, "text": "We're doing inference by means of Bayes' rule, or at least we're trying to do that.", "speaker": "SPEAKER_02"}, {"start": 452.671, "end": 453.992, "text": "And indeed, that's what we did in chapter two.", "speaker": "SPEAKER_02"}, {"start": 454.032, "end": 456.015, "text": "We used Bayes' rule exactly.", "speaker": "SPEAKER_02"}, {"start": 456.555, "end": 460.62, "text": "We made certain assumptions about how we were going to be doing things.", "speaker": "SPEAKER_02"}, {"start": 460.66, "end": 463.984, "text": "We discretized the state space.", "speaker": "SPEAKER_02"}, {"start": 464.164, "end": 465.886, "text": "We were able to do exact marginalization.", "speaker": "SPEAKER_02"}, {"start": 466.507, "end": 470.492, "text": "We waved our hands a lot, and we got rid of a lot of problems.", "speaker": "SPEAKER_02"}, {"start": 470.978, "end": 472.419, "text": " I see there's some people in the waiting room still.", "speaker": "SPEAKER_02"}, {"start": 472.539, "end": 474.601, "text": "Let me just admit some people.", "speaker": "SPEAKER_02"}, {"start": 475.802, "end": 477.684, "text": "But that was crucial.", "speaker": "SPEAKER_02"}, {"start": 478.245, "end": 480.446, "text": "We assumed that we could do exact basis rule.", "speaker": "SPEAKER_02"}, {"start": 483.229, "end": 484.35, "text": "Do I have the waiting room?", "speaker": "SPEAKER_02"}, {"start": 485.191, "end": 486.432, "text": "Let's have a see here, participants.", "speaker": "SPEAKER_02"}, {"start": 487.673, "end": 488.614, "text": "Admit, admit, admit.", "speaker": "SPEAKER_02"}, {"start": 488.714, "end": 493.818, "text": "All right, very good, very good.", "speaker": "SPEAKER_02"}, {"start": 496.801, "end": 497.742, "text": "Excellent, all right.", "speaker": "SPEAKER_02"}, {"start": 497.762, "end": 498.943, "text": "I shall share my screen again.", "speaker": "SPEAKER_02"}, {"start": 505.454, "end": 506.095, "text": " Very nice.", "speaker": "SPEAKER_02"}, {"start": 506.115, "end": 512.32, "text": "So we have this notion that what we're doing is we're trying to implement Bayes' rule, at least ideally.", "speaker": "SPEAKER_02"}, {"start": 513.101, "end": 518.366, "text": "That is what it is to be doing inference inside the agent, by means of our model.", "speaker": "SPEAKER_02"}, {"start": 518.386, "end": 525.813, "text": "And then out here in the real world, in the general process, there is some actual relationship between observations and hidden states.", "speaker": "SPEAKER_02"}, {"start": 526.293, "end": 528.155, "text": "We never know this exactly.", "speaker": "SPEAKER_02"}, {"start": 528.956, "end": 532.018, "text": "But of course, as modelers, we're in the unique position of being able to specify this.", "speaker": "SPEAKER_02"}, {"start": 532.078, "end": 534.921, "text": "So for our really trivial example,", "speaker": "SPEAKER_02"}, {"start": 535.137, "end": 543.066, "text": " we said, okay, and typically throughout all of chapter, well, part one, the relationship between hidden states and observations has been linear.", "speaker": "SPEAKER_02"}, {"start": 544.147, "end": 546.61, "text": "Okay, that's one kind of relationship, very basic kind of relationship.", "speaker": "SPEAKER_02"}, {"start": 547.43, "end": 551.875, "text": "And that's been the case in the model and in the process.", "speaker": "SPEAKER_02"}, {"start": 553.217, "end": 558.943, "text": "So that was chapter two, where we actually sort of, you know, rubber meets the road.", "speaker": "SPEAKER_02"}, {"start": 559.003, "end": 560.244, "text": "We did some real calculations.", "speaker": "SPEAKER_02"}, {"start": 560.304, "end": 564.469, "text": "We saw what it is to update beliefs by means of Bayes' rule.", "speaker": "SPEAKER_02"}, {"start": 564.955, "end": 588.229, "text": " okay and i should note just by and by the by that what we're doing we haven't made a big song and dance about this yet but all of our beliefs have been with respect to or by means of continuous probability distributions so these these bell curves okay gaussian distributions um that's and we're going to continue this in part two", "speaker": "SPEAKER_02"}, {"start": 588.209, "end": 600.422, "text": " and going on, but we're also going to see a distinction that will come up later when we put action in for simple models where instead of reasoning about these bell curves, we're going to reason about categorical distributions.", "speaker": "SPEAKER_02"}, {"start": 601.263, "end": 606.128, "text": "Instead of a continuous interval, we're going to be dealing with discrete things, but that's going to come later.", "speaker": "SPEAKER_02"}, {"start": 607.83, "end": 612.355, "text": "The big takeaway from Chapter 2 is this here, I think at least.", "speaker": "SPEAKER_02"}, {"start": 613.556, "end": 617.22, "text": "We're reasoning about things in terms of Bayes' rule,", "speaker": "SPEAKER_02"}, {"start": 617.74, "end": 621.203, "text": " We're updating the posterior over hidden states on the basis of sensory evidence.", "speaker": "SPEAKER_02"}, {"start": 621.363, "end": 621.724, "text": "Very nice.", "speaker": "SPEAKER_02"}, {"start": 622.825, "end": 624.987, "text": "And we're doing this exactly.", "speaker": "SPEAKER_02"}, {"start": 625.027, "end": 626.648, "text": "We're doing exact Bayesian inference in chapter two.", "speaker": "SPEAKER_02"}, {"start": 627.549, "end": 636.197, "text": "And the big thing that we saw is that if we do this, there are certain relationships that hold between the prior, the likelihood, and the posterior.", "speaker": "SPEAKER_02"}, {"start": 638.579, "end": 647.427, "text": "So what we saw is that depending on the precision, the degree of spread or uncertainty that you have,", "speaker": "SPEAKER_02"}, {"start": 647.693, "end": 673.009, "text": " about your prior and your likelihood that affects the precision or uncertainty of your posterior beliefs so and you know canonically you can sort of mix and match but imagine you've got a very precise prior belief okay and a kind of spread out likelihood of observations this in some sense encodes an agent that's very stubborn one that doesn't really want to", "speaker": "SPEAKER_02"}, {"start": 673.647, "end": 675.17, "text": " its observation model very much.", "speaker": "SPEAKER_02"}, {"start": 675.891, "end": 677.574, "text": "It wants to trust its prior beliefs.", "speaker": "SPEAKER_02"}, {"start": 678.014, "end": 687.029, "text": "And so you see, OK, well, if we do a Bayesian update with a model that looks like this, the posterior that we get is pretty close to our prior belief already.", "speaker": "SPEAKER_02"}, {"start": 688.311, "end": 694.922, "text": "So that's encoding an agent that's stubborn or that doesn't want to change its beliefs very much, kind of a closed-minded agent.", "speaker": "SPEAKER_02"}, {"start": 695.704, "end": 699.57, "text": "On the complete other side of the spectrum, if your precision", "speaker": "SPEAKER_02"}, {"start": 699.82, "end": 729.051, "text": " for your your prior is quite spread out so the agent doesn't really have a very fixed idea about what it's what the real world is going to be like but it's going to it's very precisely attending to sensory evidence then when we do a bayesian update you can see that the posterior is much much more like the uh the likelihood model so it's trusting the observations it's getting in more than it's trusting its prior beliefs okay and it's precisely this precision", "speaker": "SPEAKER_02"}, {"start": 729.385, "end": 755.241, "text": " or variance or uncertainty on either the prior or the likelihood and how they how they relate with each other that affect the precision or the uncertainty of the posterior belief so that's something that you get with basis rule you start being able to reason about the uncertainties um in our beliefs our prior beliefs and like well yeah is that a question i hear oh well", "speaker": "SPEAKER_02"}, {"start": 758.562, "end": 759.803, "text": " I hear some grumbling.", "speaker": "SPEAKER_02"}, {"start": 760.484, "end": 762.906, "text": "Hang on, stop sharing.", "speaker": "SPEAKER_02"}, {"start": 763.126, "end": 763.987, "text": "Was there a question?", "speaker": "SPEAKER_02"}, {"start": 767.03, "end": 767.811, "text": "Going once, going twice.", "speaker": "SPEAKER_02"}, {"start": 767.831, "end": 768.472, "text": "Sorry, my bad.", "speaker": "SPEAKER_02"}, {"start": 768.592, "end": 770.774, "text": "I thought I was on mute.", "speaker": "SPEAKER_02"}, {"start": 771.795, "end": 773.276, "text": "Okay, no problem.", "speaker": "SPEAKER_02"}, {"start": 775.338, "end": 776.159, "text": "All right, I'll just share.", "speaker": "SPEAKER_02"}, {"start": 776.459, "end": 783.206, "text": "Yeah, I just want to briefly go through just so that we're all on the same page, at least at a high level what we've sort of gone through.", "speaker": "SPEAKER_02"}, {"start": 784.347, "end": 784.988, "text": "So that was chapter two.", "speaker": "SPEAKER_02"}, {"start": 785.228, "end": 788.451, "text": "Chapter three then was a bit of a change of pace.", "speaker": "SPEAKER_02"}, {"start": 789.241, "end": 791.705, "text": " The idea is that, yes, very nice.", "speaker": "SPEAKER_02"}, {"start": 791.745, "end": 797.654, "text": "We've seen how we can operationalize this procedure of inference for perception.", "speaker": "SPEAKER_02"}, {"start": 798.615, "end": 804.624, "text": "OK, but there were assumptions that we made in chapter two, namely that we have a model.", "speaker": "SPEAKER_02"}, {"start": 805.005, "end": 806.668, "text": "OK, and we're going to use that model to do inference.", "speaker": "SPEAKER_02"}, {"start": 806.808, "end": 810.253, "text": "And we saw lots of different ways to do inference in chapter two.", "speaker": "SPEAKER_02"}, {"start": 810.706, "end": 813.851, "text": " But of course, and I want to just show you, where is it here?", "speaker": "SPEAKER_02"}, {"start": 813.971, "end": 817.117, "text": "So we go, this is this very central diagram here.", "speaker": "SPEAKER_02"}, {"start": 817.137, "end": 819.32, "text": "So we have a model and we did inference with our model.", "speaker": "SPEAKER_02"}, {"start": 820.322, "end": 831.681, "text": "Of course, our model is, you know, there are certain settings of parameters of our model that need to be set well, if we're going to do inference well.", "speaker": "SPEAKER_02"}, {"start": 831.982, "end": 832.242, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 832.745, "end": 835.468, "text": " So our model is made out of probability distributions.", "speaker": "SPEAKER_02"}, {"start": 835.548, "end": 837.21, "text": "We have our likelihood, we have our prior.", "speaker": "SPEAKER_02"}, {"start": 837.83, "end": 838.931, "text": "These have parameters.", "speaker": "SPEAKER_02"}, {"start": 839.552, "end": 840.773, "text": "These have variances.", "speaker": "SPEAKER_02"}, {"start": 841.814, "end": 844.817, "text": "Our observation model has parameters theta.", "speaker": "SPEAKER_02"}, {"start": 846.139, "end": 851.484, "text": "Because it's a linear model, typically, we have a slope and an intercept.", "speaker": "SPEAKER_02"}, {"start": 851.965, "end": 852.745, "text": "Those are parameters.", "speaker": "SPEAKER_02"}, {"start": 852.785, "end": 859.292, "text": "Those are things that have to be set well enough so that when we do inference with our model, we're going to be doing inference well.", "speaker": "SPEAKER_02"}, {"start": 860.2, "end": 862.722, "text": " So there's problem number one, how do you do inference?", "speaker": "SPEAKER_02"}, {"start": 863.243, "end": 866.466, "text": "But above that is problem number two, and it's kind of a harder problem.", "speaker": "SPEAKER_02"}, {"start": 866.486, "end": 868.888, "text": "What should the setting of the parameters be for my model?", "speaker": "SPEAKER_02"}, {"start": 870.31, "end": 872.112, "text": "In chapter two, we just completely ignored that problem.", "speaker": "SPEAKER_02"}, {"start": 872.192, "end": 879.779, "text": "We assumed that the parameters of our model, the settings of the dials of our model were already perfect.", "speaker": "SPEAKER_02"}, {"start": 880.119, "end": 882.962, "text": "They were set to be excellent, and then we can just do inference.", "speaker": "SPEAKER_02"}, {"start": 883.283, "end": 884.003, "text": "And we told that story.", "speaker": "SPEAKER_02"}, {"start": 885.445, "end": 888.908, "text": "But realistically, in the real world, your model has parameters and you need to know", "speaker": "SPEAKER_02"}, {"start": 889.175, "end": 890.256, "text": " how to set those parameters.", "speaker": "SPEAKER_02"}, {"start": 891.858, "end": 893.88, "text": "And we're going to see this more and more going forward.", "speaker": "SPEAKER_02"}, {"start": 893.9, "end": 904.65, "text": "But this question about how to set the parameters of your model, this is typically operationalized as what it is to be doing learning in active inference.", "speaker": "SPEAKER_02"}, {"start": 905.651, "end": 909.975, "text": "So we've got sort of fast changing things down here, hidden states.", "speaker": "SPEAKER_02"}, {"start": 910.055, "end": 910.876, "text": "How are those changing?", "speaker": "SPEAKER_02"}, {"start": 910.896, "end": 911.497, "text": "That's inference.", "speaker": "SPEAKER_02"}, {"start": 912.398, "end": 918.844, "text": "But then above that is the question, all right, how do I learn or how do I do inference on what the parameters of my model should be?", "speaker": "SPEAKER_02"}, {"start": 919.516, "end": 923.359, "text": " So they're two different but related problems, basically.", "speaker": "SPEAKER_02"}, {"start": 923.379, "end": 927.884, "text": "And realistically, you need to be able to do both of them if you're going to be doing inference well.", "speaker": "SPEAKER_02"}, {"start": 929.405, "end": 934.79, "text": "So chapter 3 was all about different ways of estimating model parameters.", "speaker": "SPEAKER_02"}, {"start": 935.791, "end": 939.394, "text": "And we saw how to do deterministic model parameter estimation.", "speaker": "SPEAKER_02"}, {"start": 939.454, "end": 941.556, "text": "We saw techniques like gradient descent.", "speaker": "SPEAKER_02"}, {"start": 942.377, "end": 943.498, "text": "It comes up again and again and again.", "speaker": "SPEAKER_02"}, {"start": 943.538, "end": 949.523, "text": "Maximum likelihood estimation, Bayesian linear regression, all very nice.", "speaker": "SPEAKER_02"}, {"start": 949.908, "end": 955.876, "text": " This culminated in something called the expectation maximization algorithm.", "speaker": "SPEAKER_02"}, {"start": 955.916, "end": 958.099, "text": "That was the jewel of Chapter 3.", "speaker": "SPEAKER_02"}, {"start": 959.961, "end": 961.383, "text": "Why is that important?", "speaker": "SPEAKER_02"}, {"start": 961.483, "end": 963.385, "text": "Why did we care about the model premise?", "speaker": "SPEAKER_02"}, {"start": 963.405, "end": 966.269, "text": "Just briefly before I answer that question.", "speaker": "SPEAKER_02"}, {"start": 967.471, "end": 976.422, "text": "In some sense, this figure here, this is a very nice depiction of the process of parameter learning that we saw in Chapter 3.", "speaker": "SPEAKER_02"}, {"start": 976.925, "end": 1001.785, "text": " broadly speaking and we're looking at a process of gradient descent here and gradient descent is a very fundamental algorithm that shows up everywhere in machine learning um and it's going to continually recur across the textbook the general idea is let's say i've got two parameters okay for my observation likelihood in this case i've got beta zero and beta one you know you could have hundreds of parameters or thousands of parameters but let's say we've got two", "speaker": "SPEAKER_02"}, {"start": 1002.322, "end": 1006.466, "text": " And we have some notion about how costly those parameters are.", "speaker": "SPEAKER_02"}, {"start": 1006.746, "end": 1010.97, "text": "So these two parameters here, they sweep out a two-dimensional space.", "speaker": "SPEAKER_02"}, {"start": 1011.97, "end": 1014.753, "text": "So a point in this space is a setting of beta 1 and beta 2.", "speaker": "SPEAKER_02"}, {"start": 1015.774, "end": 1023.701, "text": "And this surface, this curve, is some kind of notion about how costly those pairs of parameters are.", "speaker": "SPEAKER_02"}, {"start": 1025.262, "end": 1032.008, "text": "So what we can do is if we can operationalize a notion of cost,", "speaker": "SPEAKER_02"}, {"start": 1032.376, "end": 1043.612, "text": " We can start somewhere on this surface, and we can say, all right, let's just walk down the surface in the direction of greatest steepness, right?", "speaker": "SPEAKER_02"}, {"start": 1044.574, "end": 1048.68, "text": "And we're just going to keep doing that for some number of steps or iterations.", "speaker": "SPEAKER_02"}, {"start": 1049.822, "end": 1056.932, "text": "And once you get to somewhere that's sort of vaguely flat, a local minimum, maybe a global minimum, but typically a local minimum,", "speaker": "SPEAKER_02"}, {"start": 1057.587, "end": 1058.968, "text": " We're going to say, all right, well, you know what?", "speaker": "SPEAKER_02"}, {"start": 1059.389, "end": 1063.953, "text": "The setting of those parameters where the cost is basically flat, those are going to be good.", "speaker": "SPEAKER_02"}, {"start": 1064.954, "end": 1067.196, "text": "And those are the parameters that we're going to use for inference.", "speaker": "SPEAKER_02"}, {"start": 1068.297, "end": 1073.421, "text": "That is, in so many words, the basis of gradient descent.", "speaker": "SPEAKER_02"}, {"start": 1073.521, "end": 1075.623, "text": "And that's going to show up again and again and again.", "speaker": "SPEAKER_02"}, {"start": 1076.744, "end": 1080.148, "text": "But it's different from the problem of inference.", "speaker": "SPEAKER_02"}, {"start": 1081.329, "end": 1086.193, "text": "However, there are interpretations about what it means to learn.", "speaker": "SPEAKER_02"}, {"start": 1086.933, "end": 1113.227, "text": " such that we can treat parameters as themselves another kind of hidden state and we can do inference on those okay so that that idea is going to show up in chapter five at least in disguise when it comes to this idea of separation of temporal scales so you've got fast changing states down here and slower changing states up here and these are your parameters and they're coupled to one another all right so the last thing on chapter three", "speaker": "SPEAKER_02"}, {"start": 1114.608, "end": 1142.505, "text": " very nice that's a nice story about how to do parameter learning but you need to do parameter learning and inference kind of at the same time right basically and there's there are chicken and an egg okay you need you need to be able to do inference in order to be able to learn parameters but you need the good parameters in order to be able to do inference so it's it's a very vicious problem and the expectation maximization algorithm uh is the thing that solves this chicken and egg problem where", "speaker": "SPEAKER_02"}, {"start": 1142.958, "end": 1149.467, "text": " We assume that we can take the expectation about what parameters should be given beliefs about hidden states.", "speaker": "SPEAKER_02"}, {"start": 1149.748, "end": 1156.117, "text": "And then on the basis of that, we can maximize the evidence for certain parameters and we can kind of do them jointly.", "speaker": "SPEAKER_02"}, {"start": 1156.137, "end": 1157.098, "text": "And that solves the problem.", "speaker": "SPEAKER_02"}, {"start": 1157.94, "end": 1165.19, "text": "So we can do joint hidden state inference and parameter learning at the end of chapter three.", "speaker": "SPEAKER_02"}, {"start": 1165.21, "end": 1166.812, "text": "And that's the whole point of chapter three.", "speaker": "SPEAKER_02"}, {"start": 1166.853, "end": 1169.837, "text": "That's the really central thing that it's about.", "speaker": "SPEAKER_02"}, {"start": 1170.818, "end": 1172.02, "text": "But there's a problem.", "speaker": "SPEAKER_02"}, {"start": 1172.287, "end": 1185.922, "text": " The big problem is that the expectation maximization algorithm, at least of the kind that we've seen here, assumes it assumes that we can know that we can we can evaluate the true posterior.", "speaker": "SPEAKER_02"}, {"start": 1186.203, "end": 1187.967, "text": "So if I go back to Chapter two.", "speaker": "SPEAKER_02"}, {"start": 1189.617, "end": 1218.831, "text": " assumes that we can exactly calculate the true posterior belief okay and unfortunately that assumption is just usually wrong it's usually the case that we simply cannot evaluate the true posterior for any number of reasons the big reason is because marginalizing out the um model evidence is usually very very hard if not impossible and so you just can't use base's rule so we're kind of stuck you know we know abstractly what we", "speaker": "SPEAKER_02"}, {"start": 1218.963, "end": 1244.802, "text": " should be doing if we wanted to solve the problem of hidden state inference and parameter learning but we can't actually do that so chapter four this is really where we kick into high gear in terms of the you know mathematical foundation of active inference itself this is why we say look we can't calculate the exact posterior so we're going to approximate it we're going to find a posterior that's close enough to the real posterior okay", "speaker": "SPEAKER_02"}, {"start": 1245.659, "end": 1273.945, "text": " the process by which we do this is variational inference so we're going to assume that the real posterior takes a certain shape or is a certain kind of probability and usually we're going to assume that the posterior is a gaussian right or a bell curve that's an assumption the real posterior might be something wiggly it might be something very strange okay but we're going to assume that uh at least it's good enough to assume that the true posterior is is a gaussian", "speaker": "SPEAKER_02"}, {"start": 1274.988, "end": 1276.67, "text": " I don't have figures here just yet.", "speaker": "SPEAKER_02"}, {"start": 1277.851, "end": 1284.898, "text": "And then there's a question about, all right, we're gonna try and make that Gaussian, that bell curve as close as possible to the real posterior.", "speaker": "SPEAKER_02"}, {"start": 1285.839, "end": 1286.48, "text": "And how do we do that?", "speaker": "SPEAKER_02"}, {"start": 1287.1, "end": 1295.489, "text": "Well, there's a quantity called variational free energy, which bounds the surprisal, bounds the negative log of the model elements.", "speaker": "SPEAKER_02"}, {"start": 1296.33, "end": 1297.711, "text": "And this thing we can calculate.", "speaker": "SPEAKER_02"}, {"start": 1299.253, "end": 1303.577, "text": "So on the basis of that, we can do variational free energy minimization,", "speaker": "SPEAKER_02"}, {"start": 1304.215, "end": 1309.664, "text": " And there's various ways of talking about how to do this in terms of free form, mean field, fixed form, mean field.", "speaker": "SPEAKER_02"}, {"start": 1309.684, "end": 1311.887, "text": "We didn't end up talking about those very much at all, actually.", "speaker": "SPEAKER_02"}, {"start": 1313.089, "end": 1315.693, "text": "And this is something we can do.", "speaker": "SPEAKER_02"}, {"start": 1316.074, "end": 1320.2, "text": "So this is actually how we're going to do our approximate Bayesian inference.", "speaker": "SPEAKER_02"}, {"start": 1320.22, "end": 1321.062, "text": "Admit, admit, admit.", "speaker": "SPEAKER_02"}, {"start": 1323.385, "end": 1325.769, "text": "I see, Andrew, you have a hand up.", "speaker": "SPEAKER_02"}, {"start": 1325.789, "end": 1326.931, "text": "Sorry, I didn't notice that before.", "speaker": "SPEAKER_02"}, {"start": 1327.051, "end": 1328.333, "text": "Would you like to jump in?", "speaker": "SPEAKER_02"}, {"start": 1329.157, "end": 1329.537, "text": " Oh, yeah.", "speaker": "SPEAKER_04"}, {"start": 1329.838, "end": 1330.618, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 1331.099, "end": 1331.539, "text": "No worries.", "speaker": "SPEAKER_04"}, {"start": 1331.599, "end": 1332.881, "text": "It is going to be brief.", "speaker": "SPEAKER_04"}, {"start": 1332.921, "end": 1334.282, "text": "I didn't want to slow your momentum.", "speaker": "SPEAKER_04"}, {"start": 1335.303, "end": 1342.95, "text": "But yeah, just on on parameters, I think for those who are still wondering, like, what what are these extra parameters?", "speaker": "SPEAKER_04"}, {"start": 1343.03, "end": 1351.479, "text": "If we've been talking about things like perception and taking in a, you know, a sensory observation and then inferring a hidden state.", "speaker": "SPEAKER_04"}, {"start": 1351.539, "end": 1353.661, "text": "So what are these parameter things sort of about?", "speaker": "SPEAKER_04"}, {"start": 1353.781, "end": 1355.302, "text": "You've already addressed that, Fraser.", "speaker": "SPEAKER_04"}, {"start": 1355.342, "end": 1357.965, "text": "But it's just to say.", "speaker": "SPEAKER_04"}, {"start": 1358.148, "end": 1387.429, "text": " um you know we we take in an observation we sort of update a belief so if i were to use more colloquial terms here to think provide more of an intuition um it's sort of like you know we see our y and we update our x um but how do we do that right so so if you think of of the the you know the y is like what you perceive and then and then x is your beliefs then parameters are sort of like", "speaker": "SPEAKER_04"}, {"start": 1387.409, "end": 1407.679, "text": " your understanding they're they're what help relate the x and the y right because without some kind of understanding so to speak of how um you can link up your your beliefs with what you see or what you observe um like without any kind of parameters like we have no way of relating those two things right", "speaker": "SPEAKER_04"}, {"start": 1407.659, "end": 1414.472, "text": " So it's the parameters that kind of lead to this equating different sorts of things.", "speaker": "SPEAKER_04"}, {"start": 1414.553, "end": 1421.606, "text": "It's like, oh, well, X is simply, you know, we take our Y and then we run it through our understanding.", "speaker": "SPEAKER_04"}, {"start": 1421.747, "end": 1425.955, "text": "We run it through our parameters and the sort of combinations of these things, right?", "speaker": "SPEAKER_04"}, {"start": 1425.935, "end": 1429.744, "text": " So that's sort of what's essential about this idea of parameters.", "speaker": "SPEAKER_04"}, {"start": 1430.104, "end": 1440.307, "text": "And then, of course, why we have these things like chapter 2 and chapter 3, right next to each other, because they're, they're so sort of inextricably tied up with each other.", "speaker": "SPEAKER_04"}, {"start": 1440.728, "end": 1443.775, "text": "And then, as Frazier was going into now.", "speaker": "SPEAKER_04"}, {"start": 1443.755, "end": 1473.542, "text": " variational inference then becomes this sort of tractable means of running that sort of process of going from our y to our x as well as being able to invert that to go from an x to a y via parameters and being able able to to update those parameters over time that is they don't have to your understanding of how states and and observations also does not stay fixed that's why we have not just inference but we also have", "speaker": "SPEAKER_04"}, {"start": 1473.522, "end": 1479.259, "text": " But, yeah, that was also just supposed to be, like, kind of a brief point along the way.", "speaker": "SPEAKER_04"}, {"start": 1479.74, "end": 1483.431, "text": "I think your explanation thus far has been really great tying things together.", "speaker": "SPEAKER_04"}, {"start": 1484.288, "end": 1484.909, "text": " Yeah, no, fair enough.", "speaker": "SPEAKER_02"}, {"start": 1484.949, "end": 1485.489, "text": "Thank you, Andrew.", "speaker": "SPEAKER_02"}, {"start": 1485.509, "end": 1492.438, "text": "I mean, we've seen along the way various representations of our model in terms of maybe a graphical representation.", "speaker": "SPEAKER_02"}, {"start": 1492.458, "end": 1507.957, "text": "This is one representation here where very quickly, this is exactly expressing the idea that, look, you know, there are hidden states that generate observations, but our model about how observations are generated from hidden states is dependent on parameters.", "speaker": "SPEAKER_02"}, {"start": 1508.657, "end": 1512.422, "text": "And we can also have beliefs about parameters, okay, that are", "speaker": "SPEAKER_02"}, {"start": 1512.402, "end": 1539.692, "text": " parameterized by further other parameters and we can update our belief about parameters in addition to getting states so you can kind of do inference on parameters as what it means to be doing learning so that's an interpretation that shows up again and again at least later but um we've seen the beginnings of it here so uh yeah so okay so chapter four this is this is you know i i say where we're", "speaker": "SPEAKER_02"}, {"start": 1539.841, "end": 1548.072, "text": " introducing how it is that we're really practically going to do approximate Bayesian inference by means of variational free energy minimization.", "speaker": "SPEAKER_02"}, {"start": 1548.092, "end": 1557.664, "text": "We spent a long time motivating where it comes from, what VFE even means, why it bounds surprise, and why that's a good thing.", "speaker": "SPEAKER_02"}, {"start": 1559.206, "end": 1566.836, "text": "But the upshot is that once we can formulate our inference problem in terms of variational free energy, this is something that we can solve practically.", "speaker": "SPEAKER_02"}, {"start": 1567.617, "end": 1569.139, "text": "That gives us the mathematical machinery", "speaker": "SPEAKER_02"}, {"start": 1569.423, "end": 1576.209, "text": " to be able to do the inference problems, attack the inference problems we want to attack, in addition to being able to learn parameters.", "speaker": "SPEAKER_02"}, {"start": 1577.47, "end": 1584.236, "text": "And that's going to be the mathematical foundation for the rest of the book, really.", "speaker": "SPEAKER_02"}, {"start": 1584.457, "end": 1587.579, "text": "So that gives us the story of perception so far.", "speaker": "SPEAKER_02"}, {"start": 1588.2, "end": 1593.164, "text": "But the story about how action comes in and how actions are selected is a very, very similar story.", "speaker": "SPEAKER_02"}, {"start": 1593.264, "end": 1597.668, "text": "We're, quote unquote, just going to be minimizing variation of free energy.", "speaker": "SPEAKER_02"}, {"start": 1597.688, "end": 1599.43, "text": "But we're going to see a little bit of a wrinkle", "speaker": "SPEAKER_02"}, {"start": 1600.085, "end": 1625.552, "text": " it means to be doing planning and expected free energy but again that's coming later much later actually now in some sense okay so you know chapter five comes next right now that's still part of part one but really i tend to think that chapters one to four is sort of spiritually part one you know so i think part one kind of ends at chapter four in some sense chapter five is kind of a historical", "speaker": "SPEAKER_02"}, {"start": 1626.578, "end": 1632.225, "text": " or motivating additional perspective on what chapter 4 is about.", "speaker": "SPEAKER_02"}, {"start": 1632.786, "end": 1637.131, "text": "So chapter 5 is about this other thing called predictive coding or predictive processing.", "speaker": "SPEAKER_02"}, {"start": 1637.151, "end": 1653.312, "text": "And really, the whole idea is that we can re-express what it is to do variational free energy minimization in chapter 4 in terms of another process called prediction error minimization.", "speaker": "SPEAKER_02"}, {"start": 1653.545, "end": 1657.483, "text": " or more precisely, precision weighted prediction error minimization.", "speaker": "SPEAKER_02"}, {"start": 1657.504, "end": 1661.884, "text": "This is just a different perspective on what it is to do via theme minimization.", "speaker": "SPEAKER_02"}, {"start": 1662.168, "end": 1674.825, "text": " And historically, it sort of comes from a different or slightly adjacent tradition in neuroscience and the Bayesian brain hypothesis, where we're explicitly reasoning about these things called prediction errors.", "speaker": "SPEAKER_02"}, {"start": 1674.845, "end": 1679.852, "text": "So it's not really anything new or additional.", "speaker": "SPEAKER_02"}, {"start": 1679.892, "end": 1684.378, "text": "It's kind of, as I say, just a different perspective on what it is to be doing via theme minimization.", "speaker": "SPEAKER_02"}, {"start": 1684.578, "end": 1685.339, "text": "It's an important one.", "speaker": "SPEAKER_02"}, {"start": 1685.88, "end": 1688.884, "text": "It shows up in implementations every now and again.", "speaker": "SPEAKER_02"}, {"start": 1689.1, "end": 1709.758, "text": " um and it's it itself has been given um the ability to to do action selection which is very important but it's it's not it's not a constituent like core idea to active inference itself it's really just a different perspective so that is all of part one", "speaker": "SPEAKER_02"}, {"start": 1710.667, "end": 1711.088, "text": " Very nice.", "speaker": "SPEAKER_02"}, {"start": 1711.208, "end": 1715.155, "text": "I'll just mention before I forget, Andrew has, and where are they, Andrew, your slides?", "speaker": "SPEAKER_02"}, {"start": 1715.756, "end": 1729.08, "text": "So I've put together these comprehensive sort of notes and so on, where you've got your, what are the key ideas in the chapter, what's core, what's not, and then I have the same sort of thing for the sections.", "speaker": "SPEAKER_02"}, {"start": 1729.955, "end": 1733.425, "text": " But Andrew, you've got your slides, which are even more.", "speaker": "SPEAKER_02"}, {"start": 1733.665, "end": 1740.805, "text": "So this is quite detailed, but I like your slides because they compress a lot of information in a very tight way, which is helpful, I think.", "speaker": "SPEAKER_02"}, {"start": 1741.267, "end": 1743.011, "text": "Are they in the code?", "speaker": "SPEAKER_02"}, {"start": 1743.312, "end": 1744.335, "text": "I always forget where they are.", "speaker": "SPEAKER_02"}, {"start": 1744.349, "end": 1770.238, "text": " yeah yeah no worries um i've kind of pushed daniel to put them somewhere more obvious and he's still not doing it uh so so if you go to figures um okay yeah uh which was just a couple above where you're currently at um so we've moved on yeah yeah so so if you go to figures and then you actually scroll all the way to the bottom like actually after", "speaker": "SPEAKER_04"}, {"start": 1770.218, "end": 1773.643, "text": " Um, these, yeah, they're tucked away there.", "speaker": "SPEAKER_04"}, {"start": 1773.683, "end": 1778.049, "text": "And then also on the, the initial landing page.", "speaker": "SPEAKER_04"}, {"start": 1779.512, "end": 1786.922, "text": "Daniel's put a link to what he calls supplemental materials, and that will just take you directly to the slides as well.", "speaker": "SPEAKER_04"}, {"start": 1787.924, "end": 1791.089, "text": "But, yeah, we're going to try to put those somewhere.", "speaker": "SPEAKER_04"}, {"start": 1791.129, "end": 1791.529, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 1791.589, "end": 1793.492, "text": "Supplemental.", "speaker": "SPEAKER_04"}, {"start": 1793.742, "end": 1817.05, "text": " so yeah hopefully we could get those in a more obvious spot and the big thing is that you know they're because they're not complete um you know I think it's fine to just repeatedly share that they're sort of in progress right um but yeah the the the key idea is that uh one not everyone has the textbook and so this is like in addition to the coda this is attempting to", "speaker": "SPEAKER_04"}, {"start": 1817.03, "end": 1844.809, "text": " kind of put together a lot of information from the textbook without directly reproducing the textbook and kind of giving trying to reduce some of the redundancy because for those who have been following through the chapters you've actually seen like we even have one-to-one direct repetitions of some of the equations from chapter to chapter because of how tightly they're linked like it is warranted that there's some kind of repetition going on from chapter to chapter", "speaker": "SPEAKER_04"}, {"start": 1844.789, "end": 1850.82, "text": " So it's just intended to focus on particular concepts.", "speaker": "SPEAKER_04"}, {"start": 1851.962, "end": 1859.797, "text": "Some of the earlier slides will go into even more basic sort of things that we're getting into that even the textbook doesn't touch on.", "speaker": "SPEAKER_04"}, {"start": 1859.857, "end": 1864.806, "text": "The textbook kind of assumes that people might be taking more of an engineering direction.", "speaker": "SPEAKER_04"}, {"start": 1865.056, "end": 1872.07, "text": " something like an equivalent to undergraduate coursework in statistics or otherwise.", "speaker": "SPEAKER_04"}, {"start": 1872.511, "end": 1878.162, "text": "So in some of the earlier slides, I say like, well, what is a continuous distribution?", "speaker": "SPEAKER_04"}, {"start": 1878.543, "end": 1881.268, "text": "And what is a probability density function?", "speaker": "SPEAKER_04"}, {"start": 1881.288, "end": 1883.773, "text": "And all these other sorts of things that we rely upon.", "speaker": "SPEAKER_04"}, {"start": 1883.753, "end": 1897.288, "text": " through the textbook that don't necessarily get spelled out for you, and what are some more intuitive ways of thinking about these sorts of things, especially for those who are not coming from an engineering background but hopefully want to get the most they can out of the book.", "speaker": "SPEAKER_04"}, {"start": 1897.308, "end": 1908.84, "text": "And then you get, you know, we talked about how gradient descent itself is such an involved process, and yet it really gets kind of summarized within two pages in the textbook.", "speaker": "SPEAKER_04"}, {"start": 1908.86, "end": 1913.305, "text": "So having a slide dedicated to that, you know,", "speaker": "SPEAKER_04"}, {"start": 1913.285, "end": 1925.822, "text": " I make some references to the earlier 2022 textbook as well to sort of show some of the continuity between that, which was not focused on engineers, versus this textbook, too.", "speaker": "SPEAKER_04"}, {"start": 1926.903, "end": 1937.537, "text": "So, yeah, hopefully it provides any kind of additional value to those who are wanting to get the most out of this textbook, whether they own it or not.", "speaker": "SPEAKER_04"}, {"start": 1938.952, "end": 1946.341, "text": " Yeah, I really appreciate this because we've got at the moment 20 slides here, maybe one or two slides on average per chapter.", "speaker": "SPEAKER_02"}, {"start": 1947.202, "end": 1954.73, "text": "If you wanted a first point of contact with the material other than the book, obviously my notes and so on, there's a lot of that.", "speaker": "SPEAKER_02"}, {"start": 1955.571, "end": 1963.02, "text": "There's still less in the actual book, but this is quite a nice way to maybe not mistake the forest for the trees.", "speaker": "SPEAKER_02"}, {"start": 1963.721, "end": 1965.583, "text": "This is a nice...", "speaker": "SPEAKER_02"}, {"start": 1966.07, "end": 1986.289, "text": " synoptic overall kind of feel for what the what the concepts are and where they come from i mean it's a bit hilarious saying that given how dense each slide is you know you've done a very good job andrew of uh i think walking that tightrope but do uh do take a look at andrew's slides they are a nice um initial global view about what we're talking about", "speaker": "SPEAKER_02"}, {"start": 1986.691, "end": 1993.566, "text": " And then hopefully you can sort of titrate between there and maybe my notes as to what level of abstraction you want.", "speaker": "SPEAKER_02"}, {"start": 1993.586, "end": 1997.114, "text": "And that is specifically, as you say, for people who maybe don't have the book.", "speaker": "SPEAKER_02"}, {"start": 1997.134, "end": 1999.8, "text": "Obviously, the book is always, you know, that's the final port of call.", "speaker": "SPEAKER_02"}, {"start": 2000.261, "end": 2006.455, "text": "But for those who don't have the book, hopefully you can get the same kind of benefit or at least a similar level of benefit from those who do.", "speaker": "SPEAKER_02"}, {"start": 2006.857, "end": 2019.027, "text": " so that is the end of part one we're going to be moving into part two which is explicitly about how we put action into this story and that really is going to be active inference right so chapters", "speaker": "SPEAKER_02"}, {"start": 2019.412, "end": 2048.068, "text": " well for the first chapter of part two generalized filtering for perception we're not going to immediately jump into what it means to be doing action we're still going to be within the perceptions of story but it's going to introduce this fundamental algorithm called generalized filtering which we can use to do perception but we can also use to do action okay and then chapter seven is active generalized filtering uh so really and then you know once we do that we can then tell the story about learning again", "speaker": "SPEAKER_02"}, {"start": 2048.453, "end": 2072.638, "text": " tension and hierarchical models again okay so we told the story of hidden states or hidden state inference in two we told the story of learning in three and then we told the story of approximations of this in four and hierarchical um approaches in five that same flavor is going to be recapitulated in two again okay but just with the story of action coming in so six seven eight or maybe seven eight nine", "speaker": "SPEAKER_02"}, {"start": 2073.395, "end": 2103.3, "text": " these are kind of the core chapters of the book so it's going to be really cool once we get to those i guess i'll open it up for questions for live questions if there are any um please do not hesitate maybe i'll stop sharing just now i'd love to if there aren't any live questions immediately or if people would rather put them in the coda that is certainly advisable um i'm trying to make an effort of going through the coda more um with more frequency than i have before so", "speaker": "SPEAKER_02"}, {"start": 2103.871, "end": 2107.757, "text": " Yes, if there are questions, now would be an excellent time.", "speaker": "SPEAKER_02"}, {"start": 2107.777, "end": 2110.121, "text": "Carrie, I see your actual hand up.", "speaker": "SPEAKER_02"}, {"start": 2110.681, "end": 2113.225, "text": "My real hand, my real face in real time.", "speaker": "SPEAKER_01"}, {"start": 2114.748, "end": 2127.347, "text": "I was really struck by a comment you were making in reference to chapter three and learning in terms of the stubborn agent and the open-minded agent.", "speaker": "SPEAKER_01"}, {"start": 2127.487, "end": 2130.672, "text": "And those are my framing of things.", "speaker": "SPEAKER_01"}, {"start": 2130.652, "end": 2158.06, "text": " um and it was those three prior likelihood posterior you know um graphs and um because this um this question around um what are we attending to in our environment was sort of alluded to in your description and i i just wanted to maybe", "speaker": "SPEAKER_01"}, {"start": 2158.63, "end": 2170.75, "text": " give a little bit more space to talk about it, because I think it's a really interesting element of active inference that I wanna, you know, investigate a little bit", "speaker": "SPEAKER_01"}, {"start": 2171.624, "end": 2190.713, "text": " more thoroughly, especially with the idea of the stubborn agent being fixed on that, like what happens when you get stuck in thinking what you already know versus how do you create conditions that someone might be able to take something in", "speaker": "SPEAKER_01"}, {"start": 2190.693, "end": 2219.653, "text": " to um to do more broad learning and it sounds like or at least what i was hearing for the first time in a new way um was that it had something to do with that likelihood function where you're more precisely observing certain things um do you know are you still looking for the the graph no i found it i just i wanted to the graph is in chapter two uh i'll actually just name it so it's ticket 2.11 this one here yeah", "speaker": "SPEAKER_01"}, {"start": 2220.106, "end": 2247.312, "text": " yeah so anyway i don't i don't have a specific question i just wanted to to sort of dog ear how cool this is to me nice yeah well i i completely agree and i i spent a little bit of time as as i did in chapter in the the review for chapter two hopping on about why this is a cool thing because we've actually seen in chapter five with predictive coding this story has been told again", "speaker": "SPEAKER_02"}, {"start": 2247.629, "end": 2257.603, "text": " And it's been told, we've elaborated the idea, well, we've investigated the idea more precisely about, okay, well, what should I attend to?", "speaker": "SPEAKER_02"}, {"start": 2258.184, "end": 2259.005, "text": "Okay.", "speaker": "SPEAKER_03"}, {"start": 2259.025, "end": 2263.832, "text": "Because the whole point of chapter five with the study of prediction errors is there's prediction errors all over the place.", "speaker": "SPEAKER_02"}, {"start": 2263.852, "end": 2265.895, "text": "You know, there's prediction errors here, prediction errors there, whatever.", "speaker": "SPEAKER_02"}, {"start": 2266.556, "end": 2269.941, "text": "But only some of them are worth paying attention to, typically.", "speaker": "SPEAKER_02"}, {"start": 2270.478, "end": 2281.195, "text": " And that's where in that chapter, in chapter five, we motivate this idea of precision weighted prediction errors being the important thing, not just the prediction errors themselves.", "speaker": "SPEAKER_02"}, {"start": 2282.136, "end": 2295.557, "text": "So that's a more sophisticated notion about, because I think from your question, what you're gesturing to is there is this question about what we should attend to in some sense.", "speaker": "SPEAKER_02"}, {"start": 2295.875, "end": 2297.517, "text": " And there's various ways of attacking that.", "speaker": "SPEAKER_02"}, {"start": 2297.537, "end": 2305.445, "text": "At the coarsest level, if we just sort of say, all right, I've got prior beliefs about what I think I'm going to observe, I have a likelihood mapping.", "speaker": "SPEAKER_02"}, {"start": 2306.446, "end": 2314.095, "text": "The dance or the interplay between those two things affect how I'm updating my beliefs in terms of precision and such like.", "speaker": "SPEAKER_02"}, {"start": 2315.196, "end": 2325.527, "text": "That's one way that's the question about how I'm attending and what I'm attending to affects my belief updates.", "speaker": "SPEAKER_02"}, {"start": 2326.402, "end": 2352.004, "text": " but in a in a similar vein it's not quite the same idea but in a similar vein with chapter five we then additionally layer on top this notion that there are certain kinds of prediction errors that are that are worth more than others okay and it's the precise ones that we want to talk about so there's a there's an example i like uh fog don't know if i'll be able to find it in time but", "speaker": "SPEAKER_02"}, {"start": 2353.013, "end": 2357.718, "text": " A motivating example is that, you know, let's say you're a car in some sort of foggy environment.", "speaker": "SPEAKER_02"}, {"start": 2359.08, "end": 2361.622, "text": "You're perceiving things.", "speaker": "SPEAKER_02"}, {"start": 2362.023, "end": 2363.244, "text": "Okay, maybe this is a nice one here.", "speaker": "SPEAKER_02"}, {"start": 2364.085, "end": 2372.414, "text": "You can obviously observe, by means of your likelihood model and prior belief, you can update your belief about what you're observing in the fog.", "speaker": "SPEAKER_02"}, {"start": 2373.495, "end": 2382.885, "text": "But it would be better to look at clearings in the fog than to just look at anywhere, because those clearings give you very precise information.", "speaker": "SPEAKER_02"}, {"start": 2383.557, "end": 2385.819, "text": " So you've got some prior belief about what may be in the fog.", "speaker": "SPEAKER_02"}, {"start": 2386.279, "end": 2398.51, "text": "But if you can attend to areas in the fog that are going to give you precise feedback, that's going to be better than attending to areas in the fog that are not going to give you as precise feedback.", "speaker": "SPEAKER_02"}, {"start": 2399.151, "end": 2408.079, "text": "So that's where this notion of precision weighting comes in to add to the story about this question of stubbornness of the agent or not.", "speaker": "SPEAKER_02"}, {"start": 2408.139, "end": 2411.482, "text": "So Andrew, is there anything you'd like to share on that front?", "speaker": "SPEAKER_02"}, {"start": 2411.502, "end": 2412.703, "text": "Because I know you've", "speaker": "SPEAKER_02"}, {"start": 2413.139, "end": 2415.822, "text": " you may have some additional context there being a social science guy.", "speaker": "SPEAKER_02"}, {"start": 2420.046, "end": 2428.394, "text": "And before you jump in, Andrew, just real quick, is the likelihood function, what's the point of view of the likelihood function?", "speaker": "SPEAKER_01"}, {"start": 2428.574, "end": 2443.028, "text": "Is that likelihood function for the stubborn agent based on the fact he's like, is those two likelihood functions, the same environment perceived differently based on your priors?", "speaker": "SPEAKER_01"}, {"start": 2444.257, "end": 2444.657, "text": " Well, yeah.", "speaker": "SPEAKER_02"}, {"start": 2444.698, "end": 2448.402, "text": "So the idea is that you could have exactly the same likelihood.", "speaker": "SPEAKER_02"}, {"start": 2448.442, "end": 2449.643, "text": "In this case, we don't.", "speaker": "SPEAKER_02"}, {"start": 2450.464, "end": 2456.992, "text": "And where the only difference is the precision of the prior belief, and that would affect the belief update.", "speaker": "SPEAKER_02"}, {"start": 2457.212, "end": 2465.522, "text": "OK, so you could hold the likelihoods function, the likelihood mapping equal in both cases and vary the prior precision.", "speaker": "SPEAKER_02"}, {"start": 2466.228, "end": 2470.561, "text": " Or you could hold the prior to be the same and vary the likelihood function.", "speaker": "SPEAKER_02"}, {"start": 2470.581, "end": 2472.466, "text": "And you could see this same relationship here.", "speaker": "SPEAKER_02"}, {"start": 2472.486, "end": 2479.326, "text": "So is that what you're asking in terms of maybe you've got two agents and they have the same likelihood mapping, but they have different prior beliefs?", "speaker": "SPEAKER_02"}, {"start": 2482.192, "end": 2483.013, "text": " That's certainly possible.", "speaker": "SPEAKER_02"}, {"start": 2483.233, "end": 2484.074, "text": "That's very, very possible.", "speaker": "SPEAKER_02"}, {"start": 2486.958, "end": 2489.441, "text": "This is one of the best figures.", "speaker": "SPEAKER_04"}, {"start": 2490.642, "end": 2502.917, "text": "I think this figure is one of the most illustrative in the textbook that we've seen thus far, as much as I appreciate the descending the contours, et cetera, whenever we look at gradient descent and all that.", "speaker": "SPEAKER_04"}, {"start": 2502.897, "end": 2505.461, "text": " But I think for building an intuition, yeah.", "speaker": "SPEAKER_04"}, {"start": 2505.542, "end": 2511.171, "text": "So with this figure, it's just like the main idea is we have two rows of plots.", "speaker": "SPEAKER_04"}, {"start": 2511.271, "end": 2518.003, "text": "They show the same things, but the top row is for one particular agent in a particular situation.", "speaker": "SPEAKER_04"}, {"start": 2518.27, "end": 2521.815, "text": " And then the bottom row is for a different agent.", "speaker": "SPEAKER_04"}, {"start": 2522.616, "end": 2528.905, "text": "So both of these agents could be in the same environment and even observing the same exact things.", "speaker": "SPEAKER_04"}, {"start": 2529.666, "end": 2543.586, "text": "But the one in the top row had a higher precision on its prior, meaning it has a much sharper and potentially more confident view of its own prior beliefs.", "speaker": "SPEAKER_04"}, {"start": 2543.566, "end": 2553.902, "text": " Whereas in the bottom row, it's the opposite situation that the agent has a much sharper, more precise belief and confidence in its likelihood.", "speaker": "SPEAKER_04"}, {"start": 2554.342, "end": 2555.344, "text": "So what's that mean?", "speaker": "SPEAKER_04"}, {"start": 2555.404, "end": 2560.191, "text": "It means that the prior beliefs that you sort of bring to the table", "speaker": "SPEAKER_04"}, {"start": 2560.171, "end": 2563.116, "text": " Are are going to be sort of a bias, right?", "speaker": "SPEAKER_04"}, {"start": 2563.136, "end": 2571.249, "text": "They're going to be a sort of bias of, like, before I even see any sensory information before I even take any observations into account.", "speaker": "SPEAKER_04"}, {"start": 2571.67, "end": 2577.619, "text": "What do I think the hidden state is such as, you know, in this example, it's like the hidden state.", "speaker": "SPEAKER_04"}, {"start": 2577.66, "end": 2582.427, "text": "It has to do with food availability where the observations are light intensity.", "speaker": "SPEAKER_04"}, {"start": 2582.568, "end": 2582.888, "text": "Right?", "speaker": "SPEAKER_04"}, {"start": 2582.868, "end": 2592.873, "text": " So it's likelihood, the likelihood on the other hand, as opposed to the prior, the likelihood is what relates hidden states and observations, right?", "speaker": "SPEAKER_04"}, {"start": 2593.254, "end": 2595.78, "text": "So that's sort of the way that you translate", "speaker": "SPEAKER_04"}, {"start": 2595.963, "end": 2602.551, "text": " the world and what you observe into your beliefs and vice versa when we invert the model.", "speaker": "SPEAKER_04"}, {"start": 2602.872, "end": 2606.676, "text": "What kind of observations should you expect based on your prior beliefs?", "speaker": "SPEAKER_04"}, {"start": 2607.237, "end": 2618.511, "text": "So if you have an agent who has very precise priors and a much less precise likelihood, they're going to rely much more on their beliefs that they already bring in advance.", "speaker": "SPEAKER_04"}, {"start": 2618.491, "end": 2633.513, "text": " Uh, to to the situation, and this is such an interesting thing as far as, like, cognitive phenomena and thinking about how not just that, like, this sort of micro scale of neurons, but how potentially, you know, entire organisms, including people.", "speaker": "SPEAKER_04"}, {"start": 2633.493, "end": 2648.515, "text": " operate whenever someone has such a strong prior belief, it can sort of dominate their ability to interpret the world around them because they're going to down weight the observations that they're actually taking in.", "speaker": "SPEAKER_04"}, {"start": 2648.495, "end": 2652.443, "text": " And those, those, those strong, precise priors that they bring to the table.", "speaker": "SPEAKER_04"}, {"start": 2652.483, "end": 2652.763, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 2652.843, "end": 2655.529, "text": "Are going to to skew how they take things.", "speaker": "SPEAKER_04"}, {"start": 2655.669, "end": 2656.951, "text": "And so so.", "speaker": "SPEAKER_04"}, {"start": 2656.992, "end": 2665.508, "text": "That relates to your original question regarding, like, what we actually attend to, because if you have very high, these sort of the.", "speaker": "SPEAKER_04"}, {"start": 2665.488, "end": 2670.815, "text": " In computational psychiatry, the phrase hyper-precise priors has been brought up.", "speaker": "SPEAKER_04"}, {"start": 2670.915, "end": 2688.257, "text": "I think the hyper is just sort of trying to make the point that the prior beliefs are so extreme, sort of exacerbated, that they can lead to all sorts of interesting, potentially unfortunate misinterpretations of sensory information.", "speaker": "SPEAKER_04"}, {"start": 2689.499, "end": 2692.643, "text": "So someone who doesn't trust", "speaker": "SPEAKER_04"}, {"start": 2692.623, "end": 2710.581, "text": " The observations they receive, so if it has to do with 2 people communicating with 1 another, if you have very, you know, a very strong prior belief that you don't like this other person, or there's something else going on there that that you're really going to rely on your own prior beliefs about this other person you're attending to.", "speaker": "SPEAKER_04"}, {"start": 2710.561, "end": 2720.597, "text": " rather than actually listening to what they say, the kinds of things you can actually observe from them, then you're not going to update your beliefs very much about them.", "speaker": "SPEAKER_04"}, {"start": 2720.838, "end": 2724.263, "text": "You're going to keep your beliefs centered around the priors that you bring in.", "speaker": "SPEAKER_04"}, {"start": 2724.704, "end": 2730.453, "text": "You can think of notions of things like stereotyping and prejudice start to play a role here.", "speaker": "SPEAKER_04"}, {"start": 2731.275, "end": 2733.919, "text": "And then also we can think of things like", "speaker": "SPEAKER_04"}, {"start": 2734.355, "end": 2741.108, "text": " The reverse situation is where someone has a very precise likelihood and very imprecise priors.", "speaker": "SPEAKER_04"}, {"start": 2741.549, "end": 2746.699, "text": "And in that case, you're going to be very reactive to the environment around you.", "speaker": "SPEAKER_04"}, {"start": 2746.739, "end": 2753.933, "text": "You're going to be very focused on the observations you bring in, and you're going to be rapidly updating your beliefs, depending on how", "speaker": "SPEAKER_04"}, {"start": 2753.913, "end": 2756.197, "text": " much those observations change.", "speaker": "SPEAKER_04"}, {"start": 2756.557, "end": 2761.164, "text": "Your prior beliefs can actually add a sense of stability, right?", "speaker": "SPEAKER_04"}, {"start": 2761.225, "end": 2767.575, "text": "Like if something weird happens in your environment, you might say, oh, I don't know if I trust that observation I received.", "speaker": "SPEAKER_04"}, {"start": 2767.895, "end": 2771.08, "text": "I can rely on my prior beliefs to kind of keep me grounded here.", "speaker": "SPEAKER_04"}, {"start": 2771.12, "end": 2772.883, "text": "If you have very", "speaker": "SPEAKER_04"}, {"start": 2773.184, "end": 2778.649, "text": " Imprecise priors than that saying that, you know, oh, I, I have no grounding.", "speaker": "SPEAKER_04"}, {"start": 2778.93, "end": 2790.14, "text": "I'm hyper reactive to the environment around me in some ways in some of the psychiatric literature that's been related to tentatively some of the sorts of things we see in autism.", "speaker": "SPEAKER_04"}, {"start": 2790.781, "end": 2801.031, "text": "Actually, it's a, it's related to, you know, the idea of almost like, hyperactive, sensory, excuse me, hyperactive.", "speaker": "SPEAKER_04"}, {"start": 2801.247, "end": 2811.226, "text": " processing of sensory stimuli, such that your prior beliefs that would kind of make you more grounded in the situation.", "speaker": "SPEAKER_04"}, {"start": 2811.286, "end": 2815.214, "text": "We're not just purely talking about how someone consciously thinks.", "speaker": "SPEAKER_04"}, {"start": 2815.294, "end": 2822.007, "text": "We're also talking about the way that their brain specifically is, you know, kind of interacting with this information, what", "speaker": "SPEAKER_04"}, {"start": 2822.24, "end": 2824.622, "text": " that's where the predictive coding comes in, right?", "speaker": "SPEAKER_04"}, {"start": 2824.662, "end": 2833.071, "text": "That's where the sorts of things that we're seeing later in these later chapters that we've seen so far, they start grounding us a little bit more and starts relating more to neurobiology.", "speaker": "SPEAKER_04"}, {"start": 2834.112, "end": 2849.848, "text": "But yeah, to be so strongly reactive to sensory information in the sense of like having this kind of hyperactive likelihood that dominates your priors, yeah, it can be very disorienting for someone.", "speaker": "SPEAKER_04"}, {"start": 2850.008, "end": 2852.23, "text": "It's just being very reactive.", "speaker": "SPEAKER_04"}, {"start": 2852.21, "end": 2868.591, "text": " I don't want to overspeak because I'm sure that some of what I've said may or may not have been totally clear, but it's worth reflecting on that figure and the idea of what is it about precision in the priors versus the likelihood, right?", "speaker": "SPEAKER_04"}, {"start": 2868.751, "end": 2871.535, "text": "And as we get into later chapters, we'll look at", "speaker": "SPEAKER_04"}, {"start": 2871.802, "end": 2883.263, "text": " how we can have precisions on other things too such as decision making you know what happens whenever someone has very imprecise sort of deliberate uh uh well", "speaker": "SPEAKER_04"}, {"start": 2883.53, "end": 2892.624, "text": " the technical phrase will be policy inference, but basically what happens when you have low precision on your ability to make decisions.", "speaker": "SPEAKER_04"}, {"start": 2892.664, "end": 2903.32, "text": "It would be like someone who is able to come up with decisions they could make, but they might not trust their own ability to follow through on those decisions, right?", "speaker": "SPEAKER_04"}, {"start": 2904.062, "end": 2907.647, "text": "So those would be people who have something like low confidence", "speaker": "SPEAKER_04"}, {"start": 2907.627, "end": 2910.054, "text": " in their own decision making.", "speaker": "SPEAKER_04"}, {"start": 2910.075, "end": 2918.38, "text": "And that's the kind of thing that we might start seeing exhibited in cases of like depression and other kinds of depressive", "speaker": "SPEAKER_04"}, {"start": 2918.933, "end": 2921.437, "text": " symptoms, or at least we'll call behaviors, right?", "speaker": "SPEAKER_04"}, {"start": 2921.718, "end": 2931.113, "text": "Like someone who's still able to think through, they're still able to observe different things, but they might not be able to have a clearer sense of agency in a situation.", "speaker": "SPEAKER_04"}, {"start": 2931.273, "end": 2935.841, "text": "So they have very low precision on their ability to make decisions.", "speaker": "SPEAKER_04"}, {"start": 2936.201, "end": 2940.809, "text": "So yeah, we'll find precision sort of all over the place in future.", "speaker": "SPEAKER_04"}, {"start": 2940.949, "end": 2945.937, "text": "And it's something very interesting to model for sure, because it has all kinds of repercussions", "speaker": "SPEAKER_04"}, {"start": 2945.917, "end": 2973.797, "text": " uh for for the kinds of varieties of behaviors that we see in people uh in organisms otherwise um yeah yeah no it's it's all pervasive that's for sure um gian cuomo did you have a i hope that's how i pronounce your name did you have a question i saw your hand up before i yeah just very briefly it's it's it's it's i don't want to drive it far from here but again say you have very strong priors so you have very concentrated so your prayers have uh you know you're", "speaker": "SPEAKER_00"}, {"start": 2974.655, "end": 3002.383, "text": " are very pointy, you're very, you're, now, this is, is there, what kind of, what kind of architecture do you, so I can imagine an agent who is so very opinionated, who has a, because, you know, I'm generalizing maybe too much, but I'm thinking somebody who has too much, he's been, he's, I think about the human being receiving education, a very rigid education, for example, who, and so in a way he's,", "speaker": "SPEAKER_00"}, {"start": 3002.617, "end": 3028.152, "text": " is very front-loaded in terms of his opinions about the set of matters and therefore he assigned high probabilities to a very narrow set of values if you will i'm sorry and now two you know as one grows old over time i'm sorry i hope the metaphor holds uh you you know you can become sometimes wiser because you accumulate knowledge you accumulate experience you have", "speaker": "SPEAKER_00"}, {"start": 3028.621, "end": 3043.32, "text": " You know, you build in this sophisticated, layered mental and mnemonic database of other people, other relations, their feelings over time.", "speaker": "SPEAKER_00"}, {"start": 3044.862, "end": 3048.807, "text": "You've observed actions and reactions, consequences to actions.", "speaker": "SPEAKER_00"}, {"start": 3048.827, "end": 3055.455, "text": "You observed moral conundrums unfold and you become", "speaker": "SPEAKER_00"}, {"start": 3055.705, "end": 3060.511, "text": " You know, you actually, you see them and actually, so you're bringing all this in now.", "speaker": "SPEAKER_00"}, {"start": 3061.332, "end": 3077.974, "text": "This, what, what, what kind of architecture, how can, how will, how do these active inference models expand to provide, uh, a, let's say a mental temple or a gallery, uh, of say, for example, memory.", "speaker": "SPEAKER_00"}, {"start": 3077.994, "end": 3081.679, "text": "So how can the age, because I can see the agent on his own.", "speaker": "SPEAKER_00"}, {"start": 3082.182, "end": 3094.981, "text": " He has a likelihood, so he has a certain experience of how a certain observation is tied to a certain theoretical value or a certain value, and then he has an opinion about that value.", "speaker": "SPEAKER_00"}, {"start": 3095.382, "end": 3104.255, "text": "And I cannot see an inherent mechanism there for him to say, whoa, wait a moment.", "speaker": "SPEAKER_00"}, {"start": 3104.275, "end": 3106.458, "text": "My priors are a little bit too loaded.", "speaker": "SPEAKER_00"}, {"start": 3106.618, "end": 3108.261, "text": "I'm a bit too rigid.", "speaker": "SPEAKER_00"}, {"start": 3108.621, "end": 3109.843, "text": "I'm a little bit too...", "speaker": "SPEAKER_00"}, {"start": 3110.144, "end": 3118.515, "text": " Biased I have to change my priors now this and so where do you in what directions?", "speaker": "SPEAKER_00"}, {"start": 3120.277, "end": 3139.982, "text": "what kind of interactions and I guess I get circularity what kind of feedback mechanisms other than I mean this kind of these the kind of Hierarchies that I have to think about the ones that we looked at in predictive coding for example when you have maybe several agents that are working at different speeds and then they", "speaker": "SPEAKER_00"}, {"start": 3140.705, "end": 3169.918, "text": " they talk to each other or they question each other and they say hey one agent may observe the others and say hey look there is a you know there is a great majority of my peers that are actually different that have actually different priors why don't i change them to match them for example yeah yeah this so this question came up in the first of the review sessions as well this is a an enormous question and it is actually one of the open", "speaker": "SPEAKER_00"}, {"start": 3170.371, "end": 3173.895, "text": " Well, there's a way in which it's a very trivial question.", "speaker": "SPEAKER_02"}, {"start": 3173.915, "end": 3176.798, "text": "And there's a way in which it's a very, very interesting and deep question.", "speaker": "SPEAKER_02"}, {"start": 3177.579, "end": 3196.581, "text": "The way that it's a trivial question, I will answer first, which is that yes, at the end of sort of chapter five, we saw this idea of hierarchical predictive coding, where the idea is that each layer is in some sense, minimizing local variational free energy, predictions down and you get updates up, okay.", "speaker": "SPEAKER_02"}, {"start": 3196.814, "end": 3207.892, "text": " And what you can then do is you can interpret each layer as having preferences that are in this message passing loop that equilibrate with each other.", "speaker": "SPEAKER_02"}, {"start": 3208.272, "end": 3214.502, "text": "So that would be a sort of account of how it is that some local agent-like thing", "speaker": "SPEAKER_02"}, {"start": 3214.617, "end": 3221.527, "text": " One layer in this hierarchy can attenuate its preferences over time.", "speaker": "SPEAKER_02"}, {"start": 3222.808, "end": 3223.589, "text": "That would be number one.", "speaker": "SPEAKER_02"}, {"start": 3223.609, "end": 3226.554, "text": "That's a kind of very simple way of talking about things.", "speaker": "SPEAKER_02"}, {"start": 3227.034, "end": 3229.458, "text": "Another way is where you generalize this whole thing.", "speaker": "SPEAKER_02"}, {"start": 3229.478, "end": 3233.804, "text": "You want to talk about the agent as a whole interacting with other agents.", "speaker": "SPEAKER_02"}, {"start": 3233.824, "end": 3240.793, "text": "So if I go here as as many as one, I know we're coming up to time everyone, but I just want to.", "speaker": "SPEAKER_02"}, {"start": 3242.756, "end": 3243.477, "text": "For us.", "speaker": "SPEAKER_02"}, {"start": 3244.351, "end": 3252.405, "text": " You might want to tell the story about how interacting agents attenuate their preferences with respect to one another.", "speaker": "SPEAKER_02"}, {"start": 3252.445, "end": 3256.211, "text": "So there's very little work on this at the moment that I'm aware of.", "speaker": "SPEAKER_02"}, {"start": 3258.094, "end": 3261.981, "text": "But there's this idea of federated inference and belief sharing.", "speaker": "SPEAKER_02"}, {"start": 3262.001, "end": 3267.27, "text": "I'm struggling here to", "speaker": "SPEAKER_02"}, {"start": 3267.588, "end": 3273.137, "text": " to communicate how it is that this can be done, because it's very, very early days yet.", "speaker": "SPEAKER_02"}, {"start": 3275.08, "end": 3275.961, "text": "But this is a question.", "speaker": "SPEAKER_02"}, {"start": 3276.021, "end": 3280.869, "text": "And the more general question that I think lurks behind your question is, where do my priors come from?", "speaker": "SPEAKER_02"}, {"start": 3281.97, "end": 3283.052, "text": "Where do my prior beliefs come from?", "speaker": "SPEAKER_02"}, {"start": 3283.092, "end": 3284.314, "text": "Of course, as engineers,", "speaker": "SPEAKER_02"}, {"start": 3284.598, "end": 3311.533, "text": " we're trying to solve a particular problem we're in a very privileged position where we can just say okay well you know i want i want the thermostat to be at a certain temperature this temperature here that's your prior very nice okay but in a lot of problems a lot of situations it's very unclear what the prior should be um and so that question about how it is that we can actually specify priors or how it is that we can recognize when our prior is a bad prior that's a very hard question i tend to think", "speaker": "SPEAKER_02"}, {"start": 3312.07, "end": 3325.005, "text": " that it is precisely in these notions of hierarchy, hierarchical predictive coding, and more generally compositional notions about how agents interface with one another, that this question will be best addressed.", "speaker": "SPEAKER_02"}, {"start": 3325.025, "end": 3334.295, "text": "So I think if you were interested in that question, it would be useful to really delve more into the history of predictive coding and this idea of hierarchical predictive coding.", "speaker": "SPEAKER_02"}, {"start": 3334.315, "end": 3339.541, "text": "So I hope that is somewhat intelligible, or maybe you already understood all of that.", "speaker": "SPEAKER_02"}, {"start": 3339.721, "end": 3341.183, "text": "Those are my thoughts, at least.", "speaker": "SPEAKER_02"}, {"start": 3344.335, "end": 3347.884, "text": " Do add to the coda as well, these questions.", "speaker": "SPEAKER_02"}, {"start": 3347.904, "end": 3354.761, "text": "I think I'll do that because it's very good for other people to see these questions and maybe myself and Andrew and Daniel answering them alongside.", "speaker": "SPEAKER_02"}, {"start": 3354.801, "end": 3360.876, "text": "So maybe if there's one more question, we can take that live and then we'll have to sign off.", "speaker": "SPEAKER_02"}, {"start": 3360.856, "end": 3363.018, "text": " We're probably going to do another week of review.", "speaker": "SPEAKER_02"}, {"start": 3363.258, "end": 3365.84, "text": "OK, so another week of review for part one.", "speaker": "SPEAKER_02"}, {"start": 3366.281, "end": 3370.104, "text": "If you do want that, please tell us that you would like that.", "speaker": "SPEAKER_02"}, {"start": 3370.144, "end": 3373.147, "text": "If you don't want that, please also tell us it does.", "speaker": "SPEAKER_02"}, {"start": 3373.687, "end": 3377.691, "text": "The previous session, the previous group were very much in favor of another week of review.", "speaker": "SPEAKER_02"}, {"start": 3378.912, "end": 3380.213, "text": "But it's up to you guys.", "speaker": "SPEAKER_02"}, {"start": 3380.273, "end": 3381.254, "text": "It's up to us basically.", "speaker": "SPEAKER_02"}, {"start": 3381.274, "end": 3385.518, "text": "So I think as things stand, we're going to next week is going to be yet more review.", "speaker": "SPEAKER_02"}, {"start": 3385.558, "end": 3389.541, "text": "Hopefully we can have some more questions and even some additions to the code.", "speaker": "SPEAKER_02"}, {"start": 3389.581, "end": 3390.282, "text": "That would be nice.", "speaker": "SPEAKER_02"}, {"start": 3390.302, "end": 3390.542, "text": "Andrew.", "speaker": "SPEAKER_02"}, {"start": 3392.8, "end": 3406.095, "text": " As a final point, there was an interesting question asked in the chat regarding what we were talking about with precisions on likelihood and priors.", "speaker": "SPEAKER_04"}, {"start": 3406.516, "end": 3411.962, "text": "Does this have any relationship with some of these cognitive phenomena?", "speaker": "SPEAKER_04"}, {"start": 3412.042, "end": 3416.547, "text": "I'll be honest, I've not seen these terms used too often.", "speaker": "SPEAKER_04"}, {"start": 3416.527, "end": 3419.834, "text": " what are called the halo, horn, and golem effects.", "speaker": "SPEAKER_04"}, {"start": 3420.776, "end": 3426.948, "text": "I mean, I've seen many different sorts of terminologies from different areas of psychology, of sociology, etc.", "speaker": "SPEAKER_04"}, {"start": 3428.151, "end": 3437.269, "text": "It's interesting how often we see in discussions of prejudices, of cognitive bias, of interpretation bias, we have this", "speaker": "SPEAKER_04"}, {"start": 3437.502, "end": 3442.275, "text": " this like sort of plethora of different kinds of terms and terminologies that play out.", "speaker": "SPEAKER_04"}, {"start": 3442.315, "end": 3451.14, "text": "But it's interesting from a standpoint of active inference is that we might be able to relate many of these different kinds of phenomena and potentially even integrate them", "speaker": "SPEAKER_04"}, {"start": 3451.39, "end": 3461.624, "text": " whenever we're able to have some kind of like common like premises for thinking about these sorts of things, perhaps priors and likelihoods and thinking about things like beliefs as a way to do that.", "speaker": "SPEAKER_04"}, {"start": 3461.984, "end": 3464.428, "text": "But it was a really good question.", "speaker": "SPEAKER_04"}, {"start": 3465.97, "end": 3466.671, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 3466.691, "end": 3469.134, "text": "So the Gollum effect, which was an interesting one.", "speaker": "SPEAKER_04"}, {"start": 3470.116, "end": 3470.296, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 3470.797, "end": 3472.639, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 3473.075, "end": 3500.466, "text": " yeah when we we lower our expectation it's kind of this does get into uh well we've probably already seen this phrase in the textbook um but this idea of self-evidencing um that's something about active inference one of the premises is that we kind of minimize our uncertainty and that's not just active inference it's more the the bayesian brain hypothesis and predictive processing and all the rest more generally we minimize our uncertainty", "speaker": "SPEAKER_04"}, {"start": 3500.446, "end": 3510.887, "text": " And so part of how we can minimize our uncertainty is not just to change our minds by learning new things about the world, taking in observations and updating our beliefs.", "speaker": "SPEAKER_04"}, {"start": 3511.328, "end": 3517.14, "text": "Another thing which we'll see much more of in part two of the textbook is, well, we can also act on the world, right?", "speaker": "SPEAKER_04"}, {"start": 3517.541, "end": 3520.767, "text": "Like if I'm very hungry and I don't want to be hungry, like,", "speaker": "SPEAKER_04"}, {"start": 3521.0, "end": 3525.67, "text": " If I just sit here and think like, well, I'm hungry, that must mean I'm going to be hungry all the time.", "speaker": "SPEAKER_04"}, {"start": 3525.73, "end": 3528.516, "text": "I'm not going to do anything about it, right?", "speaker": "SPEAKER_04"}, {"start": 3528.697, "end": 3529.859, "text": "That's just perception.", "speaker": "SPEAKER_04"}, {"start": 3529.919, "end": 3531.843, "text": "That's not engaging in action.", "speaker": "SPEAKER_04"}, {"start": 3531.863, "end": 3537.496, "text": "That's not getting up to go find myself some food and then eating it, right?", "speaker": "SPEAKER_04"}, {"start": 3537.576, "end": 3538.217, "text": "That's action.", "speaker": "SPEAKER_04"}, {"start": 3538.237, "end": 3539.199, "text": "That's decision-making.", "speaker": "SPEAKER_04"}, {"start": 3539.219, "end": 3540.983, "text": "That's actually doing something in the world.", "speaker": "SPEAKER_04"}, {"start": 3541.3, "end": 3554.027, "text": " So with this like Gollum effect sort of thing of like, oh, or the halo effect or the horn effect, it's that we can take in some amount of information about another person.", "speaker": "SPEAKER_04"}, {"start": 3554.48, "end": 3557.384, "text": " And then that impact, so that's likelihood, right?", "speaker": "SPEAKER_04"}, {"start": 3557.484, "end": 3563.393, "text": "We're taking in a sensory observation about another person, interpreting it as some belief about that person.", "speaker": "SPEAKER_04"}, {"start": 3563.774, "end": 3571.425, "text": "But then from there, once we think we have a pretty good belief about that person, what they're about, or what our beliefs about them are,", "speaker": "SPEAKER_04"}, {"start": 3571.405, "end": 3573.588, "text": " we could increase the precision on that.", "speaker": "SPEAKER_04"}, {"start": 3573.628, "end": 3577.292, "text": "It's like, well, I didn't have a very strong prior about them because I didn't know anything about them.", "speaker": "SPEAKER_04"}, {"start": 3577.392, "end": 3580.897, "text": "But now that I do, I'm going to increase that precision.", "speaker": "SPEAKER_04"}, {"start": 3580.937, "end": 3583.2, "text": "Again, I'm saying this colloquially.", "speaker": "SPEAKER_04"}, {"start": 3583.22, "end": 3587.645, "text": "I'm not saying that we're consciously thinking I'm going to increase my precision, right?", "speaker": "SPEAKER_04"}, {"start": 3589.047, "end": 3592.03, "text": "But it might just be that that's how you approach the world.", "speaker": "SPEAKER_04"}, {"start": 3592.091, "end": 3593.312, "text": "That's your model of the world.", "speaker": "SPEAKER_04"}, {"start": 3593.292, "end": 3607.875, "text": " once i learn at least some sufficient amount of information about someone i've now established a prior about it and i feel pretty good about it confident about it so there's a high precision on it uh in terms of the math and then um you know that that can overshadow", "speaker": "SPEAKER_04"}, {"start": 3608.176, "end": 3635.803, "text": " any future observations you receive from the person like oh this person uh must be very you know i i saw them uh you know do something nice they that i watched them give a donation to to a non-profit or something they must be a very good person they must be you know uh interested in philanthropy and other things right and all these other things that tie into other prior beliefs that you have about other people or about how the world works too they can all get related", "speaker": "SPEAKER_04"}, {"start": 3635.783, "end": 3643.41, "text": " But then once that prior belief becomes really, really sharp, then what might happen is you observe something not so good about that person.", "speaker": "SPEAKER_04"}, {"start": 3643.55, "end": 3646.693, "text": "Oh, they, I don't know, they drop trash on the ground.", "speaker": "SPEAKER_04"}, {"start": 3646.794, "end": 3656.563, "text": "You can see I'm really trying to reach for everyday examples here, but they litter or they, oh, they don't seem like they care too much about their family.", "speaker": "SPEAKER_04"}, {"start": 3656.643, "end": 3659.546, "text": "They tend to ignore phone calls, like these other things.", "speaker": "SPEAKER_04"}, {"start": 3659.666, "end": 3665.031, "text": "But if you have that very strong prior belief already established that they're a good person, you might ignore", "speaker": "SPEAKER_04"}, {"start": 3665.686, "end": 3670.573, "text": " future observations about them, and you might not update your prior beliefs any sooner.", "speaker": "SPEAKER_04"}, {"start": 3670.653, "end": 3674.178, "text": "And I think that's what that halo effect is, right?", "speaker": "SPEAKER_04"}, {"start": 3674.259, "end": 3677.403, "text": "We receive some minimal amount of information about someone.", "speaker": "SPEAKER_04"}, {"start": 3677.443, "end": 3678.885, "text": "They seem really good.", "speaker": "SPEAKER_04"}, {"start": 3679.306, "end": 3684.433, "text": "It might even skew our judgment and decision-making towards them where we just assume they're good all the time.", "speaker": "SPEAKER_04"}, {"start": 3684.493, "end": 3688.96, "text": "What's really a mathematical way of saying that is that we've", "speaker": "SPEAKER_04"}, {"start": 3689.176, "end": 3696.787, "text": " We initially had a likelihood that let us establish a prior belief about this person we just met.", "speaker": "SPEAKER_04"}, {"start": 3697.408, "end": 3707.522, "text": "And then after that, our precision on that prior belief is so strong that we're no longer able to update our beliefs about that person going forward quite as much.", "speaker": "SPEAKER_04"}, {"start": 3707.823, "end": 3722.44, "text": " And then with the Gollum effect, that's especially interesting one, because, yeah, if you if you witness someone make a mistake on the job once or twice when they're new at it, if you quickly assume if you update your prior belief and say, oh, they must not be good at this job.", "speaker": "SPEAKER_04"}, {"start": 3722.892, "end": 3735.432, "text": " And then you change your own decision making towards them where you repeatedly, I don't know, put them in positions to, you know, to you say like, oh, they're always, you know, they're not very good at this job.", "speaker": "SPEAKER_04"}, {"start": 3735.472, "end": 3739.699, "text": "They're probably never going to be good at this job, which is a rather extreme thought to have.", "speaker": "SPEAKER_04"}, {"start": 3739.799, "end": 3741.923, "text": "But that kind of thing can start to form.", "speaker": "SPEAKER_04"}, {"start": 3742.624, "end": 3747.192, "text": "You never let them do things that would help them to improve at the job.", "speaker": "SPEAKER_04"}, {"start": 3747.232, "end": 3749.355, "text": "You give them other kinds of tasks instead.", "speaker": "SPEAKER_04"}, {"start": 3749.622, "end": 3757.61, "text": " You go and tell other people behind closed doors, oh, we shouldn't have, this person's not very good at those particular tasks, so don't have them do that, right?", "speaker": "SPEAKER_04"}, {"start": 3757.931, "end": 3765.018, "text": "You're starting to make decisions that are realizing that belief you have, that they're not so good at the job, right?", "speaker": "SPEAKER_04"}, {"start": 3765.038, "end": 3766.82, "text": "And so they never get a chance to improve.", "speaker": "SPEAKER_04"}, {"start": 3766.86, "end": 3768.662, "text": "They don't get to keep working on those tasks.", "speaker": "SPEAKER_04"}, {"start": 3768.722, "end": 3778.212, "text": "You don't talk to them as if sharing further knowledge or information with them is actually going to be helpful because you just assume that they're going to be bad at it in the first place.", "speaker": "SPEAKER_04"}, {"start": 3778.428, "end": 3791.396, "text": " These are all the sorts of things that come out of this sort of like what happens when we have prior beliefs about someone that gets shorn from actually taking in new observations and updating our beliefs about them through a likelihood.", "speaker": "SPEAKER_04"}, {"start": 3793.0, "end": 3796.207, "text": "Yeah, they're all super interesting dynamics.", "speaker": "SPEAKER_04"}, {"start": 3796.267, "end": 3798.732, "text": "It's worth, I think,", "speaker": "SPEAKER_04"}, {"start": 3799.488, "end": 3804.414, "text": " Frazier shared a book by Jacob Howey.", "speaker": "SPEAKER_04"}, {"start": 3804.995, "end": 3820.113, "text": "Yeah, really good reading that's a little less on the equation-heavy side and much more on the thinking through these things in a more social way or thinking about humans and how they operate.", "speaker": "SPEAKER_04"}, {"start": 3820.394, "end": 3823.097, "text": "So yeah, it's all just really fascinating.", "speaker": "SPEAKER_04"}, {"start": 3823.137, "end": 3823.858, "text": "I'm happy to", "speaker": "SPEAKER_04"}, {"start": 3824.277, "end": 3828.324, "text": " talk to anyone about these sorts of things as well going forward.", "speaker": "SPEAKER_04"}, {"start": 3828.464, "end": 3835.135, "text": "Another really good one is The Experience Machine by Andy Clark.", "speaker": "SPEAKER_04"}, {"start": 3835.516, "end": 3840.244, "text": "Andy Clark's very big in this sort of space as well.", "speaker": "SPEAKER_04"}, {"start": 3840.443, "end": 3844.908, "text": " So yeah, it's good stuff.", "speaker": "SPEAKER_04"}, {"start": 3845.209, "end": 3853.499, "text": "And then for those who are more interested in social phenomena, like how do we apply these things to like where it's getting closer and closer.", "speaker": "SPEAKER_04"}, {"start": 3853.539, "end": 3861.368, "text": "And then from my standpoint, some of the work I do, I think of like, well, how do people interact with others in these sorts of ways?", "speaker": "SPEAKER_04"}, {"start": 3861.448, "end": 3865.974, "text": "And how does that reflect back on one's own emotional inference?", "speaker": "SPEAKER_04"}, {"start": 3866.594, "end": 3869.758, "text": "How does that relate to one's own understanding of the", "speaker": "SPEAKER_04"}, {"start": 3869.738, "end": 3872.601, "text": " the world and how does that relate to things like trust?", "speaker": "SPEAKER_04"}, {"start": 3872.761, "end": 3878.006, "text": "How does it allow for trusting in one's own self and their own decision making in relation to others?", "speaker": "SPEAKER_04"}, {"start": 3878.567, "end": 3880.889, "text": "And then, you know, that's where that happens.", "speaker": "SPEAKER_04"}, {"start": 3880.909, "end": 3883.072, "text": "Where interoception comes in, right?", "speaker": "SPEAKER_02"}, {"start": 3883.092, "end": 3892.281, "text": "Where perhaps interoception, because there's this whole kind of story about, you know, we're predicting the environment, but also we're trying to predict ourselves in some way.", "speaker": "SPEAKER_02"}, {"start": 3892.421, "end": 3894.403, "text": "So, you know, what am I doing right now?", "speaker": "SPEAKER_02"}, {"start": 3895.264, "end": 3896.726, "text": "You know, I'm trivially, where's my heartbeat?", "speaker": "SPEAKER_02"}, {"start": 3896.746, "end": 3897.186, "text": "Things like this.", "speaker": "SPEAKER_02"}, {"start": 3897.266, "end": 3898.007, "text": "So, you know,", "speaker": "SPEAKER_02"}, {"start": 3898.223, "end": 3899.064, "text": " Yeah, absolutely.", "speaker": "SPEAKER_04"}, {"start": 3899.104, "end": 3902.929, "text": "Depending on where we draw the line around that, that sort of Markov blanket.", "speaker": "SPEAKER_04"}, {"start": 3903.089, "end": 3908.595, "text": "I mean, if you think of it as like, oh, not, you know, think of an infant who's like just born.", "speaker": "SPEAKER_04"}, {"start": 3908.615, "end": 3914.823, "text": "Like, not only do they have to navigate what we typically think of as an environment where it's like, OK, we have the ground around them.", "speaker": "SPEAKER_04"}, {"start": 3914.843, "end": 3916.224, "text": "We have the room around them.", "speaker": "SPEAKER_04"}, {"start": 3916.265, "end": 3917.706, "text": "We have other people around them.", "speaker": "SPEAKER_04"}, {"start": 3918.727, "end": 3920.89, "text": "But we also have our own bodies.", "speaker": "SPEAKER_04"}, {"start": 3920.87, "end": 3923.916, "text": " We have to learn how to use our bodies.", "speaker": "SPEAKER_04"}, {"start": 3924.537, "end": 3932.152, "text": "There's no assumption here that a child understands every single potential joint position they can make with their arm.", "speaker": "SPEAKER_04"}, {"start": 3932.894, "end": 3937.663, "text": "There's a lot of experimentation, randomly moving and throwing things and all the rest.", "speaker": "SPEAKER_04"}, {"start": 3937.643, "end": 3948.799, "text": " So learning about one's body and reading one's own body, which relates to notions of interoception and proprioception, proprioception being more on the side of motor control.", "speaker": "SPEAKER_04"}, {"start": 3949.18, "end": 3955.669, "text": "And I've had a nice conversation some time ago with Carrie about neuroception and perception of things like pain.", "speaker": "SPEAKER_04"}, {"start": 3955.649, "end": 3963.626, "text": " So, I mean, all of those things can be viewed as part of a broader set of inference and learning processes.", "speaker": "SPEAKER_04"}, {"start": 3965.009, "end": 3970.941, "text": "Yeah, so what happens whenever you interact with someone, but you are interacting", "speaker": "SPEAKER_04"}, {"start": 3970.921, "end": 3991.883, "text": " currently experiencing pain or you're experiencing other kinds of things in your body that make it, you know, interacting with others, you know, there are things going on within you that are distracting or have a higher precision on them to where your brain is attending to that more that can range between someone with, you know, some kind of previous trauma or", "speaker": "SPEAKER_04"}, {"start": 3992.15, "end": 4002.945, "text": " One of my specialties is sort of post-traumatic stress disorder, especially post-traumatic stress disorder in the sense of experiencing a traumatic event.", "speaker": "SPEAKER_04"}, {"start": 4003.226, "end": 4015.763, "text": "So this isn't CPTSD, but later in life, like you have one way of viewing the world and then you have some kind of intense event that strongly changes your view of the world after.", "speaker": "SPEAKER_04"}, {"start": 4015.883, "end": 4019.929, "text": "Like how does your generative model of the world, how you operate in it,", "speaker": "SPEAKER_04"}, {"start": 4020.112, "end": 4030.24, "text": " How does that change before and after that intense event and what kind of things end up getting modulated and how do we view that in terms of how precisions get modulated?", "speaker": "SPEAKER_04"}, {"start": 4030.507, "end": 4035.956, "text": " and all the rest, and how does that relate to things like neurotransmitters, like noradrenaline?", "speaker": "SPEAKER_04"}, {"start": 4035.996, "end": 4048.077, "text": "So it starts having implications for both pharmacological treatment, but then also with the things we're talking about and interacting with others for mental health treatment in the more general sense of therapy.", "speaker": "SPEAKER_04"}, {"start": 4048.137, "end": 4049.819, "text": "Like, how should we talk to someone?", "speaker": "SPEAKER_04"}, {"start": 4050.26, "end": 4054.928, "text": "Should we talk to someone who's been through an intensive trauma and then after that say, oh, no, you're fine?", "speaker": "SPEAKER_04"}, {"start": 4054.908, "end": 4078.975, "text": " everything's going to be safe you just had your mind changed from this like it's much more complex than that right like people being able to exercise there you go i'm gonna say the whole question is very much about okay well you know is the is the problem at the level of the setting of parameters and therefore we just need to tweak some of that and then the inference is going to follow right are the parameters fine and the info you know so there's this issue of the hierarchy", "speaker": "SPEAKER_02"}, {"start": 4079.258, "end": 4082.764, "text": " that we need to be mindful of when we're modeling inactive inference.", "speaker": "SPEAKER_02"}, {"start": 4082.844, "end": 4085.569, "text": "We can't just say, well, I got a model that's doing inference, and that's it.", "speaker": "SPEAKER_02"}, {"start": 4086.05, "end": 4094.084, "text": "There's this whole issue about how you set up the whole hierarchy from learning to inference, or even model construction to learning to inference.", "speaker": "SPEAKER_02"}, {"start": 4094.104, "end": 4098.752, "text": "That's the stack that we need to be appreciative of when we're doing the fundamentals.", "speaker": "SPEAKER_02"}, {"start": 4098.732, "end": 4125.685, "text": " i just wanted to we are we are basically at time um for those who wanted to ask questions but didn't uh get the chance please do put them in the coda andrew and myself and daniel are going to be very attentive to that i'm hopefully going to put together uh well danger daniel and myself are going to be more active with respect to the code implementations and some of the animations and things like that so next week i believe we'll come back and we'll do another round of review hopefully maybe with some", "speaker": "SPEAKER_02"}, {"start": 4125.665, "end": 4135.857, "text": " specific reference to particular examples, maybe in the code, where we can really sort of get our hands dirty, maybe just show how some of these things work and things like that.", "speaker": "SPEAKER_02"}, {"start": 4135.877, "end": 4138.5, "text": "So that's the plan going forward.", "speaker": "SPEAKER_02"}, {"start": 4138.52, "end": 4140.742, "text": "If you guys want to change it, it's up to us.", "speaker": "SPEAKER_02"}, {"start": 4140.802, "end": 4141.463, "text": "We can do what we want.", "speaker": "SPEAKER_02"}, {"start": 4142.124, "end": 4144.186, "text": "But thank you very much for everyone here.", "speaker": "SPEAKER_02"}, {"start": 4144.486, "end": 4147.59, "text": "So I'll stop the YouTube recording now.", "speaker": "SPEAKER_02"}, {"start": 4150.673, "end": 4150.934, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 4151.294, "end": 4152.195, "text": "Goodbye, YouTube people.", "speaker": "SPEAKER_02"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.txt new file mode 100644 index 000000000..7a72659e8 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_024/transcript.txt @@ -0,0 +1,1262 @@ +SPEAKER_02: +all right hello everyone so we're here it says july 3rd 2026 we're doing the um part one review basically the fundamentals of active inference uh so we've gone through chapters one through five that uh constitute the first half of the book part one there's three parts uh so today's really kind of a session opportunity to ask questions about the entirety of part one anything that anyone would like cleared up or elaborated upon + +before we move into part two. + +And that's going to be, part two is going to be active inference proper, where we actually begin to look at action, how action is selected, how we plan, how we do, quote unquote, active inference. + +So that's kind of what we're doing today. + +I shall share my screen. + +I'll give, I think, maybe just a little bit of review on what we've done, where we've gone, where we are. + +And then I'll open up the floor for questions. + +from you guys for anything at all related to part one. + +And hopefully that will kind of set us up to being able to begin to see where we're going to have to go for part two. + +So I guess very basically, we've seen this picture here. + +This is, I think, the first figure in the book from the introduction. + +This is the outline of the entire book. + +And again, hopefully everyone can see my screen. + +Please do make a lot of noise if you can't. + +So part one, chapters one to five. + +This is, you know, the book itself is called The Fundamentals of Active Inference. + +But part one is really the fundamentals of the fundamentals. + +What do you even need to understand in terms of the constituent mathematics and formalisms and ideas that surround active inference and where it comes from in order to be able to understand active inference itself? + +So that's part one. + +I'll give a brief overview of each chapter and the core ideas. + +So what we had to do initially, + +is we had to come to part one here. + +The entirety of part one is completely concerned with the problem of perception only. + +So with active infants, we have this fundamental notion of perception and action. + +But we're just going to completely assume actions, and we're going to look at perception only. + +So we're going to have to try and formalize what it means to be doing perception. + +And that's the entirety of part one, really. + +And that's going to carry through to the remainder of the book as well. + +so chapter one this was all about uh perception as inference okay so the whole point of chapter one was that we have this question what is perception how do we do perception and how are we going to think about perception in the context of the modeling paradigm that is active inference and the idea fundamentally is that perception is inference we have beliefs about hidden states of the world uh we do not directly observe hidden states we only observe + +through our sensory apparatus. + +And on the basis of those operations, we can update our beliefs about hidden sensory, well, hidden states of the world. + +And that process is inference. + +So inference is just the process of changing your beliefs about something, right? + +And that's what we're going to mean by perception. + +And there's a whole story to be woven here. + +Well, before we get to that story, the most fundamental distinction is this distinction here. + +Where is it? + +Yeah, this one. + +This, again and again and again, is the most fundamental thing. + +And we're going to see something like this diagram repeat throughout the book. + +We've already seen it repeat throughout the chapters that we've gone through so far. + +So the fundamental separation, the really bedrock starting point, is this idea that we have the environment, which is in some sense the real world, and we have the agent, which is somehow contained in the real world but not identical to it. + +And they are separated from one another. + +Okay, so there's some process out there, there's some dynamics, there's some physics, whatever that's quote unquote out there. + +And the agent is trying to predict what is going on in the environment. + +And it can only do that on the basis of observations. + +so observations come in and never the agent never gets access to the real world the real environment but on the basis of those observations the agent has a model of the real world and using that model it's able to update predictions or beliefs about the real environment okay and those predictions and beliefs we see in chapter two + +we formulate this idea of perception as hidden state inference using Bayes' rule. + +So chapter one was all about what is perception? + +It's this process of inference. + +And we're doing inference because we're separated from our environment in the sense that + +there's a boundary between the agent and the environment and chapter two that makes this precise mathematically okay so we actually start doing some mathematical operations here in chapter two so we we know we're doing inference but specifically chapter two says we're going to be doing Bayesian inference okay so and the the recurring the recurring you know running example throughout this book is this really really really simple example where we have some notion of an agent that's farmed for food + +So the idea is that some food out there in the environment has a certain size. + +That's the state of the real world. + +I'll get this picture up here. + +So the state of the real world is out here in the environment, there's some state. + +It's food sizes and arbitrary food sizes from 1 to 5, let's say. + +And those generate observations. + +And the observations are light intensities. + +So the observations are related to the hidden states in some way. + +We don't actually know really how the observations are related to the hidden states. + +But what we say is we're going to assume that there is some kind of relationship. + +And our assumptions about the relationship between the hidden states and the observations, between the food sizes and the light intensities, those assumptions are going to be our generative model. + +And we're explicitly reasoning probabilistically. + +So the entire North Star + +the whole sort of guiding apparatus of what we're doing is we're trying to, as best we can, implement Bayes' rule, which is the rule that tells us how to update our beliefs about hidden states on the basis of sensory evidence. + +And Bayes' rule tells us this precisely and exactly. + +So what do you need? + +You need a generative model that encodes our assumptions about the relationships between hidden states X and observations Y. + +And we've seen that we can factorize this generative model in terms of our likelihood. + +So we have a probability distribution of our observations given hidden states and prior beliefs about hidden states. + +Those two things together constitute our generative model. + +And then the last piece, the missing ingredient that we need + +is our model evidence. + +This is the probability of observing something independent of any state. + +And those three things together, they compose Bayes' rule. + +And then we can get our posterior belief about hidden states on the basis of sensory evidence. + +And then bingo, bango, bongo, we're doing inference. + +Very nice. + +So that fills in the picture of what's going on + +at least abstractly, mathematically, ideally, inside the agent. + +We're doing inference by means of Bayes' rule, or at least we're trying to do that. + +And indeed, that's what we did in chapter two. + +We used Bayes' rule exactly. + +We made certain assumptions about how we were going to be doing things. + +We discretized the state space. + +We were able to do exact marginalization. + +We waved our hands a lot, and we got rid of a lot of problems. + +I see there's some people in the waiting room still. + +Let me just admit some people. + +But that was crucial. + +We assumed that we could do exact basis rule. + +Do I have the waiting room? + +Let's have a see here, participants. + +Admit, admit, admit. + +All right, very good, very good. + +Excellent, all right. + +I shall share my screen again. + +Very nice. + +So we have this notion that what we're doing is we're trying to implement Bayes' rule, at least ideally. + +That is what it is to be doing inference inside the agent, by means of our model. + +And then out here in the real world, in the general process, there is some actual relationship between observations and hidden states. + +We never know this exactly. + +But of course, as modelers, we're in the unique position of being able to specify this. + +So for our really trivial example, + +we said, okay, and typically throughout all of chapter, well, part one, the relationship between hidden states and observations has been linear. + +Okay, that's one kind of relationship, very basic kind of relationship. + +And that's been the case in the model and in the process. + +So that was chapter two, where we actually sort of, you know, rubber meets the road. + +We did some real calculations. + +We saw what it is to update beliefs by means of Bayes' rule. + +okay and i should note just by and by the by that what we're doing we haven't made a big song and dance about this yet but all of our beliefs have been with respect to or by means of continuous probability distributions so these these bell curves okay gaussian distributions um that's and we're going to continue this in part two + +and going on, but we're also going to see a distinction that will come up later when we put action in for simple models where instead of reasoning about these bell curves, we're going to reason about categorical distributions. + +Instead of a continuous interval, we're going to be dealing with discrete things, but that's going to come later. + +The big takeaway from Chapter 2 is this here, I think at least. + +We're reasoning about things in terms of Bayes' rule, + +We're updating the posterior over hidden states on the basis of sensory evidence. + +Very nice. + +And we're doing this exactly. + +We're doing exact Bayesian inference in chapter two. + +And the big thing that we saw is that if we do this, there are certain relationships that hold between the prior, the likelihood, and the posterior. + +So what we saw is that depending on the precision, the degree of spread or uncertainty that you have, + +about your prior and your likelihood that affects the precision or uncertainty of your posterior beliefs so and you know canonically you can sort of mix and match but imagine you've got a very precise prior belief okay and a kind of spread out likelihood of observations this in some sense encodes an agent that's very stubborn one that doesn't really want to + +its observation model very much. + +It wants to trust its prior beliefs. + +And so you see, OK, well, if we do a Bayesian update with a model that looks like this, the posterior that we get is pretty close to our prior belief already. + +So that's encoding an agent that's stubborn or that doesn't want to change its beliefs very much, kind of a closed-minded agent. + +On the complete other side of the spectrum, if your precision + +for your your prior is quite spread out so the agent doesn't really have a very fixed idea about what it's what the real world is going to be like but it's going to it's very precisely attending to sensory evidence then when we do a bayesian update you can see that the posterior is much much more like the uh the likelihood model so it's trusting the observations it's getting in more than it's trusting its prior beliefs okay and it's precisely this precision + +or variance or uncertainty on either the prior or the likelihood and how they how they relate with each other that affect the precision or the uncertainty of the posterior belief so that's something that you get with basis rule you start being able to reason about the uncertainties um in our beliefs our prior beliefs and like well yeah is that a question i hear oh well + +I hear some grumbling. + +Hang on, stop sharing. + +Was there a question? + +Going once, going twice. + +Sorry, my bad. + +I thought I was on mute. + +Okay, no problem. + +All right, I'll just share. + +Yeah, I just want to briefly go through just so that we're all on the same page, at least at a high level what we've sort of gone through. + +So that was chapter two. + +Chapter three then was a bit of a change of pace. + +The idea is that, yes, very nice. + +We've seen how we can operationalize this procedure of inference for perception. + +OK, but there were assumptions that we made in chapter two, namely that we have a model. + +OK, and we're going to use that model to do inference. + +And we saw lots of different ways to do inference in chapter two. + +But of course, and I want to just show you, where is it here? + +So we go, this is this very central diagram here. + +So we have a model and we did inference with our model. + +Of course, our model is, you know, there are certain settings of parameters of our model that need to be set well, if we're going to do inference well. + +Okay. + +So our model is made out of probability distributions. + +We have our likelihood, we have our prior. + +These have parameters. + +These have variances. + +Our observation model has parameters theta. + +Because it's a linear model, typically, we have a slope and an intercept. + +Those are parameters. + +Those are things that have to be set well enough so that when we do inference with our model, we're going to be doing inference well. + +So there's problem number one, how do you do inference? + +But above that is problem number two, and it's kind of a harder problem. + +What should the setting of the parameters be for my model? + +In chapter two, we just completely ignored that problem. + +We assumed that the parameters of our model, the settings of the dials of our model were already perfect. + +They were set to be excellent, and then we can just do inference. + +And we told that story. + +But realistically, in the real world, your model has parameters and you need to know + +how to set those parameters. + +And we're going to see this more and more going forward. + +But this question about how to set the parameters of your model, this is typically operationalized as what it is to be doing learning in active inference. + +So we've got sort of fast changing things down here, hidden states. + +How are those changing? + +That's inference. + +But then above that is the question, all right, how do I learn or how do I do inference on what the parameters of my model should be? + +So they're two different but related problems, basically. + +And realistically, you need to be able to do both of them if you're going to be doing inference well. + +So chapter 3 was all about different ways of estimating model parameters. + +And we saw how to do deterministic model parameter estimation. + +We saw techniques like gradient descent. + +It comes up again and again and again. + +Maximum likelihood estimation, Bayesian linear regression, all very nice. + +This culminated in something called the expectation maximization algorithm. + +That was the jewel of Chapter 3. + +Why is that important? + +Why did we care about the model premise? + +Just briefly before I answer that question. + +In some sense, this figure here, this is a very nice depiction of the process of parameter learning that we saw in Chapter 3. + +broadly speaking and we're looking at a process of gradient descent here and gradient descent is a very fundamental algorithm that shows up everywhere in machine learning um and it's going to continually recur across the textbook the general idea is let's say i've got two parameters okay for my observation likelihood in this case i've got beta zero and beta one you know you could have hundreds of parameters or thousands of parameters but let's say we've got two + +And we have some notion about how costly those parameters are. + +So these two parameters here, they sweep out a two-dimensional space. + +So a point in this space is a setting of beta 1 and beta 2. + +And this surface, this curve, is some kind of notion about how costly those pairs of parameters are. + +So what we can do is if we can operationalize a notion of cost, + +We can start somewhere on this surface, and we can say, all right, let's just walk down the surface in the direction of greatest steepness, right? + +And we're just going to keep doing that for some number of steps or iterations. + +And once you get to somewhere that's sort of vaguely flat, a local minimum, maybe a global minimum, but typically a local minimum, + +We're going to say, all right, well, you know what? + +The setting of those parameters where the cost is basically flat, those are going to be good. + +And those are the parameters that we're going to use for inference. + +That is, in so many words, the basis of gradient descent. + +And that's going to show up again and again and again. + +But it's different from the problem of inference. + +However, there are interpretations about what it means to learn. + +such that we can treat parameters as themselves another kind of hidden state and we can do inference on those okay so that that idea is going to show up in chapter five at least in disguise when it comes to this idea of separation of temporal scales so you've got fast changing states down here and slower changing states up here and these are your parameters and they're coupled to one another all right so the last thing on chapter three + +very nice that's a nice story about how to do parameter learning but you need to do parameter learning and inference kind of at the same time right basically and there's there are chicken and an egg okay you need you need to be able to do inference in order to be able to learn parameters but you need the good parameters in order to be able to do inference so it's it's a very vicious problem and the expectation maximization algorithm uh is the thing that solves this chicken and egg problem where + +We assume that we can take the expectation about what parameters should be given beliefs about hidden states. + +And then on the basis of that, we can maximize the evidence for certain parameters and we can kind of do them jointly. + +And that solves the problem. + +So we can do joint hidden state inference and parameter learning at the end of chapter three. + +And that's the whole point of chapter three. + +That's the really central thing that it's about. + +But there's a problem. + +The big problem is that the expectation maximization algorithm, at least of the kind that we've seen here, assumes it assumes that we can know that we can we can evaluate the true posterior. + +So if I go back to Chapter two. + +assumes that we can exactly calculate the true posterior belief okay and unfortunately that assumption is just usually wrong it's usually the case that we simply cannot evaluate the true posterior for any number of reasons the big reason is because marginalizing out the um model evidence is usually very very hard if not impossible and so you just can't use base's rule so we're kind of stuck you know we know abstractly what we + +should be doing if we wanted to solve the problem of hidden state inference and parameter learning but we can't actually do that so chapter four this is really where we kick into high gear in terms of the you know mathematical foundation of active inference itself this is why we say look we can't calculate the exact posterior so we're going to approximate it we're going to find a posterior that's close enough to the real posterior okay + +the process by which we do this is variational inference so we're going to assume that the real posterior takes a certain shape or is a certain kind of probability and usually we're going to assume that the posterior is a gaussian right or a bell curve that's an assumption the real posterior might be something wiggly it might be something very strange okay but we're going to assume that uh at least it's good enough to assume that the true posterior is is a gaussian + +I don't have figures here just yet. + +And then there's a question about, all right, we're gonna try and make that Gaussian, that bell curve as close as possible to the real posterior. + +And how do we do that? + +Well, there's a quantity called variational free energy, which bounds the surprisal, bounds the negative log of the model elements. + +And this thing we can calculate. + +So on the basis of that, we can do variational free energy minimization, + +And there's various ways of talking about how to do this in terms of free form, mean field, fixed form, mean field. + +We didn't end up talking about those very much at all, actually. + +And this is something we can do. + +So this is actually how we're going to do our approximate Bayesian inference. + +Admit, admit, admit. + +I see, Andrew, you have a hand up. + +Sorry, I didn't notice that before. + +Would you like to jump in? + + +SPEAKER_04: +Oh, yeah. + +Yeah. + +No worries. + +It is going to be brief. + +I didn't want to slow your momentum. + +But yeah, just on on parameters, I think for those who are still wondering, like, what what are these extra parameters? + +If we've been talking about things like perception and taking in a, you know, a sensory observation and then inferring a hidden state. + +So what are these parameter things sort of about? + +You've already addressed that, Fraser. + +But it's just to say. + +um you know we we take in an observation we sort of update a belief so if i were to use more colloquial terms here to think provide more of an intuition um it's sort of like you know we see our y and we update our x um but how do we do that right so so if you think of of the the you know the y is like what you perceive and then and then x is your beliefs then parameters are sort of like + +your understanding they're they're what help relate the x and the y right because without some kind of understanding so to speak of how um you can link up your your beliefs with what you see or what you observe um like without any kind of parameters like we have no way of relating those two things right + +So it's the parameters that kind of lead to this equating different sorts of things. + +It's like, oh, well, X is simply, you know, we take our Y and then we run it through our understanding. + +We run it through our parameters and the sort of combinations of these things, right? + +So that's sort of what's essential about this idea of parameters. + +And then, of course, why we have these things like chapter 2 and chapter 3, right next to each other, because they're, they're so sort of inextricably tied up with each other. + +And then, as Frazier was going into now. + +variational inference then becomes this sort of tractable means of running that sort of process of going from our y to our x as well as being able to invert that to go from an x to a y via parameters and being able able to to update those parameters over time that is they don't have to your understanding of how states and and observations also does not stay fixed that's why we have not just inference but we also have + +But, yeah, that was also just supposed to be, like, kind of a brief point along the way. + +I think your explanation thus far has been really great tying things together. + + +SPEAKER_02: +Yeah, no, fair enough. + +Thank you, Andrew. + +I mean, we've seen along the way various representations of our model in terms of maybe a graphical representation. + +This is one representation here where very quickly, this is exactly expressing the idea that, look, you know, there are hidden states that generate observations, but our model about how observations are generated from hidden states is dependent on parameters. + +And we can also have beliefs about parameters, okay, that are + +parameterized by further other parameters and we can update our belief about parameters in addition to getting states so you can kind of do inference on parameters as what it means to be doing learning so that's an interpretation that shows up again and again at least later but um we've seen the beginnings of it here so uh yeah so okay so chapter four this is this is you know i i say where we're + +introducing how it is that we're really practically going to do approximate Bayesian inference by means of variational free energy minimization. + +We spent a long time motivating where it comes from, what VFE even means, why it bounds surprise, and why that's a good thing. + +But the upshot is that once we can formulate our inference problem in terms of variational free energy, this is something that we can solve practically. + +That gives us the mathematical machinery + +to be able to do the inference problems, attack the inference problems we want to attack, in addition to being able to learn parameters. + +And that's going to be the mathematical foundation for the rest of the book, really. + +So that gives us the story of perception so far. + +But the story about how action comes in and how actions are selected is a very, very similar story. + +We're, quote unquote, just going to be minimizing variation of free energy. + +But we're going to see a little bit of a wrinkle + +it means to be doing planning and expected free energy but again that's coming later much later actually now in some sense okay so you know chapter five comes next right now that's still part of part one but really i tend to think that chapters one to four is sort of spiritually part one you know so i think part one kind of ends at chapter four in some sense chapter five is kind of a historical + +or motivating additional perspective on what chapter 4 is about. + +So chapter 5 is about this other thing called predictive coding or predictive processing. + +And really, the whole idea is that we can re-express what it is to do variational free energy minimization in chapter 4 in terms of another process called prediction error minimization. + +or more precisely, precision weighted prediction error minimization. + +This is just a different perspective on what it is to do via theme minimization. + +And historically, it sort of comes from a different or slightly adjacent tradition in neuroscience and the Bayesian brain hypothesis, where we're explicitly reasoning about these things called prediction errors. + +So it's not really anything new or additional. + +It's kind of, as I say, just a different perspective on what it is to be doing via theme minimization. + +It's an important one. + +It shows up in implementations every now and again. + +um and it's it itself has been given um the ability to to do action selection which is very important but it's it's not it's not a constituent like core idea to active inference itself it's really just a different perspective so that is all of part one + +Very nice. + +I'll just mention before I forget, Andrew has, and where are they, Andrew, your slides? + +So I've put together these comprehensive sort of notes and so on, where you've got your, what are the key ideas in the chapter, what's core, what's not, and then I have the same sort of thing for the sections. + +But Andrew, you've got your slides, which are even more. + +So this is quite detailed, but I like your slides because they compress a lot of information in a very tight way, which is helpful, I think. + +Are they in the code? + +I always forget where they are. + + +SPEAKER_04: +yeah yeah no worries um i've kind of pushed daniel to put them somewhere more obvious and he's still not doing it uh so so if you go to figures um okay yeah uh which was just a couple above where you're currently at um so we've moved on yeah yeah so so if you go to figures and then you actually scroll all the way to the bottom like actually after + +Um, these, yeah, they're tucked away there. + +And then also on the, the initial landing page. + +Daniel's put a link to what he calls supplemental materials, and that will just take you directly to the slides as well. + +But, yeah, we're going to try to put those somewhere. + +Yeah. + +Supplemental. + +so yeah hopefully we could get those in a more obvious spot and the big thing is that you know they're because they're not complete um you know I think it's fine to just repeatedly share that they're sort of in progress right um but yeah the the the key idea is that uh one not everyone has the textbook and so this is like in addition to the coda this is attempting to + +kind of put together a lot of information from the textbook without directly reproducing the textbook and kind of giving trying to reduce some of the redundancy because for those who have been following through the chapters you've actually seen like we even have one-to-one direct repetitions of some of the equations from chapter to chapter because of how tightly they're linked like it is warranted that there's some kind of repetition going on from chapter to chapter + +So it's just intended to focus on particular concepts. + +Some of the earlier slides will go into even more basic sort of things that we're getting into that even the textbook doesn't touch on. + +The textbook kind of assumes that people might be taking more of an engineering direction. + +something like an equivalent to undergraduate coursework in statistics or otherwise. + +So in some of the earlier slides, I say like, well, what is a continuous distribution? + +And what is a probability density function? + +And all these other sorts of things that we rely upon. + +through the textbook that don't necessarily get spelled out for you, and what are some more intuitive ways of thinking about these sorts of things, especially for those who are not coming from an engineering background but hopefully want to get the most they can out of the book. + +And then you get, you know, we talked about how gradient descent itself is such an involved process, and yet it really gets kind of summarized within two pages in the textbook. + +So having a slide dedicated to that, you know, + +I make some references to the earlier 2022 textbook as well to sort of show some of the continuity between that, which was not focused on engineers, versus this textbook, too. + +So, yeah, hopefully it provides any kind of additional value to those who are wanting to get the most out of this textbook, whether they own it or not. + + +SPEAKER_02: +Yeah, I really appreciate this because we've got at the moment 20 slides here, maybe one or two slides on average per chapter. + +If you wanted a first point of contact with the material other than the book, obviously my notes and so on, there's a lot of that. + +There's still less in the actual book, but this is quite a nice way to maybe not mistake the forest for the trees. + +This is a nice... + +synoptic overall kind of feel for what the what the concepts are and where they come from i mean it's a bit hilarious saying that given how dense each slide is you know you've done a very good job andrew of uh i think walking that tightrope but do uh do take a look at andrew's slides they are a nice um initial global view about what we're talking about + +And then hopefully you can sort of titrate between there and maybe my notes as to what level of abstraction you want. + +And that is specifically, as you say, for people who maybe don't have the book. + +Obviously, the book is always, you know, that's the final port of call. + +But for those who don't have the book, hopefully you can get the same kind of benefit or at least a similar level of benefit from those who do. + +so that is the end of part one we're going to be moving into part two which is explicitly about how we put action into this story and that really is going to be active inference right so chapters + +well for the first chapter of part two generalized filtering for perception we're not going to immediately jump into what it means to be doing action we're still going to be within the perceptions of story but it's going to introduce this fundamental algorithm called generalized filtering which we can use to do perception but we can also use to do action okay and then chapter seven is active generalized filtering uh so really and then you know once we do that we can then tell the story about learning again + +tension and hierarchical models again okay so we told the story of hidden states or hidden state inference in two we told the story of learning in three and then we told the story of approximations of this in four and hierarchical um approaches in five that same flavor is going to be recapitulated in two again okay but just with the story of action coming in so six seven eight or maybe seven eight nine + +these are kind of the core chapters of the book so it's going to be really cool once we get to those i guess i'll open it up for questions for live questions if there are any um please do not hesitate maybe i'll stop sharing just now i'd love to if there aren't any live questions immediately or if people would rather put them in the coda that is certainly advisable um i'm trying to make an effort of going through the coda more um with more frequency than i have before so + +Yes, if there are questions, now would be an excellent time. + +Carrie, I see your actual hand up. + + +SPEAKER_01: +My real hand, my real face in real time. + +I was really struck by a comment you were making in reference to chapter three and learning in terms of the stubborn agent and the open-minded agent. + +And those are my framing of things. + +um and it was those three prior likelihood posterior you know um graphs and um because this um this question around um what are we attending to in our environment was sort of alluded to in your description and i i just wanted to maybe + +give a little bit more space to talk about it, because I think it's a really interesting element of active inference that I wanna, you know, investigate a little bit + +more thoroughly, especially with the idea of the stubborn agent being fixed on that, like what happens when you get stuck in thinking what you already know versus how do you create conditions that someone might be able to take something in + +to um to do more broad learning and it sounds like or at least what i was hearing for the first time in a new way um was that it had something to do with that likelihood function where you're more precisely observing certain things um do you know are you still looking for the the graph no i found it i just i wanted to the graph is in chapter two uh i'll actually just name it so it's ticket 2.11 this one here yeah + + +SPEAKER_02: +yeah so anyway i don't i don't have a specific question i just wanted to to sort of dog ear how cool this is to me nice yeah well i i completely agree and i i spent a little bit of time as as i did in chapter in the the review for chapter two hopping on about why this is a cool thing because we've actually seen in chapter five with predictive coding this story has been told again + +And it's been told, we've elaborated the idea, well, we've investigated the idea more precisely about, okay, well, what should I attend to? + + +SPEAKER_03: +Okay. + + +SPEAKER_02: +Because the whole point of chapter five with the study of prediction errors is there's prediction errors all over the place. + +You know, there's prediction errors here, prediction errors there, whatever. + +But only some of them are worth paying attention to, typically. + +And that's where in that chapter, in chapter five, we motivate this idea of precision weighted prediction errors being the important thing, not just the prediction errors themselves. + +So that's a more sophisticated notion about, because I think from your question, what you're gesturing to is there is this question about what we should attend to in some sense. + +And there's various ways of attacking that. + +At the coarsest level, if we just sort of say, all right, I've got prior beliefs about what I think I'm going to observe, I have a likelihood mapping. + +The dance or the interplay between those two things affect how I'm updating my beliefs in terms of precision and such like. + +That's one way that's the question about how I'm attending and what I'm attending to affects my belief updates. + +but in a in a similar vein it's not quite the same idea but in a similar vein with chapter five we then additionally layer on top this notion that there are certain kinds of prediction errors that are that are worth more than others okay and it's the precise ones that we want to talk about so there's a there's an example i like uh fog don't know if i'll be able to find it in time but + +A motivating example is that, you know, let's say you're a car in some sort of foggy environment. + +You're perceiving things. + +Okay, maybe this is a nice one here. + +You can obviously observe, by means of your likelihood model and prior belief, you can update your belief about what you're observing in the fog. + +But it would be better to look at clearings in the fog than to just look at anywhere, because those clearings give you very precise information. + +So you've got some prior belief about what may be in the fog. + +But if you can attend to areas in the fog that are going to give you precise feedback, that's going to be better than attending to areas in the fog that are not going to give you as precise feedback. + +So that's where this notion of precision weighting comes in to add to the story about this question of stubbornness of the agent or not. + +So Andrew, is there anything you'd like to share on that front? + +Because I know you've + +you may have some additional context there being a social science guy. + + +SPEAKER_01: +And before you jump in, Andrew, just real quick, is the likelihood function, what's the point of view of the likelihood function? + +Is that likelihood function for the stubborn agent based on the fact he's like, is those two likelihood functions, the same environment perceived differently based on your priors? + + +SPEAKER_02: +Well, yeah. + +So the idea is that you could have exactly the same likelihood. + +In this case, we don't. + +And where the only difference is the precision of the prior belief, and that would affect the belief update. + +OK, so you could hold the likelihoods function, the likelihood mapping equal in both cases and vary the prior precision. + +Or you could hold the prior to be the same and vary the likelihood function. + +And you could see this same relationship here. + +So is that what you're asking in terms of maybe you've got two agents and they have the same likelihood mapping, but they have different prior beliefs? + +That's certainly possible. + +That's very, very possible. + + +SPEAKER_04: +This is one of the best figures. + +I think this figure is one of the most illustrative in the textbook that we've seen thus far, as much as I appreciate the descending the contours, et cetera, whenever we look at gradient descent and all that. + +But I think for building an intuition, yeah. + +So with this figure, it's just like the main idea is we have two rows of plots. + +They show the same things, but the top row is for one particular agent in a particular situation. + +And then the bottom row is for a different agent. + +So both of these agents could be in the same environment and even observing the same exact things. + +But the one in the top row had a higher precision on its prior, meaning it has a much sharper and potentially more confident view of its own prior beliefs. + +Whereas in the bottom row, it's the opposite situation that the agent has a much sharper, more precise belief and confidence in its likelihood. + +So what's that mean? + +It means that the prior beliefs that you sort of bring to the table + +Are are going to be sort of a bias, right? + +They're going to be a sort of bias of, like, before I even see any sensory information before I even take any observations into account. + +What do I think the hidden state is such as, you know, in this example, it's like the hidden state. + +It has to do with food availability where the observations are light intensity. + +Right? + +So it's likelihood, the likelihood on the other hand, as opposed to the prior, the likelihood is what relates hidden states and observations, right? + +So that's sort of the way that you translate + +the world and what you observe into your beliefs and vice versa when we invert the model. + +What kind of observations should you expect based on your prior beliefs? + +So if you have an agent who has very precise priors and a much less precise likelihood, they're going to rely much more on their beliefs that they already bring in advance. + +Uh, to to the situation, and this is such an interesting thing as far as, like, cognitive phenomena and thinking about how not just that, like, this sort of micro scale of neurons, but how potentially, you know, entire organisms, including people. + +operate whenever someone has such a strong prior belief, it can sort of dominate their ability to interpret the world around them because they're going to down weight the observations that they're actually taking in. + +And those, those, those strong, precise priors that they bring to the table. + +Yeah. + +Are going to to skew how they take things. + +And so so. + +That relates to your original question regarding, like, what we actually attend to, because if you have very high, these sort of the. + +In computational psychiatry, the phrase hyper-precise priors has been brought up. + +I think the hyper is just sort of trying to make the point that the prior beliefs are so extreme, sort of exacerbated, that they can lead to all sorts of interesting, potentially unfortunate misinterpretations of sensory information. + +So someone who doesn't trust + +The observations they receive, so if it has to do with 2 people communicating with 1 another, if you have very, you know, a very strong prior belief that you don't like this other person, or there's something else going on there that that you're really going to rely on your own prior beliefs about this other person you're attending to. + +rather than actually listening to what they say, the kinds of things you can actually observe from them, then you're not going to update your beliefs very much about them. + +You're going to keep your beliefs centered around the priors that you bring in. + +You can think of notions of things like stereotyping and prejudice start to play a role here. + +And then also we can think of things like + +The reverse situation is where someone has a very precise likelihood and very imprecise priors. + +And in that case, you're going to be very reactive to the environment around you. + +You're going to be very focused on the observations you bring in, and you're going to be rapidly updating your beliefs, depending on how + +much those observations change. + +Your prior beliefs can actually add a sense of stability, right? + +Like if something weird happens in your environment, you might say, oh, I don't know if I trust that observation I received. + +I can rely on my prior beliefs to kind of keep me grounded here. + +If you have very + +Imprecise priors than that saying that, you know, oh, I, I have no grounding. + +I'm hyper reactive to the environment around me in some ways in some of the psychiatric literature that's been related to tentatively some of the sorts of things we see in autism. + +Actually, it's a, it's related to, you know, the idea of almost like, hyperactive, sensory, excuse me, hyperactive. + +processing of sensory stimuli, such that your prior beliefs that would kind of make you more grounded in the situation. + +We're not just purely talking about how someone consciously thinks. + +We're also talking about the way that their brain specifically is, you know, kind of interacting with this information, what + +that's where the predictive coding comes in, right? + +That's where the sorts of things that we're seeing later in these later chapters that we've seen so far, they start grounding us a little bit more and starts relating more to neurobiology. + +But yeah, to be so strongly reactive to sensory information in the sense of like having this kind of hyperactive likelihood that dominates your priors, yeah, it can be very disorienting for someone. + +It's just being very reactive. + +I don't want to overspeak because I'm sure that some of what I've said may or may not have been totally clear, but it's worth reflecting on that figure and the idea of what is it about precision in the priors versus the likelihood, right? + +And as we get into later chapters, we'll look at + +how we can have precisions on other things too such as decision making you know what happens whenever someone has very imprecise sort of deliberate uh uh well + +the technical phrase will be policy inference, but basically what happens when you have low precision on your ability to make decisions. + +It would be like someone who is able to come up with decisions they could make, but they might not trust their own ability to follow through on those decisions, right? + +So those would be people who have something like low confidence + +in their own decision making. + +And that's the kind of thing that we might start seeing exhibited in cases of like depression and other kinds of depressive + +symptoms, or at least we'll call behaviors, right? + +Like someone who's still able to think through, they're still able to observe different things, but they might not be able to have a clearer sense of agency in a situation. + +So they have very low precision on their ability to make decisions. + +So yeah, we'll find precision sort of all over the place in future. + +And it's something very interesting to model for sure, because it has all kinds of repercussions + + +SPEAKER_00: +uh for for the kinds of varieties of behaviors that we see in people uh in organisms otherwise um yeah yeah no it's it's all pervasive that's for sure um gian cuomo did you have a i hope that's how i pronounce your name did you have a question i saw your hand up before i yeah just very briefly it's it's it's it's i don't want to drive it far from here but again say you have very strong priors so you have very concentrated so your prayers have uh you know you're + +are very pointy, you're very, you're, now, this is, is there, what kind of, what kind of architecture do you, so I can imagine an agent who is so very opinionated, who has a, because, you know, I'm generalizing maybe too much, but I'm thinking somebody who has too much, he's been, he's, I think about the human being receiving education, a very rigid education, for example, who, and so in a way he's, + +is very front-loaded in terms of his opinions about the set of matters and therefore he assigned high probabilities to a very narrow set of values if you will i'm sorry and now two you know as one grows old over time i'm sorry i hope the metaphor holds uh you you know you can become sometimes wiser because you accumulate knowledge you accumulate experience you have + +You know, you build in this sophisticated, layered mental and mnemonic database of other people, other relations, their feelings over time. + +You've observed actions and reactions, consequences to actions. + +You observed moral conundrums unfold and you become + +You know, you actually, you see them and actually, so you're bringing all this in now. + +This, what, what, what kind of architecture, how can, how will, how do these active inference models expand to provide, uh, a, let's say a mental temple or a gallery, uh, of say, for example, memory. + +So how can the age, because I can see the agent on his own. + +He has a likelihood, so he has a certain experience of how a certain observation is tied to a certain theoretical value or a certain value, and then he has an opinion about that value. + +And I cannot see an inherent mechanism there for him to say, whoa, wait a moment. + +My priors are a little bit too loaded. + +I'm a bit too rigid. + +I'm a little bit too... + +Biased I have to change my priors now this and so where do you in what directions? + +what kind of interactions and I guess I get circularity what kind of feedback mechanisms other than I mean this kind of these the kind of Hierarchies that I have to think about the ones that we looked at in predictive coding for example when you have maybe several agents that are working at different speeds and then they + +they talk to each other or they question each other and they say hey one agent may observe the others and say hey look there is a you know there is a great majority of my peers that are actually different that have actually different priors why don't i change them to match them for example yeah yeah this so this question came up in the first of the review sessions as well this is a an enormous question and it is actually one of the open + + +SPEAKER_02: +Well, there's a way in which it's a very trivial question. + +And there's a way in which it's a very, very interesting and deep question. + +The way that it's a trivial question, I will answer first, which is that yes, at the end of sort of chapter five, we saw this idea of hierarchical predictive coding, where the idea is that each layer is in some sense, minimizing local variational free energy, predictions down and you get updates up, okay. + +And what you can then do is you can interpret each layer as having preferences that are in this message passing loop that equilibrate with each other. + +So that would be a sort of account of how it is that some local agent-like thing + +One layer in this hierarchy can attenuate its preferences over time. + +That would be number one. + +That's a kind of very simple way of talking about things. + +Another way is where you generalize this whole thing. + +You want to talk about the agent as a whole interacting with other agents. + +So if I go here as as many as one, I know we're coming up to time everyone, but I just want to. + +For us. + +You might want to tell the story about how interacting agents attenuate their preferences with respect to one another. + +So there's very little work on this at the moment that I'm aware of. + +But there's this idea of federated inference and belief sharing. + +I'm struggling here to + +to communicate how it is that this can be done, because it's very, very early days yet. + +But this is a question. + +And the more general question that I think lurks behind your question is, where do my priors come from? + +Where do my prior beliefs come from? + +Of course, as engineers, + +we're trying to solve a particular problem we're in a very privileged position where we can just say okay well you know i want i want the thermostat to be at a certain temperature this temperature here that's your prior very nice okay but in a lot of problems a lot of situations it's very unclear what the prior should be um and so that question about how it is that we can actually specify priors or how it is that we can recognize when our prior is a bad prior that's a very hard question i tend to think + +that it is precisely in these notions of hierarchy, hierarchical predictive coding, and more generally compositional notions about how agents interface with one another, that this question will be best addressed. + +So I think if you were interested in that question, it would be useful to really delve more into the history of predictive coding and this idea of hierarchical predictive coding. + +So I hope that is somewhat intelligible, or maybe you already understood all of that. + +Those are my thoughts, at least. + +Do add to the coda as well, these questions. + +I think I'll do that because it's very good for other people to see these questions and maybe myself and Andrew and Daniel answering them alongside. + +So maybe if there's one more question, we can take that live and then we'll have to sign off. + +We're probably going to do another week of review. + +OK, so another week of review for part one. + +If you do want that, please tell us that you would like that. + +If you don't want that, please also tell us it does. + +The previous session, the previous group were very much in favor of another week of review. + +But it's up to you guys. + +It's up to us basically. + +So I think as things stand, we're going to next week is going to be yet more review. + +Hopefully we can have some more questions and even some additions to the code. + +That would be nice. + +Andrew. + + +SPEAKER_04: +As a final point, there was an interesting question asked in the chat regarding what we were talking about with precisions on likelihood and priors. + +Does this have any relationship with some of these cognitive phenomena? + +I'll be honest, I've not seen these terms used too often. + +what are called the halo, horn, and golem effects. + +I mean, I've seen many different sorts of terminologies from different areas of psychology, of sociology, etc. + +It's interesting how often we see in discussions of prejudices, of cognitive bias, of interpretation bias, we have this + +this like sort of plethora of different kinds of terms and terminologies that play out. + +But it's interesting from a standpoint of active inference is that we might be able to relate many of these different kinds of phenomena and potentially even integrate them + +whenever we're able to have some kind of like common like premises for thinking about these sorts of things, perhaps priors and likelihoods and thinking about things like beliefs as a way to do that. + +But it was a really good question. + +Yeah. + +So the Gollum effect, which was an interesting one. + +Yeah. + +Yeah. + +yeah when we we lower our expectation it's kind of this does get into uh well we've probably already seen this phrase in the textbook um but this idea of self-evidencing um that's something about active inference one of the premises is that we kind of minimize our uncertainty and that's not just active inference it's more the the bayesian brain hypothesis and predictive processing and all the rest more generally we minimize our uncertainty + +And so part of how we can minimize our uncertainty is not just to change our minds by learning new things about the world, taking in observations and updating our beliefs. + +Another thing which we'll see much more of in part two of the textbook is, well, we can also act on the world, right? + +Like if I'm very hungry and I don't want to be hungry, like, + +If I just sit here and think like, well, I'm hungry, that must mean I'm going to be hungry all the time. + +I'm not going to do anything about it, right? + +That's just perception. + +That's not engaging in action. + +That's not getting up to go find myself some food and then eating it, right? + +That's action. + +That's decision-making. + +That's actually doing something in the world. + +So with this like Gollum effect sort of thing of like, oh, or the halo effect or the horn effect, it's that we can take in some amount of information about another person. + +And then that impact, so that's likelihood, right? + +We're taking in a sensory observation about another person, interpreting it as some belief about that person. + +But then from there, once we think we have a pretty good belief about that person, what they're about, or what our beliefs about them are, + +we could increase the precision on that. + +It's like, well, I didn't have a very strong prior about them because I didn't know anything about them. + +But now that I do, I'm going to increase that precision. + +Again, I'm saying this colloquially. + +I'm not saying that we're consciously thinking I'm going to increase my precision, right? + +But it might just be that that's how you approach the world. + +That's your model of the world. + +once i learn at least some sufficient amount of information about someone i've now established a prior about it and i feel pretty good about it confident about it so there's a high precision on it uh in terms of the math and then um you know that that can overshadow + +any future observations you receive from the person like oh this person uh must be very you know i i saw them uh you know do something nice they that i watched them give a donation to to a non-profit or something they must be a very good person they must be you know uh interested in philanthropy and other things right and all these other things that tie into other prior beliefs that you have about other people or about how the world works too they can all get related + +But then once that prior belief becomes really, really sharp, then what might happen is you observe something not so good about that person. + +Oh, they, I don't know, they drop trash on the ground. + +You can see I'm really trying to reach for everyday examples here, but they litter or they, oh, they don't seem like they care too much about their family. + +They tend to ignore phone calls, like these other things. + +But if you have that very strong prior belief already established that they're a good person, you might ignore + +future observations about them, and you might not update your prior beliefs any sooner. + +And I think that's what that halo effect is, right? + +We receive some minimal amount of information about someone. + +They seem really good. + +It might even skew our judgment and decision-making towards them where we just assume they're good all the time. + +What's really a mathematical way of saying that is that we've + +We initially had a likelihood that let us establish a prior belief about this person we just met. + +And then after that, our precision on that prior belief is so strong that we're no longer able to update our beliefs about that person going forward quite as much. + +And then with the Gollum effect, that's especially interesting one, because, yeah, if you if you witness someone make a mistake on the job once or twice when they're new at it, if you quickly assume if you update your prior belief and say, oh, they must not be good at this job. + +And then you change your own decision making towards them where you repeatedly, I don't know, put them in positions to, you know, to you say like, oh, they're always, you know, they're not very good at this job. + +They're probably never going to be good at this job, which is a rather extreme thought to have. + +But that kind of thing can start to form. + +You never let them do things that would help them to improve at the job. + +You give them other kinds of tasks instead. + +You go and tell other people behind closed doors, oh, we shouldn't have, this person's not very good at those particular tasks, so don't have them do that, right? + +You're starting to make decisions that are realizing that belief you have, that they're not so good at the job, right? + +And so they never get a chance to improve. + +They don't get to keep working on those tasks. + +You don't talk to them as if sharing further knowledge or information with them is actually going to be helpful because you just assume that they're going to be bad at it in the first place. + +These are all the sorts of things that come out of this sort of like what happens when we have prior beliefs about someone that gets shorn from actually taking in new observations and updating our beliefs about them through a likelihood. + +Yeah, they're all super interesting dynamics. + +It's worth, I think, + +Frazier shared a book by Jacob Howey. + +Yeah, really good reading that's a little less on the equation-heavy side and much more on the thinking through these things in a more social way or thinking about humans and how they operate. + +So yeah, it's all just really fascinating. + +I'm happy to + +talk to anyone about these sorts of things as well going forward. + +Another really good one is The Experience Machine by Andy Clark. + +Andy Clark's very big in this sort of space as well. + +So yeah, it's good stuff. + +And then for those who are more interested in social phenomena, like how do we apply these things to like where it's getting closer and closer. + +And then from my standpoint, some of the work I do, I think of like, well, how do people interact with others in these sorts of ways? + +And how does that reflect back on one's own emotional inference? + +How does that relate to one's own understanding of the + +the world and how does that relate to things like trust? + +How does it allow for trusting in one's own self and their own decision making in relation to others? + +And then, you know, that's where that happens. + + +SPEAKER_02: +Where interoception comes in, right? + +Where perhaps interoception, because there's this whole kind of story about, you know, we're predicting the environment, but also we're trying to predict ourselves in some way. + +So, you know, what am I doing right now? + +You know, I'm trivially, where's my heartbeat? + +Things like this. + +So, you know, + + +SPEAKER_04: +Yeah, absolutely. + +Depending on where we draw the line around that, that sort of Markov blanket. + +I mean, if you think of it as like, oh, not, you know, think of an infant who's like just born. + +Like, not only do they have to navigate what we typically think of as an environment where it's like, OK, we have the ground around them. + +We have the room around them. + +We have other people around them. + +But we also have our own bodies. + +We have to learn how to use our bodies. + +There's no assumption here that a child understands every single potential joint position they can make with their arm. + +There's a lot of experimentation, randomly moving and throwing things and all the rest. + +So learning about one's body and reading one's own body, which relates to notions of interoception and proprioception, proprioception being more on the side of motor control. + +And I've had a nice conversation some time ago with Carrie about neuroception and perception of things like pain. + +So, I mean, all of those things can be viewed as part of a broader set of inference and learning processes. + +Yeah, so what happens whenever you interact with someone, but you are interacting + +currently experiencing pain or you're experiencing other kinds of things in your body that make it, you know, interacting with others, you know, there are things going on within you that are distracting or have a higher precision on them to where your brain is attending to that more that can range between someone with, you know, some kind of previous trauma or + +One of my specialties is sort of post-traumatic stress disorder, especially post-traumatic stress disorder in the sense of experiencing a traumatic event. + +So this isn't CPTSD, but later in life, like you have one way of viewing the world and then you have some kind of intense event that strongly changes your view of the world after. + +Like how does your generative model of the world, how you operate in it, + +How does that change before and after that intense event and what kind of things end up getting modulated and how do we view that in terms of how precisions get modulated? + +and all the rest, and how does that relate to things like neurotransmitters, like noradrenaline? + +So it starts having implications for both pharmacological treatment, but then also with the things we're talking about and interacting with others for mental health treatment in the more general sense of therapy. + +Like, how should we talk to someone? + +Should we talk to someone who's been through an intensive trauma and then after that say, oh, no, you're fine? + + +SPEAKER_02: +everything's going to be safe you just had your mind changed from this like it's much more complex than that right like people being able to exercise there you go i'm gonna say the whole question is very much about okay well you know is the is the problem at the level of the setting of parameters and therefore we just need to tweak some of that and then the inference is going to follow right are the parameters fine and the info you know so there's this issue of the hierarchy + +that we need to be mindful of when we're modeling inactive inference. + +We can't just say, well, I got a model that's doing inference, and that's it. + +There's this whole issue about how you set up the whole hierarchy from learning to inference, or even model construction to learning to inference. + +That's the stack that we need to be appreciative of when we're doing the fundamentals. + +i just wanted to we are we are basically at time um for those who wanted to ask questions but didn't uh get the chance please do put them in the coda andrew and myself and daniel are going to be very attentive to that i'm hopefully going to put together uh well danger daniel and myself are going to be more active with respect to the code implementations and some of the animations and things like that so next week i believe we'll come back and we'll do another round of review hopefully maybe with some + +specific reference to particular examples, maybe in the code, where we can really sort of get our hands dirty, maybe just show how some of these things work and things like that. + +So that's the plan going forward. + +If you guys want to change it, it's up to us. + +We can do what we want. + +But thank you very much for everyone here. + +So I'll stop the YouTube recording now. + +All right. + +Goodbye, YouTube people. diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.json new file mode 100644 index 000000000..20c0c0fb1 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.json @@ -0,0 +1 @@ +[{"video_id": "V46B4Xy1PeQ", "segments": [{"start": 3.203, "end": 20.267, "text": " all right welcome everyone we're in on seven seven for what it's worth part one review and code later this week will be another part one review session with fraser then we will be heading into part two of the book", "speaker": "SPEAKER_02"}, {"start": 20.416, "end": 26.905, "text": " So today we'll do sort of some mixture of two things, depending on what people are curious about and want to see.", "speaker": "SPEAKER_02"}, {"start": 26.965, "end": 35.837, "text": "I'm going to start with a brief, just agent July, 2026 overview free mode.", "speaker": "SPEAKER_02"}, {"start": 36.899, "end": 41.785, "text": "And then we can talk more about the idea and the flow of the whole book.", "speaker": "SPEAKER_02"}, {"start": 41.805, "end": 44.009, "text": "Look more at the code if we want, or go to the questions.", "speaker": "SPEAKER_02"}, {"start": 44.75, "end": 48.735, "text": "So I just wanted to start with cloning the fundamentals repo.", "speaker": "SPEAKER_02"}, {"start": 49.424, "end": 57.732, "text": " So Active Inference Institute, fundamentals, GitHub repo, and clone and open that folder, and I'm using cursor.", "speaker": "SPEAKER_02"}, {"start": 58.773, "end": 64.419, "text": "Then I'm gonna hit Hermes update in the terminal.", "speaker": "SPEAKER_02"}, {"start": 64.439, "end": 78.233, "text": "And while that's, and there are commits every few minutes, so it's a good practice, but while it's updating, you can get the Hermes agents open source harness for LLMs, or use any other,", "speaker": "SPEAKER_02"}, {"start": 79.158, "end": 82.381, "text": " LLM harness if you already have one that you like.", "speaker": "SPEAKER_02"}, {"start": 84.123, "end": 95.595, "text": "Right now, the model called HY3 is free, but it's just an example of free, quite good modern models that are constantly on offer from OpenRouter.", "speaker": "SPEAKER_02"}, {"start": 96.376, "end": 100.36, "text": "Just go to models, prompt pricing, set it to free.", "speaker": "SPEAKER_02"}, {"start": 100.98, "end": 102.121, "text": "These are all free models.", "speaker": "SPEAKER_02"}, {"start": 102.182, "end": 106.566, "text": "Sometimes they have some limits for how many things you can do.", "speaker": "SPEAKER_02"}, {"start": 107.339, "end": 109.241, "text": " OK, Hermes is installed.", "speaker": "SPEAKER_02"}, {"start": 110.162, "end": 114.926, "text": "So in Fundamentals repo, there's two ways to interact with it.", "speaker": "SPEAKER_02"}, {"start": 115.947, "end": 118.129, "text": "And there's a web interface.", "speaker": "SPEAKER_02"}, {"start": 118.93, "end": 123.374, "text": "And then there's interacting with it more like through an agent in the terminal that we'll look at or through Cursor.", "speaker": "SPEAKER_02"}, {"start": 125.076, "end": 136.907, "text": "The web interface pops up this list of chapter by chapter animations and visualizations.", "speaker": "SPEAKER_02"}, {"start": 137.073, "end": 141.159, "text": " and a bunch of other visualizations and animations for different topics.", "speaker": "SPEAKER_02"}, {"start": 141.679, "end": 145.785, "text": "And we can keep on building out more topics that people see relevant for the book.", "speaker": "SPEAKER_02"}, {"start": 145.845, "end": 151.774, "text": "Or you can say, prompt, make an interactive simulation for deep generative models.", "speaker": "SPEAKER_02"}, {"start": 152.655, "end": 154.437, "text": "And we'll have the chapters up here.", "speaker": "SPEAKER_02"}, {"start": 154.517, "end": 157.161, "text": "And then the extras, we can have a ton of different pages.", "speaker": "SPEAKER_02"}, {"start": 157.982, "end": 158.883, "text": "So that's cool.", "speaker": "SPEAKER_02"}, {"start": 159.144, "end": 160.846, "text": "There's animations you can look through.", "speaker": "SPEAKER_02"}, {"start": 160.866, "end": 163.57, "text": "We can develop out that web interface.", "speaker": "SPEAKER_02"}, {"start": 164.275, "end": 173.949, "text": " But on the developer side, and what's really fun to explore with agentic coding is using some kind of harness and an LLM to interact with the code base.", "speaker": "SPEAKER_02"}, {"start": 175.151, "end": 182.621, "text": "But just to run the code base from the root layer, just do dot slash run dot sh.", "speaker": "SPEAKER_02"}, {"start": 183.883, "end": 189.792, "text": "And then you can choose to run the simulations for a given chapter or launch that web interface.", "speaker": "SPEAKER_02"}, {"start": 191.194, "end": 193.557, "text": "And then when you,", "speaker": "SPEAKER_02"}, {"start": 194.06, "end": 219.831, "text": " run it you get logs of the website ongoing but with hermes we have that open source harness and a free llm model so yeah this is something that doesn't cost money and say explain where and how variational free energy", "speaker": "SPEAKER_02"}, {"start": 220.081, "end": 249.885, "text": " is calculated in the code base i'm just going to ask it both to cursor in this kind of chat window format and hermes in the terminal format which is largely pretty similar um and anyone want to add any comments or or questions or anything", "speaker": "SPEAKER_02"}, {"start": 250.54, "end": 255.43, "text": " or just share from their experience of working with different coding methods with Active Inference?", "speaker": "SPEAKER_02"}, {"start": 258.857, "end": 262.525, "text": "Maybe just briefly, I mean, I've used Py,", "speaker": "SPEAKER_03"}, {"start": 263.922, "end": 272.111, "text": " So for exploratory work, it's very good if you want to quickly get the lay of the land of a repo.", "speaker": "SPEAKER_03"}, {"start": 272.171, "end": 276.596, "text": "But these tools here are not necessary for you to explore the repo itself, I take it.", "speaker": "SPEAKER_03"}, {"start": 277.637, "end": 282.903, "text": "You can just clone the repo and have a look through the code if you wanted, or go through the web interface.", "speaker": "SPEAKER_03"}, {"start": 283.464, "end": 284.185, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 284.205, "end": 288.149, "text": "Like the docs themselves are Markdown files.", "speaker": "SPEAKER_02"}, {"start": 288.169, "end": 293.355, "text": "So you could just use this as a knowledge base itself, or you could look at it just on the GitHub.", "speaker": "SPEAKER_02"}, {"start": 294.905, "end": 312.762, "text": " Or you could just download it and run the local interface with all of the pre computed outputs, or this is like the next step into like we can say use those methods to implement X Andrew.", "speaker": "SPEAKER_02"}, {"start": 315.256, "end": 315.897, "text": " Yeah, thanks.", "speaker": "SPEAKER_01"}, {"start": 315.917, "end": 338.563, "text": "I would just say for any participants who are wanting to learn something more specific or they would really like to see a particular kind of demo, definitely feel free to either add a question into the discourse on the CODA in the same way that we had other questions or ask them here in the chat.", "speaker": "SPEAKER_01"}, {"start": 338.948, "end": 354.747, "text": " I mean, from having talked to others at the Institute who are still new to coding itself and coming to this from a direction where they have no idea what cursor is and like a dozen other things that just happened here, I would completely understand if someone wanted something more", "speaker": "SPEAKER_01"}, {"start": 354.727, "end": 360.416, "text": " Focused and kind of like centered on, you know, just being able to maybe just run something in your browser.", "speaker": "SPEAKER_01"}, {"start": 361.618, "end": 367.066, "text": "So, yeah, please request those for anyone who more wants something in line with with that.", "speaker": "SPEAKER_01"}, {"start": 367.547, "end": 368.027, "text": "Otherwise.", "speaker": "SPEAKER_01"}, {"start": 368.087, "end": 378.383, "text": "Yeah, this is a very interesting code repo that Daniel is building out.", "speaker": "SPEAKER_01"}, {"start": 378.403, "end": 381.628, "text": "Okay, so.", "speaker": "SPEAKER_02"}, {"start": 381.862, "end": 394.766, "text": " That was starting with the question like format and depending on like which model you're using and various other factors, the latency to the first part of the response to the completed part of the response can be quite long.", "speaker": "SPEAKER_02"}, {"start": 395.287, "end": 404.123, "text": "Like different models give you different visibility into their so-called chain of thought, like their thinking steps, so-called for now.", "speaker": "SPEAKER_02"}, {"start": 404.727, "end": 413.186, "text": " And also their tool use, which over the last one plus years has been one of the major domains of tool and skill improvement.", "speaker": "SPEAKER_02"}, {"start": 413.507, "end": 422.247, "text": "So we can see it's not just doing next token completion in the sentence, it's doing tool search like grep.", "speaker": "SPEAKER_02"}, {"start": 422.378, "end": 446.158, "text": " then it's listing different terminal outputs so in a way it is inputting tokens of text into the terminal it's just that those are on a computer so it ends up doing things on the computer that return text that are valuable for informing its next steps so this is just happening at different speed but both the hy3 model through the hermes harness in the terminal", "speaker": "SPEAKER_02"}, {"start": 447.404, "end": 456.152, "text": " and the cursor side chat style, which can also be dispatched from the terminal, have given like different types of responses.", "speaker": "SPEAKER_02"}, {"start": 456.953, "end": 458.934, "text": "This one was kind of plain text in, plain text out.", "speaker": "SPEAKER_02"}, {"start": 460.456, "end": 467.462, "text": "In the cursor, we get some more multimedia, like visualizations like this.", "speaker": "SPEAKER_02"}, {"start": 468.583, "end": 476.23, "text": "So play around, find out what representations are useful and harness tool skill improvement", "speaker": "SPEAKER_02"}, {"start": 477.475, "end": 483.543, "text": " is one area where there's a ton of development, so it could look different for you in a different setup.", "speaker": "SPEAKER_02"}, {"start": 484.424, "end": 489.831, "text": "So that was more of an investigation question, like where and how is free energy calculated here?", "speaker": "SPEAKER_02"}, {"start": 490.472, "end": 501.946, "text": "And you could continue kind of associatively from there, like explain that like to this type of mood or person about this setting, or how would that apply to this?", "speaker": "SPEAKER_02"}, {"start": 502.747, "end": 506.352, "text": "Or what would be the steps to model this kind of situation?", "speaker": "SPEAKER_02"}, {"start": 506.45, "end": 530.145, "text": " um if it were talking to an agent that has like a web search tool like a perplexity tool or just yeah web search essentially you could ask what research has been done that relates to this area and it could kind of reference the code base in um by looking what locally was on the computer", "speaker": "SPEAKER_02"}, {"start": 531.104, "end": 542.558, "text": " Like this left panel, which might be subliminally obvious or not, depending on how much familiarity with this format, this is just the files in the finder.", "speaker": "SPEAKER_02"}, {"start": 544.3, "end": 547.945, "text": "So this is just a horizontal way to look at this.", "speaker": "SPEAKER_02"}, {"start": 550.027, "end": 552.37, "text": "So that's also related to Git.", "speaker": "SPEAKER_02"}, {"start": 552.518, "end": 558.564, "text": " which getting an intuition for can be helpful, but again, for another situation focusing on this repo.", "speaker": "SPEAKER_02"}, {"start": 559.305, "end": 568.315, "text": "So we've asked these two different instances of different harnesses, different interface, different underlying language model, questions about the code base.", "speaker": "SPEAKER_02"}, {"start": 568.975, "end": 571.718, "text": "They go about it in some different ways that are pretty interesting.", "speaker": "SPEAKER_02"}, {"start": 573.2, "end": 578.205, "text": "Now let's go into building something from the components that are here.", "speaker": "SPEAKER_02"}, {"start": 578.759, "end": 585.091, "text": " So we could look at a pre-existing example, or I've kind of started this prompt.", "speaker": "SPEAKER_02"}, {"start": 585.111, "end": 586.092, "text": "Yeah, many good comments.", "speaker": "SPEAKER_02"}, {"start": 586.113, "end": 588.337, "text": "I'm not even reading them all in the chat.", "speaker": "SPEAKER_02"}, {"start": 588.357, "end": 589.038, "text": "We'll check at the end.", "speaker": "SPEAKER_02"}, {"start": 590.14, "end": 594.809, "text": "I'm giving a prompt here to the cursor agents.", "speaker": "SPEAKER_02"}, {"start": 595.75, "end": 597.814, "text": "Let's add a top-level folder called demo.", "speaker": "SPEAKER_02"}, {"start": 598.823, "end": 612.591, "text": " Putting it all in one folder is just going to keep it clean for us versioning this or if I decide to push it or if you want to keep your experiment or something contained like in a sidecar, which is like a standalone folder inside of a repo or to the side of a repo.", "speaker": "SPEAKER_02"}, {"start": 613.413, "end": 618.283, "text": "So let's add a top level folder called demo with subfolders for different demo examples.", "speaker": "SPEAKER_02"}, {"start": 618.787, "end": 628.496, "text": " And there's matters of preference and taste in terms of systems engineering and interface engineering for software repos.", "speaker": "SPEAKER_02"}, {"start": 628.777, "end": 631.079, "text": "So I'm just walking one path through.", "speaker": "SPEAKER_02"}, {"start": 631.119, "end": 639.106, "text": "There's many ways even something of this relatively straightforward kind of software package could be done.", "speaker": "SPEAKER_02"}, {"start": 639.307, "end": 642.81, "text": "So I'm just laying out one way to do this, but test and explore other ways.", "speaker": "SPEAKER_02"}, {"start": 644.051, "end": 648.135, "text": "There's going to be this top level demo folder, and then there'll be different demos inside of it.", "speaker": "SPEAKER_02"}, {"start": 648.233, "end": 649.955, "text": " Like situation one, two, three.", "speaker": "SPEAKER_02"}, {"start": 650.636, "end": 657.083, "text": "So maybe in the chat or verbally, like let's pick two or three different examples.", "speaker": "SPEAKER_02"}, {"start": 658.164, "end": 661.007, "text": "Just write them in the chat now.", "speaker": "SPEAKER_02"}, {"start": 661.488, "end": 663.75, "text": "Sorry, I like example 2.3.", "speaker": "SPEAKER_04"}, {"start": 663.77, "end": 665.913, "text": "I think that's a really fundamental example.", "speaker": "SPEAKER_04"}, {"start": 666.874, "end": 667.354, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 667.374, "end": 670.338, "text": "I think we have the book examples.", "speaker": "SPEAKER_02"}, {"start": 670.918, "end": 673.041, "text": "I'm just going to prompt it to do some other situational example.", "speaker": "SPEAKER_02"}, {"start": 673.061, "end": 674.983, "text": "But yes, let's look at 2.3.", "speaker": "SPEAKER_02"}, {"start": 676.533, "end": 686.688, "text": " So the demo folder, and then like, you know, baseball game, person walking, thermometer with humidity sensor, like whatever situations.", "speaker": "SPEAKER_02"}, {"start": 687.089, "end": 689.773, "text": "And those will be each in their own folder under demo.", "speaker": "SPEAKER_02"}, {"start": 690.414, "end": 694.56, "text": "And then I'm saying, start with using the core methods of source to make it.", "speaker": "SPEAKER_02"}, {"start": 695.641, "end": 704.194, "text": "SRC is kind of a shorthand for where the source code methods are that get modularly tested by the tests repo or test folder.", "speaker": "SPEAKER_02"}, {"start": 704.512, "end": 711.281, "text": " So that's kind of like the atomic or nuclear components, components.", "speaker": "SPEAKER_02"}, {"start": 711.301, "end": 722.476, "text": "Then scripts or orchestrators call and configure sets of methods, but they don't necessarily define new or complex logic by themself.", "speaker": "SPEAKER_02"}, {"start": 723.297, "end": 732.409, "text": "So that's just one separation of consideration to have the underlying methods in source or something similarly named, tested,", "speaker": "SPEAKER_02"}, {"start": 732.659, "end": 752.254, "text": " at the unit test level with tests and then get called and invoked by thin orchestrators or scripts, which make it so that you can improve and version the underlying methods and kind of keep a level of abstraction in terms of like, we want to be able to call, um,", "speaker": "SPEAKER_02"}, {"start": 752.555, "end": 780.722, "text": " example 3.3 with our most updated versions of the configuration and the implementation and the logging and the visualizing whereas if chapter 3 was just scripts for just chapter 3 then we could get divergence with the same method from different chapters um but using demo we can just kind of make some other random things so just type type a random situation or something for uh", "speaker": "SPEAKER_02"}, {"start": 781.799, "end": 782.801, "text": " that we'll make them.", "speaker": "SPEAKER_02"}, {"start": 783.402, "end": 788.59, "text": "And then while it's making those demos, we'll look at one of the examples that Frazier mentioned.", "speaker": "SPEAKER_02"}, {"start": 793.418, "end": 795.341, "text": "I'll put something in the chat myself for that.", "speaker": "SPEAKER_04"}, {"start": 795.802, "end": 796.042, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 798.186, "end": 799.548, "text": "How about isocades?", "speaker": "SPEAKER_02"}, {"start": 803.735, "end": 804.516, "text": "Riding a bicycle.", "speaker": "SPEAKER_02"}, {"start": 814.52, "end": 815.981, "text": " Someone add one or type one.", "speaker": "SPEAKER_02"}, {"start": 820.086, "end": 822.588, "text": "Simulating active inference processes by message passing.", "speaker": "SPEAKER_04"}, {"start": 822.608, "end": 824.75, "text": "They have a cool Bayesian thermostat agent.", "speaker": "SPEAKER_04"}, {"start": 825.271, "end": 829.515, "text": "I'll paste that in the chat.", "speaker": "SPEAKER_04"}, {"start": 829.535, "end": 829.815, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 829.855, "end": 830.336, "text": "Drone flight.", "speaker": "SPEAKER_02"}, {"start": 834.22, "end": 839.785, "text": "Could take learnable loop, paste the blog in, say transpose this to this framework.", "speaker": "SPEAKER_02"}, {"start": 844.338, "end": 844.678, "text": " Okay.", "speaker": "SPEAKER_02"}, {"start": 845.379, "end": 845.6, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 846.401, "end": 849.004, "text": "I'm going to switch from agent mode just to show one mode.", "speaker": "SPEAKER_02"}, {"start": 849.024, "end": 865.384, "text": "Different harnesses, again, have different features, but I'm going to switch it to plan mode, which is that it's going to set off to make a work plan for how to do this request, which, as far as I understand,", "speaker": "SPEAKER_02"}, {"start": 866.663, "end": 868.406, "text": " is usually as good or better.", "speaker": "SPEAKER_02"}, {"start": 869.047, "end": 878.34, "text": "There's rarely a situation where simply launching into it immediately seems like it'd be better, except unless it's going to cost extra tokens for something that's irrelevant.", "speaker": "SPEAKER_02"}, {"start": 878.36, "end": 881.145, "text": "But it's kind of fun to see, and then we'll have the opportunity to look at the plan.", "speaker": "SPEAKER_02"}, {"start": 881.986, "end": 890.899, "text": "But things like plan, loop, goal, memory layers, these are having a lot of impact in long horizon task planning.", "speaker": "SPEAKER_02"}, {"start": 891.183, "end": 899.952, "text": " and more agentic orchestration methods because it's not like it's a rolling window of memory and it forgets where it was.", "speaker": "SPEAKER_02"}, {"start": 900.553, "end": 916.23, "text": "If it knows where the work plan is, then a hierarchically organized work plan, even that has hundreds or tens of thousands of task items with the right kind of harness and listener environment, you can just continue to go for.", "speaker": "SPEAKER_02"}, {"start": 917.651, "end": 921.055, "text": "Okay, so while it's making its plan, let's look at", "speaker": "SPEAKER_02"}, {"start": 923.11, "end": 923.891, "text": " Example 2.3.", "speaker": "SPEAKER_02"}, {"start": 931.944, "end": 932.765, "text": "Yes, go ahead, Frasier.", "speaker": "SPEAKER_02"}, {"start": 933.727, "end": 935.79, "text": "Introduce it or what's important about it.", "speaker": "SPEAKER_02"}, {"start": 936.712, "end": 945.726, "text": "Well, actually, I just want to mention, I just cloned the repo on my local machine.", "speaker": "SPEAKER_03"}, {"start": 946.043, "end": 952.812, "text": " I used UV to install everything and I ran, I ran chapter, you know, 2.3, the precision example.", "speaker": "SPEAKER_03"}, {"start": 953.793, "end": 957.318, "text": "I'm just wondering where the outputs would have gone.", "speaker": "SPEAKER_03"}, {"start": 957.338, "end": 959.721, "text": "I didn't run it in interactive mode.", "speaker": "SPEAKER_03"}, {"start": 959.741, "end": 962.365, "text": "So I just kind of ran it through the, through the terminal.", "speaker": "SPEAKER_03"}, {"start": 963.126, "end": 963.306, "text": "Yep.", "speaker": "SPEAKER_02"}, {"start": 965.229, "end": 965.389, "text": "Yep.", "speaker": "SPEAKER_02"}, {"start": 965.81, "end": 965.97, "text": "Yep.", "speaker": "SPEAKER_02"}, {"start": 966.05, "end": 967.412, "text": "Here's, here's what that looks like.", "speaker": "SPEAKER_02"}, {"start": 967.792, "end": 971.617, "text": "So I'm in the top level of fundamentals dot slash run.", "speaker": "SPEAKER_02"}, {"start": 973.9, "end": 975.122, "text": "I'm going to,", "speaker": "SPEAKER_02"}, {"start": 975.271, "end": 984.641, "text": " hit option A, run, yeah, option A or X. I'll run the chapters.", "speaker": "SPEAKER_02"}, {"start": 986.242, "end": 989.846, "text": "And then it's gonna log where it saves things.", "speaker": "SPEAKER_02"}, {"start": 990.907, "end": 993.089, "text": "So output slash figures slash chapter.", "speaker": "SPEAKER_02"}, {"start": 993.209, "end": 995.592, "text": "So output slash figures slash chapters.", "speaker": "SPEAKER_02"}, {"start": 997.173, "end": 1002.679, "text": "And that's just this folder, like example 2.3", "speaker": "SPEAKER_02"}, {"start": 1003.975, "end": 1005.357, "text": " Here's just that image.", "speaker": "SPEAKER_02"}, {"start": 1008.04, "end": 1014.227, "text": "Yeah, Sebastian wrote, if you don't have video RAM for your own LLM, I can advise to try Chinese models like Minimax.", "speaker": "SPEAKER_02"}, {"start": 1014.788, "end": 1019.874, "text": "Yeah, again, on OpenRouter, there's always, or News Portal, there's always free models.", "speaker": "SPEAKER_02"}, {"start": 1020.615, "end": 1021.536, "text": "Those are getting better and better.", "speaker": "SPEAKER_02"}, {"start": 1021.576, "end": 1027.603, "text": "And, or you can run", "speaker": "SPEAKER_02"}, {"start": 1028.647, "end": 1034.496, "text": " local LLMs or go to a friend's server, but like Ollama or other methods.", "speaker": "SPEAKER_02"}, {"start": 1036.519, "end": 1038.783, "text": "Giacomo wrote, does it log to a timestamped file?", "speaker": "SPEAKER_02"}, {"start": 1040.406, "end": 1045.333, "text": "Definitely in the terminal, you'll get a timestamp and in the metadata of the file, you'll get a timestamp.", "speaker": "SPEAKER_02"}, {"start": 1048.739, "end": 1050.161, "text": "Okay, so here we go.", "speaker": "SPEAKER_02"}, {"start": 1050.181, "end": 1054.047, "text": "Back to this cursor chat.", "speaker": "SPEAKER_02"}, {"start": 1054.263, "end": 1056.045, "text": " So we asked it to go to plant mode.", "speaker": "SPEAKER_02"}, {"start": 1056.065, "end": 1064.294, "text": "And then this is another area of both utility and frustration for the human side.", "speaker": "SPEAKER_02"}, {"start": 1064.314, "end": 1065.736, "text": "It's like, it's asking questions.", "speaker": "SPEAKER_02"}, {"start": 1065.796, "end": 1073.665, "text": "So again, at that sort of qualitative level, putting aside all that we're learning about energy based methods and all that, this is like an active inference.", "speaker": "SPEAKER_02"}, {"start": 1073.785, "end": 1079.511, "text": "Like it identified somehow that it had enough of an uncertainty to cross the threshold", "speaker": "SPEAKER_02"}, {"start": 1079.609, "end": 1083.134, "text": " to break a loop and go into asking us a question.", "speaker": "SPEAKER_02"}, {"start": 1083.935, "end": 1093.029, "text": "So, should the new demo folder be first class, like chapters and extra, menu, web, smoke, orchestrator, save, or standalone scripts only?", "speaker": "SPEAKER_02"}, {"start": 1093.99, "end": 1094.691, "text": "Full integration.", "speaker": "SPEAKER_02"}, {"start": 1095.532, "end": 1108.992, "text": "This is kind of like one reason why having a modular, signposted, composable software package, even from the beginning, is super effective, because it's like, now we can wire a whole new", "speaker": "SPEAKER_02"}, {"start": 1110.271, "end": 1116.539, "text": " format or a whole new media type because there's already multiple things ongoing.", "speaker": "SPEAKER_02"}, {"start": 1117.58, "end": 1127.232, "text": "And yes, that can be like in overkill or overly restraining mode of software.", "speaker": "SPEAKER_02"}, {"start": 1128.193, "end": 1136.403, "text": "However, finding that balance of like structural plasticity and form, that's kind of the art and the science.", "speaker": "SPEAKER_02"}, {"start": 1137.765, "end": 1139.367, "text": "All right, so here's what the plan looks like.", "speaker": "SPEAKER_02"}, {"start": 1148.24, "end": 1150.163, "text": " There's the plan and the to-dos.", "speaker": "SPEAKER_02"}, {"start": 1154.429, "end": 1161.358, "text": "At this point, we could give feedback or we can just hit build.", "speaker": "SPEAKER_02"}, {"start": 1165.104, "end": 1175.258, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 1175.278, "end": 1175.378, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 1180.184, "end": 1183.549, "text": " So that was all this, that was just the scripts running from that run.", "speaker": "SPEAKER_02"}, {"start": 1183.949, "end": 1188.857, "text": "So it logs what was output, right?", "speaker": "SPEAKER_02"}, {"start": 1188.877, "end": 1191.961, "text": "Click kill terminal, right?", "speaker": "SPEAKER_02"}, {"start": 1191.981, "end": 1192.702, "text": "Click kill term.", "speaker": "SPEAKER_02"}, {"start": 1193.624, "end": 1197.349, "text": "That's the, that's the terminal that's running the website.", "speaker": "SPEAKER_02"}, {"start": 1200.033, "end": 1203.979, "text": "This is that high three model.", "speaker": "SPEAKER_02"}, {"start": 1205.501, "end": 1208.385, "text": "So can you review the repo?", "speaker": "SPEAKER_02"}, {"start": 1209.597, "end": 1227.258, "text": " and give me a learning path given I have taken frequentist but no Bayesian statistics.", "speaker": "SPEAKER_02"}, {"start": 1237.791, "end": 1239.233, "text": "Okay, then on the right side,", "speaker": "SPEAKER_02"}, {"start": 1240.816, "end": 1243.641, "text": " We can see that plan getting built.", "speaker": "SPEAKER_02"}, {"start": 1249.411, "end": 1251.314, "text": "Because we, yeah, Fraser.", "speaker": "SPEAKER_02"}, {"start": 1252.315, "end": 1266.199, "text": "I was just gonna say, it is nice to be able to have this ability to query your, you know, LLM assistant here to be like, all right, well, you know, I'm working on this, but really how does this relate to this other thing I'm thinking about in much the manner you did just then?", "speaker": "SPEAKER_03"}, {"start": 1266.179, "end": 1293.942, "text": " um that is that's a nice workflow for someone who doesn't know as much about active inference as they would like but maybe already knows some their way around a code repository or something like that so it is it is a powerful way to do things yeah thank you yeah totally agree some of you probably will be very familiar with an ide and use it in a similar way or very familiar with ide not seen it used in this way or have your own", "speaker": "SPEAKER_03"}, {"start": 1294.951, "end": 1301.257, "text": " more ergonomic way for using it, or maybe you haven't used an IDE, but there's a variety of free ones for that too.", "speaker": "SPEAKER_02"}, {"start": 1302.698, "end": 1308.423, "text": "So because in that demo ask, we went for the full integration.", "speaker": "SPEAKER_02"}, {"start": 1308.764, "end": 1315.71, "text": "We could have like, even without much software engineering, you could imagine like, okay, we're gonna have this demos zone.", "speaker": "SPEAKER_02"}, {"start": 1317.031, "end": 1320.014, "text": "Is it gonna be like its own self-contained sandbox?", "speaker": "SPEAKER_02"}, {"start": 1320.194, "end": 1324.418, "text": "Like have all the methods it needs so that it can be picked up and moved outside and still work?", "speaker": "SPEAKER_02"}, {"start": 1325.005, "end": 1341.225, "text": " is it gonna make like light use of some aspects of the package, which would have like kind of demo specific logic in the package or in the demo folder, and then like more core methods from the main package, but then it would require that package.", "speaker": "SPEAKER_02"}, {"start": 1341.627, "end": 1352.402, "text": " or the demo could be fully integrated like the extras where they're only thin orchestrators and all the core methods go to the core software package.", "speaker": "SPEAKER_02"}, {"start": 1353.123, "end": 1360.814, "text": "So that's a good question of how integrated do you want that time cone kind of software blanket to be?", "speaker": "SPEAKER_02"}, {"start": 1362.537, "end": 1370.588, "text": "So, because we want more integration, it is writing these demo methods,", "speaker": "SPEAKER_02"}, {"start": 1371.428, "end": 1392.778, "text": " under the main src so it's most integrated you can still extract some methods but but it's connected with the core testing architecture it's not just like standalone but you could have also just gone to a blank folder and say like from scratch make me a standalone active inference drone example", "speaker": "SPEAKER_02"}, {"start": 1392.758, "end": 1405.81, "text": " Now, that's very possible now, but part of the reason for having reusable software is so that we can validate, benchmark, all these kinds of things and know that we're not just making a method that says one thing and does another.", "speaker": "SPEAKER_02"}, {"start": 1407.051, "end": 1408.052, "text": "So it's going through.", "speaker": "SPEAKER_02"}, {"start": 1408.092, "end": 1410.274, "text": "The green are additions.", "speaker": "SPEAKER_02"}, {"start": 1410.974, "end": 1414.698, "text": "The red are subtractions of numbers of lines.", "speaker": "SPEAKER_02"}, {"start": 1415.638, "end": 1420.723, "text": "So just showing its sequence of tool use and file modification.", "speaker": "SPEAKER_02"}, {"start": 1421.952, "end": 1425.137, "text": " And then it's showing here's 27 files that it's modified.", "speaker": "SPEAKER_02"}, {"start": 1427.28, "end": 1434.491, "text": "We can see those are those three examples and it's three of six done on the plan.", "speaker": "SPEAKER_02"}, {"start": 1438.016, "end": 1438.476, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 1438.496, "end": 1447.45, "text": "Now back over here to this and like Frazier brought up, like that's what is so fun to explore and figure out the right attention allocation, regimes of attention.", "speaker": "SPEAKER_02"}, {"start": 1447.991, "end": 1448.331, "text": "Like,", "speaker": "SPEAKER_02"}, {"start": 1449.037, "end": 1458.372, "text": " Do you want to focus more on just following one conversation like that prompt that we asked about a learning path for Bayesian statistics?", "speaker": "SPEAKER_02"}, {"start": 1458.973, "end": 1464.782, "text": "Or are we going to be like switching between different projects and kind of like another layer of attentional move?", "speaker": "SPEAKER_02"}, {"start": 1466.104, "end": 1468.007, "text": "I'm using the cursor IDE.", "speaker": "SPEAKER_02"}, {"start": 1469.928, "end": 1480.963, "text": " However, even this side chat AI functionality is available through a plugin, even if your IDE is not AI-based.", "speaker": "SPEAKER_02"}, {"start": 1483.025, "end": 1483.987, "text": "It's a dark mode theme.", "speaker": "SPEAKER_02"}, {"start": 1485.589, "end": 1488.713, "text": "But this kind of represents two modes", "speaker": "SPEAKER_02"}, {"start": 1488.929, "end": 1495.039, "text": " two points on that continuum in that cursor window, we're basically doing, we're modifying the core package.", "speaker": "SPEAKER_02"}, {"start": 1495.821, "end": 1500.068, "text": "Like, we're modifying surfaces and functions of the core repo.", "speaker": "SPEAKER_02"}, {"start": 1501.15, "end": 1507.02, "text": "And then in this chat window, we're not even modifying the repo at all.", "speaker": "SPEAKER_02"}, {"start": 1507.801, "end": 1510.286, "text": "We might not even be asking something related to the repo.", "speaker": "SPEAKER_02"}, {"start": 1511.027, "end": 1512.81, "text": "We could just, like...", "speaker": "SPEAKER_02"}, {"start": 1513.82, "end": 1540.401, "text": " spin up a Hermes and ask it some other question and it may or may not even use the repo for for context um and then in the middle is like you're leaving the core logic of the repo intact but you might be doing some side experiment that's an important access is like where are you actually chipping away or adding to the repos code base versus where are you just", "speaker": "SPEAKER_02"}, {"start": 1540.668, "end": 1545.936, "text": " using the repo for context or not to make some other point.", "speaker": "SPEAKER_02"}, {"start": 1550.142, "end": 1552.585, "text": "Okay, so here it ended its response.", "speaker": "SPEAKER_02"}, {"start": 1552.625, "end": 1556.471, "text": "We asked, we know about frequentist stats, but not Bayesian statistics.", "speaker": "SPEAKER_02"}, {"start": 1559.675, "end": 1570.05, "text": "This that just popped up is a terminal window that cursor spins up.", "speaker": "SPEAKER_02"}, {"start": 1570.874, "end": 1576.74, "text": " where it's gonna run the scripts that it wrote in the demo sections.", "speaker": "SPEAKER_02"}, {"start": 1586.451, "end": 1600.847, "text": "How about just report the learning path in a sequence of 50 creative song titles from the 1970s?", "speaker": "SPEAKER_02"}, {"start": 1602.953, "end": 1609.3, "text": " See where that goes.", "speaker": "SPEAKER_02"}, {"start": 1618.81, "end": 1631.003, "text": "Anyone want to add a question or anything else about how they're using coding agents or like where, what parts look familiar or would they add in, or, or it could be useful for somebody who's learning how to use this.", "speaker": "SPEAKER_02"}, {"start": 1641.867, "end": 1649.817, "text": " Certainly for people who maybe don't have much experience, it can look a bit massively confusing, especially for a virtuoso like Daniel.", "speaker": "SPEAKER_03"}, {"start": 1650.358, "end": 1652.12, "text": "But it is actually not.", "speaker": "SPEAKER_03"}, {"start": 1652.741, "end": 1661.172, "text": "It doesn't have to be as sophisticated as this or, dare I say, as confusing.", "speaker": "SPEAKER_03"}, {"start": 1661.232, "end": 1666.879, "text": "But basically, the repo is just placed where we host our examples.", "speaker": "SPEAKER_03"}, {"start": 1667.163, "end": 1671.791, "text": " and whatever other examples, and your workflow, your chosen way of interacting that is totally up to you.", "speaker": "SPEAKER_03"}, {"start": 1671.811, "end": 1675.076, "text": "So this is definitely a powerful modern way to do that.", "speaker": "SPEAKER_03"}, {"start": 1676.739, "end": 1681.907, "text": "But you can do more traditional, you know, I'm just going to clone the repository and have a look manually.", "speaker": "SPEAKER_03"}, {"start": 1682.007, "end": 1682.949, "text": "That's totally up to you.", "speaker": "SPEAKER_03"}, {"start": 1682.989, "end": 1689.9, "text": "So don't feel like this is like Howard saying you have to do things if you want to interact with the repo.", "speaker": "SPEAKER_03"}, {"start": 1691.483, "end": 1691.903, "text": "Thank you.", "speaker": "SPEAKER_02"}, {"start": 1691.943, "end": 1693.626, "text": "Totally agreed.", "speaker": "SPEAKER_02"}, {"start": 1694.264, "end": 1712.849, "text": " This is just one way of working kind of in the engine room, as it were, and help us and collaborate with everyone to co-create the interfaces that would put you in like a flow learning state.", "speaker": "SPEAKER_02"}, {"start": 1714.331, "end": 1720.9, "text": "And maybe you'd get familiar with things that don't even exist today, but in six months, they could be a super effective way to learn.", "speaker": "SPEAKER_02"}, {"start": 1722.442, "end": 1722.702, "text": "Andrew?", "speaker": "SPEAKER_02"}, {"start": 1724.977, "end": 1726.079, "text": " Yeah, thanks.", "speaker": "SPEAKER_01"}, {"start": 1727.24, "end": 1740.08, "text": "Yeah, I, on Fraser's point earlier, I did, you know, I haven't had the time and haven't received enough participant feedback quite yet to know how to build this out further.", "speaker": "SPEAKER_01"}, {"start": 1740.2, "end": 1749.094, "text": "But there is one demo for chapter two that I made some time ago, and I dropped the link to that in the chat.", "speaker": "SPEAKER_01"}, {"start": 1749.395, "end": 1754.881, "text": " And that's a very, very different approach, which is simply you open it up in your web browser.", "speaker": "SPEAKER_01"}, {"start": 1754.942, "end": 1757.024, "text": "There's no installation required.", "speaker": "SPEAKER_01"}, {"start": 1757.084, "end": 1758.826, "text": "There's no prompting LLMs.", "speaker": "SPEAKER_01"}, {"start": 1758.886, "end": 1759.647, "text": "There's none of that.", "speaker": "SPEAKER_01"}, {"start": 1760.428, "end": 1766.695, "text": "And it's just a very direct, like you can just click run all at the top and it runs all the code in your browser for you.", "speaker": "SPEAKER_01"}, {"start": 1766.735, "end": 1771.821, "text": "And then the big thing is that it has these like sliders that you can interact with.", "speaker": "SPEAKER_01"}, {"start": 1772.602, "end": 1777.528, "text": "And so the code that's there currently can just show you like,", "speaker": "SPEAKER_01"}, {"start": 1778.369, "end": 1785.978, "text": " here's what happens whenever you change the mean of the prior versus the mean of the likelihood, things like that, all the things we talk about in chapter two.", "speaker": "SPEAKER_01"}, {"start": 1786.018, "end": 1796.07, "text": "And then it also has maximum likelihood estimation with gradient descent, as well as map with gradient descent.", "speaker": "SPEAKER_01"}, {"start": 1796.55, "end": 1801.356, "text": "And so I made that with the intent of, you know, there are many people who are new here who are", "speaker": "SPEAKER_01"}, {"start": 1801.758, "end": 1826.533, "text": " coming in they're learning coding they're learning things like that they want to know how to get involved in active inference and then very quickly they get swept away by a lot of like difficult you know sort of different things to to understand and so um yeah that was the focus of that demo so if anyone you know finds that to be useful or they want to see like another more directly like interactive without necessarily like having an llm handle a bunch of", "speaker": "SPEAKER_01"}, {"start": 1826.513, "end": 1854.023, "text": " things for you and just bearing in mind that all lms are kind of prone to error so um you know there's just different ways to do things um so yeah just wanted to have that known that that's an affordance thank you magdalena um yeah my question is um i'm sorry that i don't know the answer but um does do you have a link", "speaker": "SPEAKER_01"}, {"start": 1854.492, "end": 1864.468, "text": " To a site in the active inference Institute that walks people who are really, really super new to coding and everything.", "speaker": "SPEAKER_05"}, {"start": 1864.837, "end": 1873.565, "text": " where it describes what are the basic skills that you need in order to engage with these models, right?", "speaker": "SPEAKER_05"}, {"start": 1874.505, "end": 1889.338, "text": "So what I guess I'm asking the experts is, as a beginner, I don't have an idea as to where I can stop in my basic learning and then proceed to use these models.", "speaker": "SPEAKER_05"}, {"start": 1889.959, "end": 1894.843, "text": "So is there something already in the Active Inference Institute that can walk us through that?", "speaker": "SPEAKER_05"}, {"start": 1894.823, "end": 1896.645, "text": " phase of our learning.", "speaker": "SPEAKER_05"}, {"start": 1899.428, "end": 1900.529, "text": "It's an important question.", "speaker": "SPEAKER_02"}, {"start": 1900.729, "end": 1913.022, "text": "I think there's kind of a step zero right before this video picks up with downloading cursor or an IDE and installing Hermes, installing Git and cloning the fundamentals repo.", "speaker": "SPEAKER_02"}, {"start": 1914.143, "end": 1924.594, "text": "So that is some information that somebody could either ask about or search to install the IDE.", "speaker": "SPEAKER_02"}, {"start": 1926.076, "end": 1951.493, "text": " install GitHub okay and install Hermes agent to recapitulate this exact material and then pick up right at the beginning of this meeting um and I think like we can continue to explore what artifacts and and onboarding material it just there are so many onboardings to programming because", "speaker": "SPEAKER_02"}, {"start": 1953.599, "end": 1955.862, "text": " there's an infinite number of places to come to and go to.", "speaker": "SPEAKER_02"}, {"start": 1955.882, "end": 1982.172, "text": "This in terms of like the time that I've been learning these topics, this is like on multiple orders of magnitude, different setup and applicability duration, like in a mind boggling fashion, just getting things like the packages installed", "speaker": "SPEAKER_02"}, {"start": 1983.502, "end": 2005.421, "text": " was like a quite challenging situation, let alone the natural conversation with a code base and the ability to like convert a sort of question like Giacomo's, how much of it could be ported to Jupyter Notebooks?", "speaker": "SPEAKER_02"}, {"start": 2007.263, "end": 2008.884, "text": "It's just like all of it.", "speaker": "SPEAKER_02"}, {"start": 2010.165, "end": 2012.327, "text": "Make that transformation.", "speaker": "SPEAKER_02"}, {"start": 2012.628, "end": 2014.691, "text": " That's it, it's not a huge cognitive lift.", "speaker": "SPEAKER_02"}, {"start": 2015.853, "end": 2018.697, "text": "No harm in trying.", "speaker": "SPEAKER_02"}, {"start": 2020.02, "end": 2031.818, "text": "Building expertise in what you can ask and what is like trivially tiny, it's gonna finish in 15 seconds versus what would require like a structured work plan.", "speaker": "SPEAKER_02"}, {"start": 2032.473, "end": 2035.578, "text": " or an even more advanced way to do task allocation.", "speaker": "SPEAKER_02"}, {"start": 2036.158, "end": 2042.287, "text": "That's the art and the science of understanding how to work with these synthetic intelligence systems.", "speaker": "SPEAKER_02"}, {"start": 2043.068, "end": 2050.379, "text": "But just watching what's happening right here should put a few dots down.", "speaker": "SPEAKER_02"}, {"start": 2050.439, "end": 2055.947, "text": "And then how you want to allocate your regime of attention.", "speaker": "SPEAKER_02"}, {"start": 2056.808, "end": 2060.674, "text": "Do you like one chat", "speaker": "SPEAKER_02"}, {"start": 2061.802, "end": 2068.029, "text": " looking more like a chat window.", "speaker": "SPEAKER_02"}, {"start": 2069.651, "end": 2076.999, "text": "There's a lot of different user interface services for this in terms of the more generative AI end.", "speaker": "SPEAKER_02"}, {"start": 2077.479, "end": 2087.37, "text": "You could paste the link into Perplexity or Google search on your browser with no software and just say, what is this?", "speaker": "SPEAKER_02"}, {"start": 2091.4, "end": 2095.045, "text": " perplexity and just what is this?", "speaker": "SPEAKER_02"}, {"start": 2100.172, "end": 2105.059, "text": "Or you can read the GitHub static documentation or paste it into something.", "speaker": "SPEAKER_02"}, {"start": 2105.079, "end": 2120.26, "text": "And then we're making the notebooks, which gives a good intermediate level because it's browser based, but runs the code on through running it locally.", "speaker": "SPEAKER_02"}, {"start": 2120.915, "end": 2130.588, "text": " which has the most configurability and resilience, and whether your drone demo is ready to be relied upon.", "speaker": "SPEAKER_02"}, {"start": 2130.648, "end": 2142.985, "text": "That's a second question or a fifth question or something, but is it one prompt away to make an Active Inference-esque drone?", "speaker": "SPEAKER_02"}, {"start": 2143.106, "end": 2143.446, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 2145.316, "end": 2156.089, "text": " and like just what was a bigger or a smaller jump and making visible some of these computer scientific steps that otherwise are like pretty down the road.", "speaker": "SPEAKER_02"}, {"start": 2158.191, "end": 2159.233, "text": "It just flips that.", "speaker": "SPEAKER_02"}, {"start": 2159.813, "end": 2160.094, "text": "Fraser?", "speaker": "SPEAKER_02"}, {"start": 2161.055, "end": 2161.395, "text": "Thank you.", "speaker": "SPEAKER_05"}, {"start": 2162.837, "end": 2167.122, "text": "Yeah, I just wanted to say real quick, I guess there isn't yet a good resource", "speaker": "SPEAKER_03"}, {"start": 2169.16, "end": 2184.672, "text": " the Active Infants Institute ecosystem itself yet, which would maybe walk someone through a program, a way to learn the fundamentals in terms of the technological skills and", "speaker": "SPEAKER_03"}, {"start": 2185.63, "end": 2187.272, "text": " out there that you can interface with.", "speaker": "SPEAKER_03"}, {"start": 2187.692, "end": 2190.875, "text": "The problem that you have is that there's so many of them that do exist already.", "speaker": "SPEAKER_03"}, {"start": 2191.696, "end": 2195.66, "text": "So in some sense, you just need to choose one and go with it.", "speaker": "SPEAKER_03"}, {"start": 2196.101, "end": 2197.462, "text": "And it obviously depends on who you are.", "speaker": "SPEAKER_03"}, {"start": 2197.622, "end": 2205.51, "text": "I will just say that the way that I learned Python, I obviously went to university to do computer science and maths.", "speaker": "SPEAKER_03"}, {"start": 2206.091, "end": 2207.733, "text": "There's a lot of programming there.", "speaker": "SPEAKER_03"}, {"start": 2208.513, "end": 2212.077, "text": "But the way that I actually learned Python was through this course here.", "speaker": "SPEAKER_03"}, {"start": 2215.06, "end": 2216.602, "text": " millions of people have done and they really like.", "speaker": "SPEAKER_03"}, {"start": 2216.663, "end": 2219.787, "text": "It's on a platform called Udemy.", "speaker": "SPEAKER_03"}, {"start": 2220.148, "end": 2224.936, "text": "So that would be, if you don't know any Python programming and you want to learn some Python, that's an excellent resource.", "speaker": "SPEAKER_03"}, {"start": 2225.597, "end": 2237.355, "text": "There are, in terms of skills that you would need, really you need probably some basic familiarity with Python and some basic familiarity with .", "speaker": "SPEAKER_03"}, {"start": 2237.774, "end": 2238.755, "text": " You know, that's pretty much it.", "speaker": "SPEAKER_03"}, {"start": 2238.895, "end": 2243.9, "text": "And then everything from there is really up to you in terms of the agential workflow.", "speaker": "SPEAKER_03"}, {"start": 2244.241, "end": 2252.929, "text": "This is really quite cool what Daniel is showing, but it is only one of the interpretation as to how it is that you can interact with this material.", "speaker": "SPEAKER_03"}, {"start": 2253.39, "end": 2258.855, "text": "So you basically Python syntax and like what is GitHub and how do I use it?", "speaker": "SPEAKER_03"}, {"start": 2259.436, "end": 2260.797, "text": "Really, those are the two things you need.", "speaker": "SPEAKER_03"}, {"start": 2261.118, "end": 2263.24, "text": "So there isn't yet a place that I'm aware of.", "speaker": "SPEAKER_03"}, {"start": 2263.26, "end": 2266.363, "text": "Although, Andrew, maybe you've got your hand up if there's anything you'd like to say on that.", "speaker": "SPEAKER_03"}, {"start": 2268.166, "end": 2268.807, "text": " Sure, thanks.", "speaker": "SPEAKER_01"}, {"start": 2269.107, "end": 2280.377, "text": "Yeah, I just wanted to add to sort of the list of affordances there because I, you know, I also went to university for like, you know, economics and stats and various other things that have been helpful.", "speaker": "SPEAKER_01"}, {"start": 2280.477, "end": 2298.173, "text": "But in this context, as far as coding and programming, you know, there's some free resources like W3 is a website that has a lot of common functionalities for programming languages like Python and R and probably Julia by now.", "speaker": "SPEAKER_01"}, {"start": 2298.153, "end": 2304.113, "text": " And then Datacamp was very helpful for those who need kind of a gamified approach.", "speaker": "SPEAKER_01"}, {"start": 2304.756, "end": 2308.99, "text": "And then there's Coursera, which is very helpful.", "speaker": "SPEAKER_01"}, {"start": 2309.257, "end": 2316.608, "text": " that you can kind of earn certifications from university designed online courses.", "speaker": "SPEAKER_01"}, {"start": 2317.349, "end": 2318.852, "text": "So some find those useful.", "speaker": "SPEAKER_01"}, {"start": 2318.912, "end": 2324.32, "text": "It leads to something you can add sort of to your CV while along the way, like doing hands-on projects.", "speaker": "SPEAKER_01"}, {"start": 2324.661, "end": 2329.408, "text": "But then from there, the big thing is like, can you actually directly interact with the", "speaker": "SPEAKER_01"}, {"start": 2329.388, "end": 2334.535, "text": " The inputs intermediate outputs and final outputs of the code that you're creating.", "speaker": "SPEAKER_01"}, {"start": 2335.095, "end": 2340.903, "text": "So, for those who don't know what a notebook is, a notebook is something that allows you to sort of run code.", "speaker": "SPEAKER_01"}, {"start": 2341.644, "end": 2350.755, "text": "Chunk by chunk sort of or cell by cell and that's a really nice way to both demonstrate how code works as well as, like, incrementally on your own build up.", "speaker": "SPEAKER_01"}, {"start": 2350.735, "end": 2356.662, "text": " Sometimes a lot of motivation comes from choosing some kind of project that is interesting to you.", "speaker": "SPEAKER_01"}, {"start": 2356.702, "end": 2364.01, "text": "And then from there, just sort of building up from there, including learning about all the trials and errors that come along the way.", "speaker": "SPEAKER_01"}, {"start": 2364.07, "end": 2370.838, "text": "So, because it's just the thing with LLMs is that there's a lot of, you know, the catch word of the day is vibe coding.", "speaker": "SPEAKER_01"}, {"start": 2370.878, "end": 2372.76, "text": "And it was interesting.", "speaker": "SPEAKER_01"}, {"start": 2372.74, "end": 2397.617, "text": " interesting point made in the chat that yeah the while a lot of these things look look interesting i just you know it's sometimes one can miss the forest for the trees um whenever they they do this kind of um llm based coding but yeah whatever whatever suits your style yep yep so first on the notebook question it was like we just saw it happening", "speaker": "SPEAKER_01"}, {"start": 2398.576, "end": 2403.702, "text": " from the question to this, it landed the plane on all the chapter notebooks.", "speaker": "SPEAKER_02"}, {"start": 2404.643, "end": 2415.294, "text": "Now, maybe some of the scripts need to be slightly updated in order to run in the notebook format.", "speaker": "SPEAKER_02"}, {"start": 2415.314, "end": 2420.4, "text": "But like, it is running the notebooks using the methods from the package.", "speaker": "SPEAKER_02"}, {"start": 2421.161, "end": 2427.047, "text": "So it's like, that would have been, it would have been like a semester ending discussion.", "speaker": "SPEAKER_02"}, {"start": 2427.837, "end": 2430.24, "text": " Are we gonna allocate our TA's effort?", "speaker": "SPEAKER_02"}, {"start": 2431.001, "end": 2436.066, "text": "How many hours per week is this gonna take to, are we gonna choose to do this or that?", "speaker": "SPEAKER_02"}, {"start": 2436.206, "end": 2456.95, "text": "Like just having a calm position and knowing your affordances and those kinds of topics become really important when stuff like what we just saw happens in the setting of a software package that has methods that are tested.", "speaker": "SPEAKER_02"}, {"start": 2457.115, "end": 2459.758, "text": " That is the kind of speed that you can get.", "speaker": "SPEAKER_02"}, {"start": 2460.299, "end": 2473.295, "text": "So you can be having a conversation with people with different perspectives and skills and like desires for different interfaces to work with active inference and just sit there with your laptop and work through it.", "speaker": "SPEAKER_02"}, {"start": 2473.315, "end": 2484.068, "text": "Then the second, and I think that the deeper question Gcoma wrote,", "speaker": "SPEAKER_02"}, {"start": 2484.655, "end": 2485.736, "text": " I want to do it by hand.", "speaker": "SPEAKER_02"}, {"start": 2485.776, "end": 2487.338, "text": "LLM cannot make me learn.", "speaker": "SPEAKER_02"}, {"start": 2487.999, "end": 2488.339, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 2489.02, "end": 2506.68, "text": "Like there's so much to that question, which is like, how are we going to make our perceptions, cognitions and actions support our learning and flourishing and alignment over multiple scales when we only have partial knowledge of even some of the scales.", "speaker": "SPEAKER_02"}, {"start": 2507.701, "end": 2509.043, "text": "So like,", "speaker": "SPEAKER_02"}, {"start": 2511.217, "end": 2523.617, "text": " There's a lot of ways to use the amount of different AI and the fact that this exists as like pheromone in our niche doesn't mean that you don't want to forage.", "speaker": "SPEAKER_02"}, {"start": 2525.22, "end": 2526.442, "text": "It can be like signposting.", "speaker": "SPEAKER_02"}, {"start": 2527.223, "end": 2530.709, "text": "It can be something to ignore.", "speaker": "SPEAKER_02"}, {"start": 2532.592, "end": 2537.68, "text": "Figuring out though, and then also that process of, okay,", "speaker": "SPEAKER_02"}, {"start": 2537.879, "end": 2554.774, "text": " just copied that question in i want to learn by hand you have that kind of that kind of conversation around like it's like oh i don't know what uv is okay but now you're in the game or here's a learning loop that it suggests", "speaker": "SPEAKER_02"}, {"start": 2555.952, "end": 2564.526, "text": " I mean, how about copy an example and then render it by hand into assembly code and do the CPU additions yourself?", "speaker": "SPEAKER_02"}, {"start": 2564.687, "end": 2567.311, "text": "It's like, okay, but you didn't mean that by hand.", "speaker": "SPEAKER_02"}, {"start": 2567.852, "end": 2569.214, "text": "So what did you mean by hand?", "speaker": "SPEAKER_02"}, {"start": 2569.855, "end": 2576.947, "text": "Did you mean draw something until you have a convergence understanding with the code?", "speaker": "SPEAKER_02"}, {"start": 2577.208, "end": 2580.373, "text": "Did you mean type the Python syntax yourself?", "speaker": "SPEAKER_02"}, {"start": 2581.433, "end": 2596.257, "text": " like working and kind of zigzagging to find that out, which might not even be, it might be something about refining and operationalizing a strongly held belief that you have about how you ought to learn in that moment.", "speaker": "SPEAKER_02"}, {"start": 2597.039, "end": 2602.648, "text": "Or it could be being open to like streams of action and information that, um,", "speaker": "SPEAKER_02"}, {"start": 2603.962, "end": 2607.906, "text": " are novel in terms of what you have seen.", "speaker": "SPEAKER_02"}, {"start": 2608.646, "end": 2609.868, "text": "Oh, yeah.", "speaker": "SPEAKER_02"}, {"start": 2610.048, "end": 2612.37, "text": "I mean, it's not even tongue in cheek.", "speaker": "SPEAKER_02"}, {"start": 2612.45, "end": 2613.731, "text": "I hear you on that though, Giacomo.", "speaker": "SPEAKER_02"}, {"start": 2613.751, "end": 2615.993, "text": "It's like, that is literally the question.", "speaker": "SPEAKER_02"}, {"start": 2616.274, "end": 2620.878, "text": "It's like learning and application, epistemic value and pragmatic value.", "speaker": "SPEAKER_02"}, {"start": 2621.659, "end": 2633.53, "text": "Like in your setting, are you an intern or someone in a student's role or season of their life?", "speaker": "SPEAKER_02"}, {"start": 2634.117, "end": 2658.741, "text": " and you're looking to like learn on multiple levels about the topic and breadth and depth and tooling and metacognition, or is it like a time bound deliverable where you need to do some kind of model performance or a given system, totally different modes, but like the palette and the canvas with active inference is really big.", "speaker": "SPEAKER_02"}, {"start": 2660.324, "end": 2667.597, "text": " So then we kind of entrench that learn by hand into this documentation file, learn by hand.", "speaker": "SPEAKER_02"}, {"start": 2669.04, "end": 2670.262, "text": "We can point people to that.", "speaker": "SPEAKER_02"}, {"start": 2670.342, "end": 2671.163, "text": "It's a stable link.", "speaker": "SPEAKER_02"}, {"start": 2672.826, "end": 2673.988, "text": "They can contribute.", "speaker": "SPEAKER_02"}, {"start": 2674.189, "end": 2682.163, "text": "So that kind of brings in the open source and the open science component is like that becomes somewhere that we can reference.", "speaker": "SPEAKER_02"}, {"start": 2691.442, "end": 2691.763, "text": " Yes.", "speaker": "SPEAKER_02"}, {"start": 2692.025, "end": 2693.753, "text": "Good, good comment, Andrew.", "speaker": "SPEAKER_02"}, {"start": 2693.773, "end": 2698.414, "text": "I never know exactly which one to add, but I will.", "speaker": "SPEAKER_02"}, {"start": 2699.896, "end": 2702.538, "text": " Oh, sorry that yeah, that wasn't to you.", "speaker": "SPEAKER_01"}, {"start": 2702.598, "end": 2706.562, "text": "Someone else is sharing their repo in the chat and I just it.", "speaker": "SPEAKER_01"}, {"start": 2706.822, "end": 2714.549, "text": "It's very cool that people want to contribute more learning materials, but it's just yeah, whenever you put out a repo, you should add a separate license.", "speaker": "SPEAKER_01"}, {"start": 2714.609, "end": 2721.395, "text": "I saw that a couple of scripts like half the phrase MIT license tucked in there, but still presents potential legal issues.", "speaker": "SPEAKER_01"}, {"start": 2721.475, "end": 2729.582, "text": "So for whoever is posting about that, I would strongly suggest you add a separate like license file to the repo and then that way people can properly.", "speaker": "SPEAKER_01"}, {"start": 2729.765, "end": 2759.046, "text": " use and contribute as as you seem to be wanting them to thanks yeah that's what this one that's what this repo is mit license but but i'm i'm um not i i don't know all the situations involving different licensing but it's a key it's it's an it's a key aspect of that previous point on like open science and participation for sure um okay so again we've like looked at a bunch of different modes ranging from", "speaker": "SPEAKER_02"}, {"start": 2760.258, "end": 2774.11, "text": " browsing the fundamentals repo as a website to pasting in the URL into a search browser, looking at the code and the outputs, looking at documentation, different topics.", "speaker": "SPEAKER_02"}, {"start": 2775.211, "end": 2782.138, "text": "And then if you kind of want to break the fourth wall, you clone the repo, which is to say you bring it locally.", "speaker": "SPEAKER_02"}, {"start": 2782.898, "end": 2788.223, "text": "And even if you don't have GitHub installed, you can still go to code", "speaker": "SPEAKER_02"}, {"start": 2788.793, "end": 2790.095, "text": " and then download zip.", "speaker": "SPEAKER_02"}, {"start": 2791.097, "end": 2796.406, "text": "So technically, you don't even need Git, especially if you're not going to push back.", "speaker": "SPEAKER_02"}, {"start": 2796.426, "end": 2797.869, "text": "You can just download the zip file.", "speaker": "SPEAKER_02"}, {"start": 2798.931, "end": 2800.574, "text": "It'll be a few megabytes.", "speaker": "SPEAKER_02"}, {"start": 2801.355, "end": 2804.661, "text": "Open that folder, and then you're going to be in this setting.", "speaker": "SPEAKER_02"}, {"start": 2807.025, "end": 2812.675, "text": "Open that folder in an IDE, like Cursor or VS Code or any number of other", "speaker": "SPEAKER_02"}, {"start": 2813.65, "end": 2826.946, "text": " And then you can either just right away open the terminal and do dot slash run, or you might hit a few speed bumps depending on your environment.", "speaker": "SPEAKER_02"}, {"start": 2827.747, "end": 2836.797, "text": "And for those, it may be very useful to have a coding agent who you can say, paste in, I got this error, what do I do?", "speaker": "SPEAKER_02"}, {"start": 2837.578, "end": 2841.463, "text": "Say, oh, you need to install UV or, oh, you need to install Python or something like that.", "speaker": "SPEAKER_02"}, {"start": 2846.607, "end": 2850.031, "text": " Let's see where we got with the demos.", "speaker": "SPEAKER_02"}, {"start": 2852.154, "end": 2853.455, "text": "Just one quick question, Daniel.", "speaker": "SPEAKER_03"}, {"start": 2854.016, "end": 2857.96, "text": "I saw somewhere there was some .ipymb notebooks.", "speaker": "SPEAKER_03"}, {"start": 2858.141, "end": 2858.942, "text": "I looked through the repo.", "speaker": "SPEAKER_03"}, {"start": 2859.202, "end": 2859.863, "text": "I couldn't find them.", "speaker": "SPEAKER_03"}, {"start": 2859.903, "end": 2862.566, "text": "Am I blind or something?", "speaker": "SPEAKER_03"}, {"start": 2865.209, "end": 2866.711, "text": "I just haven't committed at all.", "speaker": "SPEAKER_02"}, {"start": 2867.952, "end": 2868.052, "text": "Okay.", "speaker": "SPEAKER_04"}, {"start": 2868.273, "end": 2868.934, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 2869.174, "end": 2870.736, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 2870.836, "end": 2870.936, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 2870.976, "end": 2871.176, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 2871.356, "end": 2871.717, "text": "It will.", "speaker": "SPEAKER_02"}, {"start": 2872.658, "end": 2874.42, "text": "We'll look at what it looks like on GitHub.", "speaker": "SPEAKER_02"}, {"start": 2875.379, "end": 2877.321, "text": " It's in output notebooks.", "speaker": "SPEAKER_02"}, {"start": 2879.184, "end": 2896.105, "text": "Just some types of things like that, like work in the Kinevin sort of complicated and simple quadrants where it's like, I mean, if there were 500,000 hours to do this, then it would be possible to port it over.", "speaker": "SPEAKER_02"}, {"start": 2896.125, "end": 2903.935, "text": "It's like, well, and then there's a few little connections like, oh, it's possible to export a thin orchestrator to a notebook.", "speaker": "SPEAKER_02"}, {"start": 2904.792, "end": 2908.136, "text": " Now you know, and it's possible to transform this into that.", "speaker": "SPEAKER_02"}, {"start": 2908.797, "end": 2910.098, "text": "And that can be done on maps.", "speaker": "SPEAKER_02"}, {"start": 2910.819, "end": 2912.681, "text": "So then it's like, can you do it once?", "speaker": "SPEAKER_02"}, {"start": 2913.161, "end": 2922.432, "text": "Because if you can do it once, it's just going to be a question of, is it going to be just as fast to do it a million times or will it take a million times as long?", "speaker": "SPEAKER_02"}, {"start": 2922.672, "end": 2929.46, "text": "But if it's a million times, a millionth of a second, then that's not going to be too bad.", "speaker": "SPEAKER_02"}, {"start": 2932.724, "end": 2933.044, "text": "Um,", "speaker": "SPEAKER_02"}, {"start": 2934.661, "end": 2943.553, "text": " Okay, so here, run the demos, output, figures, demo.", "speaker": "SPEAKER_02"}, {"start": 2945.856, "end": 2955.99, "text": "So let's look at it in Finder.", "speaker": "SPEAKER_02"}, {"start": 2956.011, "end": 2957.833, "text": "Balance on a bicycle.", "speaker": "SPEAKER_02"}, {"start": 2958.286, "end": 2974.067, "text": " we can say we want to see the animation of the bicycle um or we want to specifically be looking at these two factors or we want to study um acquisition of primary bicycle riding or we want to look at you know it's like you go from there drone flight", "speaker": "SPEAKER_02"}, {"start": 2979.059, "end": 2991.296, "text": " You know, especially if you just dashed off the prompt and looked at an output like this, like part of the metacognition is how much confidence should I have that it really did do a discrete expected free energy based plan?", "speaker": "SPEAKER_02"}, {"start": 2992.558, "end": 3006.737, "text": "And if you're in an environment where those methods are really well signposted, really flexible, really clear, really documented, really used in a lot of other examples, it will likely be that method.", "speaker": "SPEAKER_02"}, {"start": 3006.937, "end": 3028.522, "text": " But if you do this in a blank slate repo or a repo that has like sort of a mosaic of different experiments at different stages of speculative functionality, then you're more likely to get a method that meets that repo in its sort of mosaic nature.", "speaker": "SPEAKER_02"}, {"start": 3029.023, "end": 3032.387, "text": "And another way to say it would be like, it's not really gonna be doing EFE.", "speaker": "SPEAKER_02"}, {"start": 3033.008, "end": 3034.71, "text": " at least EFE in any specific sense.", "speaker": "SPEAKER_02"}, {"start": 3035.331, "end": 3039.575, "text": "Maybe that zero to one draw from the center of the distribution is exactly what you need.", "speaker": "SPEAKER_02"}, {"start": 3040.416, "end": 3050.447, "text": "But if you're looking for like a specific active inference ontological relationship with your code, you would want to investigate it at great depth, but this is just a short overview.", "speaker": "SPEAKER_02"}, {"start": 3052.369, "end": 3053.37, "text": "And then isochade.", "speaker": "SPEAKER_02"}, {"start": 3056.494, "end": 3057.114, "text": "This is a path.", "speaker": "SPEAKER_02"}, {"start": 3057.154, "end": 3062.34, "text": "And like, this is actually the relationship between the pottery visual culture", "speaker": "SPEAKER_02"}, {"start": 3063.13, "end": 3071.221, "text": " constant et al paper and the active infrared paper in 2020 and 2021 was the iCicade model.", "speaker": "SPEAKER_02"}, {"start": 3071.662, "end": 3087.523, "text": "We adapted it to the ant foraging model by just looking at movement that left a trace, not just the iCicade movement, which is also path-based and can be based upon epistemic and pragmatic value, but just doesn't leave a trace on the surface.", "speaker": "SPEAKER_02"}, {"start": 3089.847, "end": 3092.31, "text": "And then you can look at those demo,", "speaker": "SPEAKER_02"}, {"start": 3093.235, "end": 3116.965, "text": " scripts and and see what they're written okay yeah oleana great comment thanks for adding that to the questions", "speaker": "SPEAKER_02"}, {"start": 3118.177, "end": 3147.458, "text": " section there's so many angles to converge upon the material with and this is like a um this is like this is like a computer science entry point like we didn't even look at any math notation equations at all we didn't engage with like um some philosophical questions it wasn't about precision psychiatry so this is like another entry point and then figuring out in your own time allocation to this", "speaker": "SPEAKER_02"}, {"start": 3147.539, "end": 3154.37, "text": " learning and application work itself, which has like so many learning moments and rewarding moments in it.", "speaker": "SPEAKER_02"}, {"start": 3162.303, "end": 3164.647, "text": "Okay.", "speaker": "SPEAKER_02"}, {"start": 3164.667, "end": 3166.83, "text": "Any last comment or question?", "speaker": "SPEAKER_02"}, {"start": 3177.695, "end": 3193.307, "text": " uh yeah just again uh yeah there's a final note for anyone if you especially those most interested in checking out code um yeah we're doing a lot of this kind of as we move along and what daniel", "speaker": "SPEAKER_01"}, {"start": 3193.287, "end": 3204.981, "text": " has produced here is a very nice, like, attempt to already at the get-go have a more universalized approach to, like, here's how we can nicely, like, you know, here are all the chapters.", "speaker": "SPEAKER_01"}, {"start": 3205.261, "end": 3207.464, "text": "Here's a repository layout.", "speaker": "SPEAKER_01"}, {"start": 3207.504, "end": 3219.959, "text": "You know, it's getting a strong step forward on, like, building out a repository that's dedicated to this rather than, you know, doing things too piecemeal and bit by bit to where things can become even more disorganized.", "speaker": "SPEAKER_01"}, {"start": 3220.28, "end": 3221.421, "text": "That said, it is a lot.", "speaker": "SPEAKER_01"}, {"start": 3221.873, "end": 3225.402, "text": " probably at once for a lot of people who are new to this approach.", "speaker": "SPEAKER_01"}, {"start": 3225.422, "end": 3234.685, "text": "So yeah, we're very open to, you know, there are different learners coming in at different levels who've never touched computer code before versus those who are just", "speaker": "SPEAKER_01"}, {"start": 3235.222, "end": 3264.992, "text": " making machine learning models all the time and just waiting for some hierarchical predictive coding model code so they can just you know shift over to that um so yeah we're very open to different kinds of questions and and and uh preferences so just let us know thank you really quickly and then aliana so um dorsa yes we can record a short video like that but go to cursor.com and install that software go to github.com", "speaker": "SPEAKER_01"}, {"start": 3265.445, "end": 3279.006, "text": " and get an account and get the desktop application or just download the code and then get Hermes agents and then use the free LLMs just like I showed with OpenRouter.", "speaker": "SPEAKER_02"}, {"start": 3279.646, "end": 3283.012, "text": "And this is the one command that you run to install that in the terminal.", "speaker": "SPEAKER_02"}, {"start": 3284.193, "end": 3292.566, "text": "Open the terminal on your computer, look for a YouTube video for how to open the terminal, but you're going to like open the terminal, install Hermes and cursor.", "speaker": "SPEAKER_02"}, {"start": 3292.698, "end": 3295.222, "text": " Log into GitHub if you have that set up or don't.", "speaker": "SPEAKER_02"}, {"start": 3296.123, "end": 3298.587, "text": "And then literally prompt and discuss from there.", "speaker": "SPEAKER_02"}, {"start": 3299.589, "end": 3302.894, "text": "Join the Discord or email us if you have some specific questions.", "speaker": "SPEAKER_02"}, {"start": 3303.595, "end": 3308.483, "text": "But that is really as consolidated as it is at right now.", "speaker": "SPEAKER_02"}, {"start": 3309.604, "end": 3310.065, "text": "Aliana?", "speaker": "SPEAKER_02"}, {"start": 3310.72, "end": 3311.601, "text": " Okay.", "speaker": "SPEAKER_00"}, {"start": 3311.621, "end": 3321.555, "text": "Once again, I just want to extend a really big shout out and thank you to you, Daniel, to Andrew, to Frasier for the immense amount of hard work that you've had to put in and incredible consistency.", "speaker": "SPEAKER_00"}, {"start": 3322.016, "end": 3325.26, "text": "And I'm noticing that this is a new cohort for fundamentals.", "speaker": "SPEAKER_00"}, {"start": 3325.581, "end": 3326.923, "text": "There are nine previous.", "speaker": "SPEAKER_00"}, {"start": 3330.327, "end": 3336.015, "text": "And you're showing up multiple times each week, which is, it's just a true test of stamina and dedication.", "speaker": "SPEAKER_00"}, {"start": 3336.035, "end": 3338.078, "text": "And I just thank you guys so much.", "speaker": "SPEAKER_00"}, {"start": 3340.353, "end": 3340.814, "text": " Thank you.", "speaker": "SPEAKER_02"}, {"start": 3341.556, "end": 3348.716, "text": "Andrew and Fraser, both ex-interns, have done amazing learning and teaching.", "speaker": "SPEAKER_02"}, {"start": 3352.908, "end": 3353.67, "text": "Appreciate it.", "speaker": "SPEAKER_02"}, {"start": 3355.456, "end": 3359.2, "text": " Yeah, so it's all on there on fundamentals.", "speaker": "SPEAKER_02"}, {"start": 3359.901, "end": 3370.211, "text": "I've added a few questions that were put in the chat to the questions block on Coda where we will add answers to them and they'll be a bit more visible for everyone to see.", "speaker": "SPEAKER_03"}, {"start": 3370.591, "end": 3374.976, "text": "They're in chapter five right now, which is not the best place, but I just sort of mentioned that.", "speaker": "SPEAKER_03"}, {"start": 3374.996, "end": 3375.236, "text": "Cool.", "speaker": "SPEAKER_02"}, {"start": 3375.857, "end": 3376.818, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 3376.838, "end": 3378.279, "text": "So enjoy.", "speaker": "SPEAKER_02"}, {"start": 3379.921, "end": 3380.341, "text": "Give a shot.", "speaker": "SPEAKER_02"}, {"start": 3380.782, "end": 3382.884, "text": "Come back with what questions you have.", "speaker": "SPEAKER_02"}, {"start": 3383.825, "end": 3384.045, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 3384.886, "end": 3385.574, "text": " Farewell, everybody.", "speaker": "SPEAKER_02"}, {"start": 3386.101, "end": 3387.133, "text": "I'll stop the recording.", "speaker": "SPEAKER_02"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.txt new file mode 100644 index 000000000..67b659f32 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_025/transcript.txt @@ -0,0 +1,1006 @@ +SPEAKER_02: +all right welcome everyone we're in on seven seven for what it's worth part one review and code later this week will be another part one review session with fraser then we will be heading into part two of the book + +So today we'll do sort of some mixture of two things, depending on what people are curious about and want to see. + +I'm going to start with a brief, just agent July, 2026 overview free mode. + +And then we can talk more about the idea and the flow of the whole book. + +Look more at the code if we want, or go to the questions. + +So I just wanted to start with cloning the fundamentals repo. + +So Active Inference Institute, fundamentals, GitHub repo, and clone and open that folder, and I'm using cursor. + +Then I'm gonna hit Hermes update in the terminal. + +And while that's, and there are commits every few minutes, so it's a good practice, but while it's updating, you can get the Hermes agents open source harness for LLMs, or use any other, + +LLM harness if you already have one that you like. + +Right now, the model called HY3 is free, but it's just an example of free, quite good modern models that are constantly on offer from OpenRouter. + +Just go to models, prompt pricing, set it to free. + +These are all free models. + +Sometimes they have some limits for how many things you can do. + +OK, Hermes is installed. + +So in Fundamentals repo, there's two ways to interact with it. + +And there's a web interface. + +And then there's interacting with it more like through an agent in the terminal that we'll look at or through Cursor. + +The web interface pops up this list of chapter by chapter animations and visualizations. + +and a bunch of other visualizations and animations for different topics. + +And we can keep on building out more topics that people see relevant for the book. + +Or you can say, prompt, make an interactive simulation for deep generative models. + +And we'll have the chapters up here. + +And then the extras, we can have a ton of different pages. + +So that's cool. + +There's animations you can look through. + +We can develop out that web interface. + +But on the developer side, and what's really fun to explore with agentic coding is using some kind of harness and an LLM to interact with the code base. + +But just to run the code base from the root layer, just do dot slash run dot sh. + +And then you can choose to run the simulations for a given chapter or launch that web interface. + +And then when you, + +run it you get logs of the website ongoing but with hermes we have that open source harness and a free llm model so yeah this is something that doesn't cost money and say explain where and how variational free energy + +is calculated in the code base i'm just going to ask it both to cursor in this kind of chat window format and hermes in the terminal format which is largely pretty similar um and anyone want to add any comments or or questions or anything + +or just share from their experience of working with different coding methods with Active Inference? + + +SPEAKER_03: +Maybe just briefly, I mean, I've used Py, + +So for exploratory work, it's very good if you want to quickly get the lay of the land of a repo. + +But these tools here are not necessary for you to explore the repo itself, I take it. + +You can just clone the repo and have a look through the code if you wanted, or go through the web interface. + + +SPEAKER_02: +Yeah. + +Like the docs themselves are Markdown files. + +So you could just use this as a knowledge base itself, or you could look at it just on the GitHub. + +Or you could just download it and run the local interface with all of the pre computed outputs, or this is like the next step into like we can say use those methods to implement X Andrew. + + +SPEAKER_01: +Yeah, thanks. + +I would just say for any participants who are wanting to learn something more specific or they would really like to see a particular kind of demo, definitely feel free to either add a question into the discourse on the CODA in the same way that we had other questions or ask them here in the chat. + +I mean, from having talked to others at the Institute who are still new to coding itself and coming to this from a direction where they have no idea what cursor is and like a dozen other things that just happened here, I would completely understand if someone wanted something more + +Focused and kind of like centered on, you know, just being able to maybe just run something in your browser. + +So, yeah, please request those for anyone who more wants something in line with with that. + +Otherwise. + +Yeah, this is a very interesting code repo that Daniel is building out. + + +SPEAKER_02: +Okay, so. + +That was starting with the question like format and depending on like which model you're using and various other factors, the latency to the first part of the response to the completed part of the response can be quite long. + +Like different models give you different visibility into their so-called chain of thought, like their thinking steps, so-called for now. + +And also their tool use, which over the last one plus years has been one of the major domains of tool and skill improvement. + +So we can see it's not just doing next token completion in the sentence, it's doing tool search like grep. + +then it's listing different terminal outputs so in a way it is inputting tokens of text into the terminal it's just that those are on a computer so it ends up doing things on the computer that return text that are valuable for informing its next steps so this is just happening at different speed but both the hy3 model through the hermes harness in the terminal + +and the cursor side chat style, which can also be dispatched from the terminal, have given like different types of responses. + +This one was kind of plain text in, plain text out. + +In the cursor, we get some more multimedia, like visualizations like this. + +So play around, find out what representations are useful and harness tool skill improvement + +is one area where there's a ton of development, so it could look different for you in a different setup. + +So that was more of an investigation question, like where and how is free energy calculated here? + +And you could continue kind of associatively from there, like explain that like to this type of mood or person about this setting, or how would that apply to this? + +Or what would be the steps to model this kind of situation? + +um if it were talking to an agent that has like a web search tool like a perplexity tool or just yeah web search essentially you could ask what research has been done that relates to this area and it could kind of reference the code base in um by looking what locally was on the computer + +Like this left panel, which might be subliminally obvious or not, depending on how much familiarity with this format, this is just the files in the finder. + +So this is just a horizontal way to look at this. + +So that's also related to Git. + +which getting an intuition for can be helpful, but again, for another situation focusing on this repo. + +So we've asked these two different instances of different harnesses, different interface, different underlying language model, questions about the code base. + +They go about it in some different ways that are pretty interesting. + +Now let's go into building something from the components that are here. + +So we could look at a pre-existing example, or I've kind of started this prompt. + +Yeah, many good comments. + +I'm not even reading them all in the chat. + +We'll check at the end. + +I'm giving a prompt here to the cursor agents. + +Let's add a top-level folder called demo. + +Putting it all in one folder is just going to keep it clean for us versioning this or if I decide to push it or if you want to keep your experiment or something contained like in a sidecar, which is like a standalone folder inside of a repo or to the side of a repo. + +So let's add a top level folder called demo with subfolders for different demo examples. + +And there's matters of preference and taste in terms of systems engineering and interface engineering for software repos. + +So I'm just walking one path through. + +There's many ways even something of this relatively straightforward kind of software package could be done. + +So I'm just laying out one way to do this, but test and explore other ways. + +There's going to be this top level demo folder, and then there'll be different demos inside of it. + +Like situation one, two, three. + +So maybe in the chat or verbally, like let's pick two or three different examples. + +Just write them in the chat now. + + +SPEAKER_04: +Sorry, I like example 2.3. + +I think that's a really fundamental example. + + +SPEAKER_02: +Okay. + +I think we have the book examples. + +I'm just going to prompt it to do some other situational example. + +But yes, let's look at 2.3. + +So the demo folder, and then like, you know, baseball game, person walking, thermometer with humidity sensor, like whatever situations. + +And those will be each in their own folder under demo. + +And then I'm saying, start with using the core methods of source to make it. + +SRC is kind of a shorthand for where the source code methods are that get modularly tested by the tests repo or test folder. + +So that's kind of like the atomic or nuclear components, components. + +Then scripts or orchestrators call and configure sets of methods, but they don't necessarily define new or complex logic by themself. + +So that's just one separation of consideration to have the underlying methods in source or something similarly named, tested, + +at the unit test level with tests and then get called and invoked by thin orchestrators or scripts, which make it so that you can improve and version the underlying methods and kind of keep a level of abstraction in terms of like, we want to be able to call, um, + +example 3.3 with our most updated versions of the configuration and the implementation and the logging and the visualizing whereas if chapter 3 was just scripts for just chapter 3 then we could get divergence with the same method from different chapters um but using demo we can just kind of make some other random things so just type type a random situation or something for uh + +that we'll make them. + +And then while it's making those demos, we'll look at one of the examples that Frazier mentioned. + + +SPEAKER_04: +I'll put something in the chat myself for that. + + +SPEAKER_02: +Okay. + +How about isocades? + +Riding a bicycle. + +Someone add one or type one. + + +SPEAKER_04: +Simulating active inference processes by message passing. + +They have a cool Bayesian thermostat agent. + +I'll paste that in the chat. + + +SPEAKER_02: +Okay. + +Drone flight. + +Could take learnable loop, paste the blog in, say transpose this to this framework. + +Okay. + +All right. + +I'm going to switch from agent mode just to show one mode. + +Different harnesses, again, have different features, but I'm going to switch it to plan mode, which is that it's going to set off to make a work plan for how to do this request, which, as far as I understand, + +is usually as good or better. + +There's rarely a situation where simply launching into it immediately seems like it'd be better, except unless it's going to cost extra tokens for something that's irrelevant. + +But it's kind of fun to see, and then we'll have the opportunity to look at the plan. + +But things like plan, loop, goal, memory layers, these are having a lot of impact in long horizon task planning. + +and more agentic orchestration methods because it's not like it's a rolling window of memory and it forgets where it was. + +If it knows where the work plan is, then a hierarchically organized work plan, even that has hundreds or tens of thousands of task items with the right kind of harness and listener environment, you can just continue to go for. + +Okay, so while it's making its plan, let's look at + +Example 2.3. + +Yes, go ahead, Frasier. + +Introduce it or what's important about it. + + +SPEAKER_03: +Well, actually, I just want to mention, I just cloned the repo on my local machine. + +I used UV to install everything and I ran, I ran chapter, you know, 2.3, the precision example. + +I'm just wondering where the outputs would have gone. + +I didn't run it in interactive mode. + +So I just kind of ran it through the, through the terminal. + + +SPEAKER_02: +Yep. + +Yep. + +Yep. + +Here's, here's what that looks like. + +So I'm in the top level of fundamentals dot slash run. + +I'm going to, + +hit option A, run, yeah, option A or X. I'll run the chapters. + +And then it's gonna log where it saves things. + +So output slash figures slash chapter. + +So output slash figures slash chapters. + +And that's just this folder, like example 2.3 + +Here's just that image. + +Yeah, Sebastian wrote, if you don't have video RAM for your own LLM, I can advise to try Chinese models like Minimax. + +Yeah, again, on OpenRouter, there's always, or News Portal, there's always free models. + +Those are getting better and better. + +And, or you can run + +local LLMs or go to a friend's server, but like Ollama or other methods. + +Giacomo wrote, does it log to a timestamped file? + +Definitely in the terminal, you'll get a timestamp and in the metadata of the file, you'll get a timestamp. + +Okay, so here we go. + +Back to this cursor chat. + +So we asked it to go to plant mode. + +And then this is another area of both utility and frustration for the human side. + +It's like, it's asking questions. + +So again, at that sort of qualitative level, putting aside all that we're learning about energy based methods and all that, this is like an active inference. + +Like it identified somehow that it had enough of an uncertainty to cross the threshold + +to break a loop and go into asking us a question. + +So, should the new demo folder be first class, like chapters and extra, menu, web, smoke, orchestrator, save, or standalone scripts only? + +Full integration. + +This is kind of like one reason why having a modular, signposted, composable software package, even from the beginning, is super effective, because it's like, now we can wire a whole new + +format or a whole new media type because there's already multiple things ongoing. + +And yes, that can be like in overkill or overly restraining mode of software. + +However, finding that balance of like structural plasticity and form, that's kind of the art and the science. + +All right, so here's what the plan looks like. + +There's the plan and the to-dos. + +At this point, we could give feedback or we can just hit build. + +Okay. + +Okay. + +So that was all this, that was just the scripts running from that run. + +So it logs what was output, right? + +Click kill terminal, right? + +Click kill term. + +That's the, that's the terminal that's running the website. + +This is that high three model. + +So can you review the repo? + +and give me a learning path given I have taken frequentist but no Bayesian statistics. + +Okay, then on the right side, + +We can see that plan getting built. + +Because we, yeah, Fraser. + + +SPEAKER_03: +I was just gonna say, it is nice to be able to have this ability to query your, you know, LLM assistant here to be like, all right, well, you know, I'm working on this, but really how does this relate to this other thing I'm thinking about in much the manner you did just then? + +um that is that's a nice workflow for someone who doesn't know as much about active inference as they would like but maybe already knows some their way around a code repository or something like that so it is it is a powerful way to do things yeah thank you yeah totally agree some of you probably will be very familiar with an ide and use it in a similar way or very familiar with ide not seen it used in this way or have your own + + +SPEAKER_02: +more ergonomic way for using it, or maybe you haven't used an IDE, but there's a variety of free ones for that too. + +So because in that demo ask, we went for the full integration. + +We could have like, even without much software engineering, you could imagine like, okay, we're gonna have this demos zone. + +Is it gonna be like its own self-contained sandbox? + +Like have all the methods it needs so that it can be picked up and moved outside and still work? + +is it gonna make like light use of some aspects of the package, which would have like kind of demo specific logic in the package or in the demo folder, and then like more core methods from the main package, but then it would require that package. + +or the demo could be fully integrated like the extras where they're only thin orchestrators and all the core methods go to the core software package. + +So that's a good question of how integrated do you want that time cone kind of software blanket to be? + +So, because we want more integration, it is writing these demo methods, + +under the main src so it's most integrated you can still extract some methods but but it's connected with the core testing architecture it's not just like standalone but you could have also just gone to a blank folder and say like from scratch make me a standalone active inference drone example + +Now, that's very possible now, but part of the reason for having reusable software is so that we can validate, benchmark, all these kinds of things and know that we're not just making a method that says one thing and does another. + +So it's going through. + +The green are additions. + +The red are subtractions of numbers of lines. + +So just showing its sequence of tool use and file modification. + +And then it's showing here's 27 files that it's modified. + +We can see those are those three examples and it's three of six done on the plan. + +Okay. + +Now back over here to this and like Frazier brought up, like that's what is so fun to explore and figure out the right attention allocation, regimes of attention. + +Like, + +Do you want to focus more on just following one conversation like that prompt that we asked about a learning path for Bayesian statistics? + +Or are we going to be like switching between different projects and kind of like another layer of attentional move? + +I'm using the cursor IDE. + +However, even this side chat AI functionality is available through a plugin, even if your IDE is not AI-based. + +It's a dark mode theme. + +But this kind of represents two modes + +two points on that continuum in that cursor window, we're basically doing, we're modifying the core package. + +Like, we're modifying surfaces and functions of the core repo. + +And then in this chat window, we're not even modifying the repo at all. + +We might not even be asking something related to the repo. + +We could just, like... + +spin up a Hermes and ask it some other question and it may or may not even use the repo for for context um and then in the middle is like you're leaving the core logic of the repo intact but you might be doing some side experiment that's an important access is like where are you actually chipping away or adding to the repos code base versus where are you just + +using the repo for context or not to make some other point. + +Okay, so here it ended its response. + +We asked, we know about frequentist stats, but not Bayesian statistics. + +This that just popped up is a terminal window that cursor spins up. + +where it's gonna run the scripts that it wrote in the demo sections. + +How about just report the learning path in a sequence of 50 creative song titles from the 1970s? + +See where that goes. + +Anyone want to add a question or anything else about how they're using coding agents or like where, what parts look familiar or would they add in, or, or it could be useful for somebody who's learning how to use this. + + +SPEAKER_03: +Certainly for people who maybe don't have much experience, it can look a bit massively confusing, especially for a virtuoso like Daniel. + +But it is actually not. + +It doesn't have to be as sophisticated as this or, dare I say, as confusing. + +But basically, the repo is just placed where we host our examples. + +and whatever other examples, and your workflow, your chosen way of interacting that is totally up to you. + +So this is definitely a powerful modern way to do that. + +But you can do more traditional, you know, I'm just going to clone the repository and have a look manually. + +That's totally up to you. + +So don't feel like this is like Howard saying you have to do things if you want to interact with the repo. + + +SPEAKER_02: +Thank you. + +Totally agreed. + +This is just one way of working kind of in the engine room, as it were, and help us and collaborate with everyone to co-create the interfaces that would put you in like a flow learning state. + +And maybe you'd get familiar with things that don't even exist today, but in six months, they could be a super effective way to learn. + +Andrew? + + +SPEAKER_01: +Yeah, thanks. + +Yeah, I, on Fraser's point earlier, I did, you know, I haven't had the time and haven't received enough participant feedback quite yet to know how to build this out further. + +But there is one demo for chapter two that I made some time ago, and I dropped the link to that in the chat. + +And that's a very, very different approach, which is simply you open it up in your web browser. + +There's no installation required. + +There's no prompting LLMs. + +There's none of that. + +And it's just a very direct, like you can just click run all at the top and it runs all the code in your browser for you. + +And then the big thing is that it has these like sliders that you can interact with. + +And so the code that's there currently can just show you like, + +here's what happens whenever you change the mean of the prior versus the mean of the likelihood, things like that, all the things we talk about in chapter two. + +And then it also has maximum likelihood estimation with gradient descent, as well as map with gradient descent. + +And so I made that with the intent of, you know, there are many people who are new here who are + +coming in they're learning coding they're learning things like that they want to know how to get involved in active inference and then very quickly they get swept away by a lot of like difficult you know sort of different things to to understand and so um yeah that was the focus of that demo so if anyone you know finds that to be useful or they want to see like another more directly like interactive without necessarily like having an llm handle a bunch of + +things for you and just bearing in mind that all lms are kind of prone to error so um you know there's just different ways to do things um so yeah just wanted to have that known that that's an affordance thank you magdalena um yeah my question is um i'm sorry that i don't know the answer but um does do you have a link + + +SPEAKER_05: +To a site in the active inference Institute that walks people who are really, really super new to coding and everything. + +where it describes what are the basic skills that you need in order to engage with these models, right? + +So what I guess I'm asking the experts is, as a beginner, I don't have an idea as to where I can stop in my basic learning and then proceed to use these models. + +So is there something already in the Active Inference Institute that can walk us through that? + +phase of our learning. + + +SPEAKER_02: +It's an important question. + +I think there's kind of a step zero right before this video picks up with downloading cursor or an IDE and installing Hermes, installing Git and cloning the fundamentals repo. + +So that is some information that somebody could either ask about or search to install the IDE. + +install GitHub okay and install Hermes agent to recapitulate this exact material and then pick up right at the beginning of this meeting um and I think like we can continue to explore what artifacts and and onboarding material it just there are so many onboardings to programming because + +there's an infinite number of places to come to and go to. + +This in terms of like the time that I've been learning these topics, this is like on multiple orders of magnitude, different setup and applicability duration, like in a mind boggling fashion, just getting things like the packages installed + +was like a quite challenging situation, let alone the natural conversation with a code base and the ability to like convert a sort of question like Giacomo's, how much of it could be ported to Jupyter Notebooks? + +It's just like all of it. + +Make that transformation. + +That's it, it's not a huge cognitive lift. + +No harm in trying. + +Building expertise in what you can ask and what is like trivially tiny, it's gonna finish in 15 seconds versus what would require like a structured work plan. + +or an even more advanced way to do task allocation. + +That's the art and the science of understanding how to work with these synthetic intelligence systems. + +But just watching what's happening right here should put a few dots down. + +And then how you want to allocate your regime of attention. + +Do you like one chat + +looking more like a chat window. + +There's a lot of different user interface services for this in terms of the more generative AI end. + +You could paste the link into Perplexity or Google search on your browser with no software and just say, what is this? + +perplexity and just what is this? + +Or you can read the GitHub static documentation or paste it into something. + +And then we're making the notebooks, which gives a good intermediate level because it's browser based, but runs the code on through running it locally. + +which has the most configurability and resilience, and whether your drone demo is ready to be relied upon. + +That's a second question or a fifth question or something, but is it one prompt away to make an Active Inference-esque drone? + +Yeah. + +and like just what was a bigger or a smaller jump and making visible some of these computer scientific steps that otherwise are like pretty down the road. + +It just flips that. + +Fraser? + + +SPEAKER_05: +Thank you. + + +SPEAKER_03: +Yeah, I just wanted to say real quick, I guess there isn't yet a good resource + +the Active Infants Institute ecosystem itself yet, which would maybe walk someone through a program, a way to learn the fundamentals in terms of the technological skills and + +out there that you can interface with. + +The problem that you have is that there's so many of them that do exist already. + +So in some sense, you just need to choose one and go with it. + +And it obviously depends on who you are. + +I will just say that the way that I learned Python, I obviously went to university to do computer science and maths. + +There's a lot of programming there. + +But the way that I actually learned Python was through this course here. + +millions of people have done and they really like. + +It's on a platform called Udemy. + +So that would be, if you don't know any Python programming and you want to learn some Python, that's an excellent resource. + +There are, in terms of skills that you would need, really you need probably some basic familiarity with Python and some basic familiarity with . + +You know, that's pretty much it. + +And then everything from there is really up to you in terms of the agential workflow. + +This is really quite cool what Daniel is showing, but it is only one of the interpretation as to how it is that you can interact with this material. + +So you basically Python syntax and like what is GitHub and how do I use it? + +Really, those are the two things you need. + +So there isn't yet a place that I'm aware of. + +Although, Andrew, maybe you've got your hand up if there's anything you'd like to say on that. + + +SPEAKER_01: +Sure, thanks. + +Yeah, I just wanted to add to sort of the list of affordances there because I, you know, I also went to university for like, you know, economics and stats and various other things that have been helpful. + +But in this context, as far as coding and programming, you know, there's some free resources like W3 is a website that has a lot of common functionalities for programming languages like Python and R and probably Julia by now. + +And then Datacamp was very helpful for those who need kind of a gamified approach. + +And then there's Coursera, which is very helpful. + +that you can kind of earn certifications from university designed online courses. + +So some find those useful. + +It leads to something you can add sort of to your CV while along the way, like doing hands-on projects. + +But then from there, the big thing is like, can you actually directly interact with the + +The inputs intermediate outputs and final outputs of the code that you're creating. + +So, for those who don't know what a notebook is, a notebook is something that allows you to sort of run code. + +Chunk by chunk sort of or cell by cell and that's a really nice way to both demonstrate how code works as well as, like, incrementally on your own build up. + +Sometimes a lot of motivation comes from choosing some kind of project that is interesting to you. + +And then from there, just sort of building up from there, including learning about all the trials and errors that come along the way. + +So, because it's just the thing with LLMs is that there's a lot of, you know, the catch word of the day is vibe coding. + +And it was interesting. + +interesting point made in the chat that yeah the while a lot of these things look look interesting i just you know it's sometimes one can miss the forest for the trees um whenever they they do this kind of um llm based coding but yeah whatever whatever suits your style yep yep so first on the notebook question it was like we just saw it happening + + +SPEAKER_02: +from the question to this, it landed the plane on all the chapter notebooks. + +Now, maybe some of the scripts need to be slightly updated in order to run in the notebook format. + +But like, it is running the notebooks using the methods from the package. + +So it's like, that would have been, it would have been like a semester ending discussion. + +Are we gonna allocate our TA's effort? + +How many hours per week is this gonna take to, are we gonna choose to do this or that? + +Like just having a calm position and knowing your affordances and those kinds of topics become really important when stuff like what we just saw happens in the setting of a software package that has methods that are tested. + +That is the kind of speed that you can get. + +So you can be having a conversation with people with different perspectives and skills and like desires for different interfaces to work with active inference and just sit there with your laptop and work through it. + +Then the second, and I think that the deeper question Gcoma wrote, + +I want to do it by hand. + +LLM cannot make me learn. + +Yeah. + +Like there's so much to that question, which is like, how are we going to make our perceptions, cognitions and actions support our learning and flourishing and alignment over multiple scales when we only have partial knowledge of even some of the scales. + +So like, + +There's a lot of ways to use the amount of different AI and the fact that this exists as like pheromone in our niche doesn't mean that you don't want to forage. + +It can be like signposting. + +It can be something to ignore. + +Figuring out though, and then also that process of, okay, + +just copied that question in i want to learn by hand you have that kind of that kind of conversation around like it's like oh i don't know what uv is okay but now you're in the game or here's a learning loop that it suggests + +I mean, how about copy an example and then render it by hand into assembly code and do the CPU additions yourself? + +It's like, okay, but you didn't mean that by hand. + +So what did you mean by hand? + +Did you mean draw something until you have a convergence understanding with the code? + +Did you mean type the Python syntax yourself? + +like working and kind of zigzagging to find that out, which might not even be, it might be something about refining and operationalizing a strongly held belief that you have about how you ought to learn in that moment. + +Or it could be being open to like streams of action and information that, um, + +are novel in terms of what you have seen. + +Oh, yeah. + +I mean, it's not even tongue in cheek. + +I hear you on that though, Giacomo. + +It's like, that is literally the question. + +It's like learning and application, epistemic value and pragmatic value. + +Like in your setting, are you an intern or someone in a student's role or season of their life? + +and you're looking to like learn on multiple levels about the topic and breadth and depth and tooling and metacognition, or is it like a time bound deliverable where you need to do some kind of model performance or a given system, totally different modes, but like the palette and the canvas with active inference is really big. + +So then we kind of entrench that learn by hand into this documentation file, learn by hand. + +We can point people to that. + +It's a stable link. + +They can contribute. + +So that kind of brings in the open source and the open science component is like that becomes somewhere that we can reference. + +Yes. + +Good, good comment, Andrew. + +I never know exactly which one to add, but I will. + + +SPEAKER_01: +Oh, sorry that yeah, that wasn't to you. + +Someone else is sharing their repo in the chat and I just it. + +It's very cool that people want to contribute more learning materials, but it's just yeah, whenever you put out a repo, you should add a separate license. + +I saw that a couple of scripts like half the phrase MIT license tucked in there, but still presents potential legal issues. + +So for whoever is posting about that, I would strongly suggest you add a separate like license file to the repo and then that way people can properly. + + +SPEAKER_02: +use and contribute as as you seem to be wanting them to thanks yeah that's what this one that's what this repo is mit license but but i'm i'm um not i i don't know all the situations involving different licensing but it's a key it's it's an it's a key aspect of that previous point on like open science and participation for sure um okay so again we've like looked at a bunch of different modes ranging from + +browsing the fundamentals repo as a website to pasting in the URL into a search browser, looking at the code and the outputs, looking at documentation, different topics. + +And then if you kind of want to break the fourth wall, you clone the repo, which is to say you bring it locally. + +And even if you don't have GitHub installed, you can still go to code + +and then download zip. + +So technically, you don't even need Git, especially if you're not going to push back. + +You can just download the zip file. + +It'll be a few megabytes. + +Open that folder, and then you're going to be in this setting. + +Open that folder in an IDE, like Cursor or VS Code or any number of other + +And then you can either just right away open the terminal and do dot slash run, or you might hit a few speed bumps depending on your environment. + +And for those, it may be very useful to have a coding agent who you can say, paste in, I got this error, what do I do? + +Say, oh, you need to install UV or, oh, you need to install Python or something like that. + +Let's see where we got with the demos. + + +SPEAKER_03: +Just one quick question, Daniel. + +I saw somewhere there was some .ipymb notebooks. + +I looked through the repo. + +I couldn't find them. + +Am I blind or something? + + +SPEAKER_02: +I just haven't committed at all. + + +SPEAKER_04: +Okay. + + +SPEAKER_02: +All right. + +Yeah. + +Okay. + +Yeah. + +It will. + +We'll look at what it looks like on GitHub. + +It's in output notebooks. + +Just some types of things like that, like work in the Kinevin sort of complicated and simple quadrants where it's like, I mean, if there were 500,000 hours to do this, then it would be possible to port it over. + +It's like, well, and then there's a few little connections like, oh, it's possible to export a thin orchestrator to a notebook. + +Now you know, and it's possible to transform this into that. + +And that can be done on maps. + +So then it's like, can you do it once? + +Because if you can do it once, it's just going to be a question of, is it going to be just as fast to do it a million times or will it take a million times as long? + +But if it's a million times, a millionth of a second, then that's not going to be too bad. + +Um, + +Okay, so here, run the demos, output, figures, demo. + +So let's look at it in Finder. + +Balance on a bicycle. + +we can say we want to see the animation of the bicycle um or we want to specifically be looking at these two factors or we want to study um acquisition of primary bicycle riding or we want to look at you know it's like you go from there drone flight + +You know, especially if you just dashed off the prompt and looked at an output like this, like part of the metacognition is how much confidence should I have that it really did do a discrete expected free energy based plan? + +And if you're in an environment where those methods are really well signposted, really flexible, really clear, really documented, really used in a lot of other examples, it will likely be that method. + +But if you do this in a blank slate repo or a repo that has like sort of a mosaic of different experiments at different stages of speculative functionality, then you're more likely to get a method that meets that repo in its sort of mosaic nature. + +And another way to say it would be like, it's not really gonna be doing EFE. + +at least EFE in any specific sense. + +Maybe that zero to one draw from the center of the distribution is exactly what you need. + +But if you're looking for like a specific active inference ontological relationship with your code, you would want to investigate it at great depth, but this is just a short overview. + +And then isochade. + +This is a path. + +And like, this is actually the relationship between the pottery visual culture + +constant et al paper and the active infrared paper in 2020 and 2021 was the iCicade model. + +We adapted it to the ant foraging model by just looking at movement that left a trace, not just the iCicade movement, which is also path-based and can be based upon epistemic and pragmatic value, but just doesn't leave a trace on the surface. + +And then you can look at those demo, + +scripts and and see what they're written okay yeah oleana great comment thanks for adding that to the questions + +section there's so many angles to converge upon the material with and this is like a um this is like this is like a computer science entry point like we didn't even look at any math notation equations at all we didn't engage with like um some philosophical questions it wasn't about precision psychiatry so this is like another entry point and then figuring out in your own time allocation to this + +learning and application work itself, which has like so many learning moments and rewarding moments in it. + +Okay. + +Any last comment or question? + + +SPEAKER_01: +uh yeah just again uh yeah there's a final note for anyone if you especially those most interested in checking out code um yeah we're doing a lot of this kind of as we move along and what daniel + +has produced here is a very nice, like, attempt to already at the get-go have a more universalized approach to, like, here's how we can nicely, like, you know, here are all the chapters. + +Here's a repository layout. + +You know, it's getting a strong step forward on, like, building out a repository that's dedicated to this rather than, you know, doing things too piecemeal and bit by bit to where things can become even more disorganized. + +That said, it is a lot. + +probably at once for a lot of people who are new to this approach. + +So yeah, we're very open to, you know, there are different learners coming in at different levels who've never touched computer code before versus those who are just + +making machine learning models all the time and just waiting for some hierarchical predictive coding model code so they can just you know shift over to that um so yeah we're very open to different kinds of questions and and and uh preferences so just let us know thank you really quickly and then aliana so um dorsa yes we can record a short video like that but go to cursor.com and install that software go to github.com + + +SPEAKER_02: +and get an account and get the desktop application or just download the code and then get Hermes agents and then use the free LLMs just like I showed with OpenRouter. + +And this is the one command that you run to install that in the terminal. + +Open the terminal on your computer, look for a YouTube video for how to open the terminal, but you're going to like open the terminal, install Hermes and cursor. + +Log into GitHub if you have that set up or don't. + +And then literally prompt and discuss from there. + +Join the Discord or email us if you have some specific questions. + +But that is really as consolidated as it is at right now. + +Aliana? + + +SPEAKER_00: +Okay. + +Once again, I just want to extend a really big shout out and thank you to you, Daniel, to Andrew, to Frasier for the immense amount of hard work that you've had to put in and incredible consistency. + +And I'm noticing that this is a new cohort for fundamentals. + +There are nine previous. + +And you're showing up multiple times each week, which is, it's just a true test of stamina and dedication. + +And I just thank you guys so much. + + +SPEAKER_02: +Thank you. + +Andrew and Fraser, both ex-interns, have done amazing learning and teaching. + +Appreciate it. + +Yeah, so it's all on there on fundamentals. + + +SPEAKER_03: +I've added a few questions that were put in the chat to the questions block on Coda where we will add answers to them and they'll be a bit more visible for everyone to see. + +They're in chapter five right now, which is not the best place, but I just sort of mentioned that. + + +SPEAKER_02: +Cool. + +All right. + +So enjoy. + +Give a shot. + +Come back with what questions you have. + +All right. + +Farewell, everybody. + +I'll stop the recording. diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.json new file mode 100644 index 000000000..771f71717 --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.json @@ -0,0 +1 @@ +[{"video_id": "FnYCaCkah4U", "segments": [{"start": 3.625, "end": 28.536, "text": " all right hello everyone so it's uh july 10th 2026 we're here at the active inference institute going through fundamentals of active inference so this is our second week of review we've we've gone through part one of the textbook we're moving into part two we're going to be kicking up with chapter six next week uh and part two is all about kind of active inference proper when we actually begin to put action into the mix", "speaker": "SPEAKER_04"}, {"start": 29.107, "end": 40.749, "text": " So this is the last session of chapter one review, maybe more general questions and things like that before we do move on to the more regular pace of things again, where we go chapter by chapter.", "speaker": "SPEAKER_04"}, {"start": 42.011, "end": 51.79, "text": "So this session today, this is an opportunity to ask questions, live questions, or maybe we can go through some of the prerecorded ones as well relating to really anything to do with part one.", "speaker": "SPEAKER_04"}, {"start": 52.259, "end": 54.662, "text": " So that's what we're all here to do.", "speaker": "SPEAKER_04"}, {"start": 54.682, "end": 59.629, "text": "I will say before we do that, there's interest.", "speaker": "SPEAKER_04"}, {"start": 59.969, "end": 64.395, "text": "And indeed, we have been big because we have gone through part one.", "speaker": "SPEAKER_04"}, {"start": 64.415, "end": 65.296, "text": "There's three parts of the book.", "speaker": "SPEAKER_04"}, {"start": 66.258, "end": 71.685, "text": "We've gotten somewhat established in terms of the general lay of the land and things like this.", "speaker": "SPEAKER_04"}, {"start": 72.188, "end": 85.683, "text": " And we now want to think about maybe ways in which you guys and participants at all levels would like to participate in experimenting with things practically.", "speaker": "SPEAKER_04"}, {"start": 86.564, "end": 96.436, "text": "So I just want to let you all know that we are actually, so Daniel, myself, Andrew, and even Sanjeev are actively thinking about the best way to sort of create those affordances.", "speaker": "SPEAKER_04"}, {"start": 96.816, "end": 101.301, "text": "As we saw with Daniel's session last time, I'll just share my screen here.", "speaker": "SPEAKER_04"}, {"start": 102.074, "end": 104.117, "text": " Probably the entire screen is best.", "speaker": "SPEAKER_04"}, {"start": 105.659, "end": 106.9, "text": "Let me see the big nested things.", "speaker": "SPEAKER_04"}, {"start": 106.92, "end": 108.042, "text": "Hopefully everyone can see my screen.", "speaker": "SPEAKER_04"}, {"start": 108.082, "end": 109.143, "text": "I'll just maximize this here.", "speaker": "SPEAKER_04"}, {"start": 110.124, "end": 111.866, "text": "So we're seeing the coder just now.", "speaker": "SPEAKER_04"}, {"start": 112.968, "end": 123.922, "text": "As we saw last time, Daniel went over the existing, there is a place where we host the code for the existing code such as it is in GitHub.", "speaker": "SPEAKER_04"}, {"start": 124.563, "end": 126.405, "text": "And this was really cool.", "speaker": "SPEAKER_04"}, {"start": 126.425, "end": 127.286, "text": "There's a lot of stuff here.", "speaker": "SPEAKER_04"}, {"start": 127.326, "end": 127.987, "text": "It's very rich.", "speaker": "SPEAKER_04"}, {"start": 129.409, "end": 131.271, "text": "Daniel's putting quite a bit of work about", "speaker": "SPEAKER_04"}, {"start": 131.707, "end": 137.738, "text": " You know how it is that you can palpate and go through this at your own pace.", "speaker": "SPEAKER_04"}, {"start": 137.758, "end": 141.746, "text": "It is a bit overwhelming at first, maybe for people who are new.", "speaker": "SPEAKER_04"}, {"start": 141.826, "end": 142.888, "text": "So it's very powerful.", "speaker": "SPEAKER_04"}, {"start": 143.028, "end": 148.298, "text": "So pretty much all of the code is here for all of the chapters.", "speaker": "SPEAKER_04"}, {"start": 148.278, "end": 149.239, "text": " Very cool stuff.", "speaker": "SPEAKER_04"}, {"start": 149.34, "end": 157.592, "text": "But one thing that we're quite enthusiastic about is surfacing maybe different levels of participation.", "speaker": "SPEAKER_04"}, {"start": 158.193, "end": 172.193, "text": "So we saw in, if I go to chapter two here, I tend to think that it's probably a good idea for, you know, there's a lot of people in this group, all different backgrounds and so on, software developers, there's even a professor or two, and then there's", "speaker": "SPEAKER_04"}, {"start": 174.957, "end": 176.547, "text": " variety of backgrounds and perspectives.", "speaker": "SPEAKER_04"}, {"start": 177.393, "end": 181.238, "text": "I am thinking it would be quite nice to", "speaker": "SPEAKER_04"}, {"start": 181.572, "end": 188.183, "text": " have a sort of opportunity for you to just explore the code very much along the lines of what Andrew did for chapter two.", "speaker": "SPEAKER_04"}, {"start": 189.385, "end": 203.689, "text": "So if we come over here, Andrew created this very wonderful notebook where you don't have, I mean, there is code, you have to run some stuff for some setup, but fundamentally what you can do is you can just sort of look at an example.", "speaker": "SPEAKER_04"}, {"start": 203.729, "end": 206.213, "text": "This is the canonical example in chapter two.", "speaker": "SPEAKER_04"}, {"start": 206.193, "end": 222.322, "text": " I think figure 2.11, or at least that's what this instantiates, where you can have a look at the general relationship that we've got here and actually play around interactively with, okay, well, what if the observation Y was 2.6?", "speaker": "SPEAKER_04"}, {"start": 222.342, "end": 225.868, "text": "How does this affect the likelihood and then the posterior distribution?", "speaker": "SPEAKER_04"}, {"start": 226.309, "end": 228.032, "text": "You can sort of do this interactively.", "speaker": "SPEAKER_04"}, {"start": 228.012, "end": 236.088, "text": " And we shared this for chapter two, but we haven't yet implemented these kinds of notebooks for subsequent chapters.", "speaker": "SPEAKER_04"}, {"start": 236.95, "end": 249.295, "text": "I think for someone really new and who doesn't have much experience, this is probably a very nice way to begin to have some ingress into being able to play around with these ideas interactively.", "speaker": "SPEAKER_04"}, {"start": 249.798, "end": 272.076, "text": " um so that'd be great so we have you know this is just looking at the effects of changing correct while changing the parameters of various distributions when it comes to bayesian inference there's a wonderful example down here as well on gradient descent some of this can be cleaned up and maybe we can hide some things that don't necessarily need to be in here or whatever but this style of interaction i think is a very", "speaker": "SPEAKER_04"}, {"start": 272.056, "end": 277.482, "text": " excellent way to begin interacting with practical examples.", "speaker": "SPEAKER_04"}, {"start": 277.502, "end": 279.184, "text": "So you don't have to set up anything.", "speaker": "SPEAKER_04"}, {"start": 279.224, "end": 281.107, "text": "You don't have to download anything.", "speaker": "SPEAKER_04"}, {"start": 281.167, "end": 285.252, "text": "You don't have to worry about GitHub repositories or anything like that.", "speaker": "SPEAKER_04"}, {"start": 285.272, "end": 293.281, "text": "You just click on this link here and you'll go to this web page and you can just directly interact in the browser, which is very, very cool.", "speaker": "SPEAKER_04"}, {"start": 294.242, "end": 299.228, "text": "Now, for those of you who do want more of an interaction side of thing,", "speaker": "SPEAKER_04"}, {"start": 299.849, "end": 314.35, "text": " If we go here, Sanjeev has actually prepared, these are quite old, but he's prepared notebooks, code notebooks that are not yet available on the book.", "speaker": "SPEAKER_04"}, {"start": 314.671, "end": 318.496, "text": "So the idea is that these would be very similar to what you're seeing over here.", "speaker": "SPEAKER_04"}, {"start": 318.536, "end": 322.482, "text": "This is a notebook style thing where you've got these cells and you can run some code.", "speaker": "SPEAKER_04"}, {"start": 323.052, "end": 343.947, "text": " also some notes you know this is marked out notes whatever fundamentally he's done a similar kind of thing where you know chapter one all the way through to I think chapter 10 but he hasn't had time really to to build this out in a way that would be professional enough to release yes to the public", "speaker": "SPEAKER_04"}, {"start": 344.804, "end": 369.782, "text": " so this style of thing is very much what he's interested in in affording and the idea would be okay maybe you can go into chapter one uh well let me see here i'm not going to actually render anything particularly interesting to look at but if we take a look at a particular notebook you know these are these are hosted here on on github so the idea would be someone who's a little bit more interested in messing around with the code itself", "speaker": "SPEAKER_04"}, {"start": 369.948, "end": 373.253, "text": " you would have the ability to have this notebook here, maybe hosted on GitHub.", "speaker": "SPEAKER_04"}, {"start": 373.713, "end": 384.389, "text": "We could very easily have our own version of this in the fundamentals repo, where they're version controlled, they're on GitHub, so you clone the repo and then kind of play around with stuff in the notebook.", "speaker": "SPEAKER_04"}, {"start": 385.29, "end": 389.917, "text": "And then there would be hooks for you to directly interface with certain parts of the code.", "speaker": "SPEAKER_04"}, {"start": 389.937, "end": 393.542, "text": "So rather than, let's say, okay, we've got a generative model here,", "speaker": "SPEAKER_04"}, {"start": 393.522, "end": 396.248, "text": " And this is our observation likelihood.", "speaker": "SPEAKER_04"}, {"start": 396.829, "end": 403.423, "text": "And oh, look, we have the function that generates our observation likelihood is a linear model.", "speaker": "SPEAKER_04"}, {"start": 403.443, "end": 409.156, "text": "And then what you could do is you could come down here and say, all right, well, my parameters for my linear model, they're beta 1 and beta 0.", "speaker": "SPEAKER_04"}, {"start": 410.278, "end": 410.699, "text": "They're here.", "speaker": "SPEAKER_04"}, {"start": 410.759, "end": 413.505, "text": "What if I change beta 1 to be", "speaker": "SPEAKER_04"}, {"start": 413.941, "end": 436.101, "text": " something else you know you directly manipulate the code and then you could sort of see by changing the structure of the model and the structure of the environment how things change so that would be a little bit more of a direct use of you know python so the idea with the explorer you know the explorer type thing you wouldn't have to know any code to be able to interact with any of this stuff okay", "speaker": "SPEAKER_04"}, {"start": 436.638, "end": 443.189, "text": " and you would still be able to have the benefit of interactivity with the example.", "speaker": "SPEAKER_04"}, {"start": 443.85, "end": 449.299, "text": "But with this sort of style of thing, you would be able to directly play around with the code.", "speaker": "SPEAKER_04"}, {"start": 449.98, "end": 462.24, "text": "And then for someone who's maybe a little bit more advanced, someone who wants to build stuff, there is actually, if I go back here into the main repo,", "speaker": "SPEAKER_04"}, {"start": 462.794, "end": 471.505, "text": " Sanjeev himself has built out all of these examples in such a way that there is abstract, there's an abstract specification of what an agent is, right?", "speaker": "SPEAKER_04"}, {"start": 471.626, "end": 476.492, "text": "So there's a class that says, here's an agent, and it's got a bunch of things it can do.", "speaker": "SPEAKER_04"}, {"start": 476.572, "end": 479.916, "text": "It can learn, and it can infer, blah, blah, blah.", "speaker": "SPEAKER_04"}, {"start": 480.918, "end": 483.802, "text": "Daniel's actually done a very similar thing in the existing repo.", "speaker": "SPEAKER_04"}, {"start": 485.163, "end": 490.27, "text": "And, oh, you know, okay, there's a class for an environment, and you could hook up an agent to an environment, you know.", "speaker": "SPEAKER_04"}, {"start": 490.992, "end": 492.174, "text": " This is empty just now.", "speaker": "SPEAKER_04"}, {"start": 492.314, "end": 494.138, "text": "Yeah, so there's a few things missing from here.", "speaker": "SPEAKER_04"}, {"start": 495.12, "end": 499.307, "text": "And that's a very yet more abstract way of doing things.", "speaker": "SPEAKER_04"}, {"start": 499.347, "end": 501.832, "text": "And typically what you do is you work with these source files.", "speaker": "SPEAKER_04"}, {"start": 502.193, "end": 504.797, "text": "You don't work with notebooks anymore.", "speaker": "SPEAKER_04"}, {"start": 505.419, "end": 510.728, "text": "If I just look at the repo here, go to Active Inference.", "speaker": "SPEAKER_04"}, {"start": 510.748, "end": 514.095, "text": "I guess it would be good to look at, what would it be good to look at?", "speaker": "SPEAKER_04"}, {"start": 514.155, "end": 515.397, "text": "Maybe go to Core.", "speaker": "SPEAKER_04"}, {"start": 516.372, "end": 519.477, "text": " There's a lot of stuff here from Daniel.", "speaker": "SPEAKER_04"}, {"start": 519.577, "end": 523.883, "text": "We haven't looked at POMDPs yet, but I just want to get the idea.", "speaker": "SPEAKER_04"}, {"start": 525.586, "end": 528.931, "text": "This is a style of problem, POMDP.", "speaker": "SPEAKER_04"}, {"start": 528.951, "end": 529.872, "text": "We're going to look at that later.", "speaker": "SPEAKER_04"}, {"start": 530.794, "end": 534.639, "text": "There's all these different methods for talking about a POMDP.", "speaker": "SPEAKER_04"}, {"start": 535.881, "end": 538.545, "text": "You would have the ability to wire this up very much like a gym.", "speaker": "SPEAKER_04"}, {"start": 538.786, "end": 544.354, "text": "Some people here are probably very familiar with gym.", "speaker": "SPEAKER_04"}, {"start": 544.823, "end": 560.723, "text": " This is a well-known, go to gymnasium website, you know, set of environments where you can say, all right, maybe I've got a lunar lander, you know, and I can control this lunar lander and at certain times I can get rewards and observations and I can progress in a state space.", "speaker": "SPEAKER_04"}, {"start": 561.31, "end": 562.131, "text": " very much like this here.", "speaker": "SPEAKER_04"}, {"start": 562.171, "end": 575.43, "text": "So this kind of thing where you can hook up an abstract agent to an abstract problem and do active inference, that kind of thing would be very, very possible to do in the sort of source code based version of things.", "speaker": "SPEAKER_04"}, {"start": 575.711, "end": 580.738, "text": "And that's very similar to what Daniel showed on Tuesday.", "speaker": "SPEAKER_04"}, {"start": 581.199, "end": 586.667, "text": "So I think it's very important to be able to service this kind of spectrum of interest.", "speaker": "SPEAKER_04"}, {"start": 587.147, "end": 588.97, "text": "Obviously, some people have more time than others.", "speaker": "SPEAKER_04"}, {"start": 589.03, "end": 590.372, "text": "Some people have more interest than others.", "speaker": "SPEAKER_04"}, {"start": 590.926, "end": 596.714, "text": " Um, but, uh, you know, there's active influence itself is so heterogeneous in terms of what it covers.", "speaker": "SPEAKER_04"}, {"start": 596.734, "end": 603.984, "text": "We've got people from psychology, people from non-equilibrium statistical physics, and they all kind of speak different languages and they all have different backgrounds.", "speaker": "SPEAKER_04"}, {"start": 604.004, "end": 608.089, "text": "So it's not a matter of, you know, Oh, you don't know enough code really.", "speaker": "SPEAKER_04"}, {"start": 608.109, "end": 611.253, "text": "You should be spending 10 hours a day to learn code, blah, blah, blah.", "speaker": "SPEAKER_04"}, {"start": 612.114, "end": 615.499, "text": "Um, but we do have to be mindful of servicing people's, um,", "speaker": "SPEAKER_04"}, {"start": 616.087, "end": 620.215, "text": " degree of prior ability.", "speaker": "SPEAKER_04"}, {"start": 620.275, "end": 622.419, "text": "So that's kind of what I'm thinking going forward.", "speaker": "SPEAKER_04"}, {"start": 623.06, "end": 632.678, "text": "We're going to hopefully build out initially some of these notebooks here that you're seeing, the interactive style of things.", "speaker": "SPEAKER_04"}, {"start": 632.658, "end": 635.702, "text": " So the fundamental example of Bayesian inference, this is really, really important.", "speaker": "SPEAKER_04"}, {"start": 637.003, "end": 638.185, "text": "So that's very central.", "speaker": "SPEAKER_04"}, {"start": 638.946, "end": 641.388, "text": "And then other things like gradient descent.", "speaker": "SPEAKER_04"}, {"start": 642.269, "end": 647.436, "text": "This one here is excellent, but we maybe would like a visualization of the actual gradient landscape and that kind of thing.", "speaker": "SPEAKER_04"}, {"start": 647.796, "end": 650.72, "text": "That would be another very canonical implementation.", "speaker": "SPEAKER_04"}, {"start": 651.501, "end": 654.865, "text": "Maybe, you know, expectation maximization and then things like variational free energy.", "speaker": "SPEAKER_04"}, {"start": 655.846, "end": 657.788, "text": "So that's kind of what we're thinking going forward.", "speaker": "SPEAKER_04"}, {"start": 658.595, "end": 667.625, "text": " I personally would be very interested in people's thoughts as to what they would like in terms of the examples that they want to see covered.", "speaker": "SPEAKER_04"}, {"start": 667.645, "end": 677.396, "text": "Obviously there's lots of examples in the book and eventually we'd like to cover them all, but it'd be good to cover the sort of canonical core of the examples so far.", "speaker": "SPEAKER_04"}, {"start": 678.197, "end": 687.707, "text": "So I'm interested in people's thoughts with respect to, you know, A, the level of, or if that sort of makes sense, the kind of explorer and,", "speaker": "SPEAKER_04"}, {"start": 688.531, "end": 695.859, "text": " well, you know, experimenter and then maybe builder levels of participation and experimentation.", "speaker": "SPEAKER_04"}, {"start": 696.86, "end": 700.524, "text": "To that end, I myself have been sort of negligent.", "speaker": "SPEAKER_04"}, {"start": 701.545, "end": 706.371, "text": "We do actually have a Discord chat here for the textbook group.", "speaker": "SPEAKER_04"}, {"start": 706.391, "end": 714.299, "text": "This is actually, as Andrew said earlier, the chat that has been used for previous textbook sessions on the PAR et al textbook.", "speaker": "SPEAKER_04"}, {"start": 716.962, "end": 717.643, "text": "So this is the", "speaker": "SPEAKER_04"}, {"start": 720.982, "end": 726.069, "text": " Although Sanjeev's recent one really does expand on a lot of stuff in detail.", "speaker": "SPEAKER_04"}, {"start": 727.451, "end": 730.475, "text": "Previous textbook sessions have been going on and they've been happening here in this chat.", "speaker": "SPEAKER_04"}, {"start": 731.296, "end": 735.842, "text": "Now, this is a space for us to be able to talk about things with respect to the book.", "speaker": "SPEAKER_04"}, {"start": 736.363, "end": 740.188, "text": "Myself and Andrew and Daniel are on here pretty regularly.", "speaker": "SPEAKER_04"}, {"start": 740.168, "end": 744.237, "text": " So it'd be an excellent way to stay in asynchronous contact with us if you're interested in that.", "speaker": "SPEAKER_04"}, {"start": 744.999, "end": 747.384, "text": "Obviously, our emails are another affordance for that.", "speaker": "SPEAKER_04"}, {"start": 747.685, "end": 752.095, "text": "So if you don't have a Discord account, you'd have to make one, obviously.", "speaker": "SPEAKER_04"}, {"start": 752.135, "end": 755.342, "text": "But for those who do have them, you can immediately hop on over here.", "speaker": "SPEAKER_04"}, {"start": 755.795, "end": 760.481, "text": " And that would be a great way to share your thoughts on things related to examples.", "speaker": "SPEAKER_04"}, {"start": 761.502, "end": 762.003, "text": "So that's that.", "speaker": "SPEAKER_04"}, {"start": 762.023, "end": 767.229, "text": "I don't know if there's any questions or anything people would like to ask with respect to the examples.", "speaker": "SPEAKER_04"}, {"start": 767.449, "end": 770.714, "text": "We are thinking about how to build them and so on.", "speaker": "SPEAKER_04"}, {"start": 771.414, "end": 779.264, "text": "But I think now we can perhaps take any questions that are related to part one.", "speaker": "SPEAKER_04"}, {"start": 779.284, "end": 783.169, "text": "I'd be very interested in assisting where I can there.", "speaker": "SPEAKER_04"}, {"start": 784.735, "end": 786.698, "text": " The floor is open for live questions.", "speaker": "SPEAKER_04"}, {"start": 786.718, "end": 793.771, "text": "If there aren't live questions that people don't want to ask them, we can go over some ones in the chat as well, in the page.", "speaker": "SPEAKER_04"}, {"start": 795.634, "end": 797.016, "text": "Just taking a moment to read the chat here.", "speaker": "SPEAKER_04"}, {"start": 797.076, "end": 800.182, "text": "Andrew, yes.", "speaker": "SPEAKER_04"}, {"start": 803.568, "end": 804.87, "text": "Yeah, just in lieu of", "speaker": "SPEAKER_03"}, {"start": 805.053, "end": 807.656, "text": " They're not being any other immediate questions right now.", "speaker": "SPEAKER_03"}, {"start": 808.617, "end": 813.842, "text": "This is a bit more of a brainstorming point that we don't have to spend too much time on.", "speaker": "SPEAKER_03"}, {"start": 814.523, "end": 824.093, "text": "But I was thinking for the code examples, you know, given that right now we're in, we're going to start part two.", "speaker": "SPEAKER_03"}, {"start": 824.213, "end": 825.474, "text": "So we've not got there yet.", "speaker": "SPEAKER_03"}, {"start": 826.355, "end": 829.879, "text": "We ended chapter five with predictive coding.", "speaker": "SPEAKER_03"}, {"start": 830.332, "end": 832.575, "text": " Um, 1 direction we could go.", "speaker": "SPEAKER_03"}, {"start": 832.595, "end": 835.92, "text": "Is that we could have a code example.", "speaker": "SPEAKER_03"}, {"start": 835.94, "end": 841.629, "text": "That sort of consolidates a lot of the things that we saw in chapters 2 through 5.", "speaker": "SPEAKER_03"}, {"start": 841.649, "end": 848.058, "text": "Not everything, obviously, there's some methods that were given that, like, kind of contradict each other.", "speaker": "SPEAKER_03"}, {"start": 848.098, "end": 850.301, "text": "They're sort of in their own respect paradigm.", "speaker": "SPEAKER_03"}, {"start": 851.023, "end": 853.386, "text": "Um, but still, um.", "speaker": "SPEAKER_03"}, {"start": 853.366, "end": 860.457, "text": " You know, in almost every chapter we see, it's like, well, here's the univariate case and then here's the multivariate case.", "speaker": "SPEAKER_03"}, {"start": 861.258, "end": 874.058, "text": "A coding example could just be geared around the multivariate case and then there's some option to only have one hidden state and one observation modality if you did want it to be univariate or something.", "speaker": "SPEAKER_03"}, {"start": 874.038, "end": 881.427, "text": " But, yeah, I'm just, you know, these are the kind of questions that we're sort of thinking about with regards to developing any code further.", "speaker": "SPEAKER_03"}, {"start": 881.487, "end": 894.563, "text": "Obviously, we'd want to exploit sort of Sanjeev's code that he shared with us to make sure that we're really kind of serving what he was going for as far as developing code goes.", "speaker": "SPEAKER_03"}, {"start": 895.384, "end": 900.45, "text": "But, yeah, like we could potentially do like a full, you know,", "speaker": "SPEAKER_03"}, {"start": 900.43, "end": 905.434, "text": " Here's code for a hierarchical predictive coding model.", "speaker": "SPEAKER_03"}, {"start": 905.454, "end": 908.897, "text": "It has all of your precision weighting and updates.", "speaker": "SPEAKER_03"}, {"start": 908.997, "end": 910.259, "text": "It has parameter learning.", "speaker": "SPEAKER_03"}, {"start": 910.279, "end": 912.36, "text": "It has all of the things.", "speaker": "SPEAKER_03"}, {"start": 913.001, "end": 929.155, "text": "And then having some kinds of interactive options for users that maybe initially lets you specify how many hidden states are there and all the rest.", "speaker": "SPEAKER_03"}, {"start": 929.557, "end": 935.425, "text": " It all has to do with how much time everyone, you know, kind of has available and how much we're able to build this out.", "speaker": "SPEAKER_03"}, {"start": 935.565, "end": 945.139, "text": "And then this is also why I've, like, in the past repeatedly mentioned, like, for anyone who wants to see something in particular, that could be really helpful.", "speaker": "SPEAKER_03"}, {"start": 945.64, "end": 954.933, "text": "Because given we only have so much time, it would be great to make sure we're spending our time on things that learners would most like to sort of focus upon.", "speaker": "SPEAKER_03"}, {"start": 954.973, "end": 956.915, "text": "But, yeah, I'll stop there.", "speaker": "SPEAKER_03"}, {"start": 957.196, "end": 958.838, "text": "I noticed Kobus has his hand up.", "speaker": "SPEAKER_03"}, {"start": 959.122, "end": 961.625, "text": " I could go for it.", "speaker": "SPEAKER_01"}, {"start": 961.865, "end": 965.309, "text": "Yeah, Andrew, that sounds like an exciting idea.", "speaker": "SPEAKER_01"}, {"start": 965.39, "end": 979.026, "text": "I like the idea of having a generic code base where you can say, I want to have so many components in my state vector, so many components in my observation vector.", "speaker": "SPEAKER_01"}, {"start": 979.046, "end": 981.389, "text": "I want to have so many layers.", "speaker": "SPEAKER_01"}, {"start": 981.749, "end": 984.433, "text": "But the code works for every possible case.", "speaker": "SPEAKER_01"}, {"start": 984.493, "end": 986.916, "text": "Now, I realize there's some downside to that.", "speaker": "SPEAKER_01"}, {"start": 986.996, "end": 988.858, "text": "It makes it more complicated.", "speaker": "SPEAKER_01"}, {"start": 989.294, "end": 990.876, "text": " less easy to understand.", "speaker": "SPEAKER_01"}, {"start": 990.977, "end": 994.542, "text": "But I think there's value in a generic code base.", "speaker": "SPEAKER_01"}, {"start": 994.562, "end": 1007.261, "text": "And maybe this can even form the beginnings of what you mentioned previously, that you want to put together a library that works for the continuous case.", "speaker": "SPEAKER_01"}, {"start": 1008.162, "end": 1016.935, "text": "And eventually, who knows, we can make it even more generic to work for both the continuous formulation as well as the discrete formulation.", "speaker": "SPEAKER_01"}, {"start": 1017.055, "end": 1018.137, "text": "So I like that idea.", "speaker": "SPEAKER_01"}, {"start": 1020.007, "end": 1021.349, "text": " No, I completely agree.", "speaker": "SPEAKER_04"}, {"start": 1021.369, "end": 1024.874, "text": "I think that is a sensible way to go long term.", "speaker": "SPEAKER_04"}, {"start": 1026.476, "end": 1026.576, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 1026.596, "end": 1027.057, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 1027.358, "end": 1029.12, "text": "Thanks for the feedback, Kobus.", "speaker": "SPEAKER_03"}, {"start": 1029.18, "end": 1043.34, "text": "Because, yeah, it's like to some degree, it would be great to make sure that we're not sort of starting over with the code for each, you know, chapter related example, because a lot of this is like we're building up over time.", "speaker": "SPEAKER_03"}, {"start": 1043.401, "end": 1047.246, "text": "Like if we were to sort of do that, that more generalized approach,", "speaker": "SPEAKER_03"}, {"start": 1047.547, "end": 1051.135, "text": " I hesitate to use the word generalized because we're seeing it so much in the textbook.", "speaker": "SPEAKER_03"}, {"start": 1051.737, "end": 1059.615, "text": "But yeah, but a more general library, like that hierarchical predictive coding model set up that I mentioned,", "speaker": "SPEAKER_03"}, {"start": 1060.135, "end": 1074.934, "text": " You know, in Chapters 6 and 7 and 8, we're going to be looking at generalized filtering and then active generalized filtering, which also have an immense amount of overlap with the hierarchical predictive coding models.", "speaker": "SPEAKER_03"}, {"start": 1074.974, "end": 1079.52, "text": "We're going to start seeing similar diagrams of multiple layers, all the rest, and precision weighting.", "speaker": "SPEAKER_03"}, {"start": 1079.64, "end": 1087.83, "text": "So then, you know, with the hierarchical predictive coding, we could then continue building off of that as a general.", "speaker": "SPEAKER_03"}, {"start": 1088.165, "end": 1093.0, "text": " Library yeah to to like, further serve, being able to move further through the chapters.", "speaker": "SPEAKER_03"}, {"start": 1093.06, "end": 1094.007, "text": "No.", "speaker": "SPEAKER_03"}, {"start": 1094.392, "end": 1101.479, "text": " know, starting all over, maybe having a more consistent singular API for users to be able to call these things.", "speaker": "SPEAKER_03"}, {"start": 1101.559, "end": 1117.775, "text": "Because, yeah, I mean, that's sort of, that's sort of where Active Inference is at on the programming side, more generally, everyone's seeking some kind of, in the same way of like what Keras, the library Keras did for deep learning, right?", "speaker": "SPEAKER_03"}, {"start": 1118.316, "end": 1120.838, "text": "It'd be nice to have sort of like a universal", "speaker": "SPEAKER_03"}, {"start": 1121.155, "end": 1123.978, "text": " library that kind of lets us do anything we'd like.", "speaker": "SPEAKER_03"}, {"start": 1124.678, "end": 1137.591, "text": "Other questions that come up are like, yeah, as far as continuous and discrete state space problems go in a singular library, it's like, well, PIMDP is currently fully open source.", "speaker": "SPEAKER_03"}, {"start": 1138.732, "end": 1149.002, "text": "I mean, we even know some of the developers, so it might be nice to, you know, we could potentially integrate with PIMTP or we could do something else entirely.", "speaker": "SPEAKER_03"}, {"start": 1149.421, "end": 1155.831, "text": " I've been working with others who are actually trying to do the discrete state space stuff from scratch.", "speaker": "SPEAKER_03"}, {"start": 1156.692, "end": 1161.239, "text": "PyMDP does have its limitations, so maybe it is worthwhile to think that through.", "speaker": "SPEAKER_03"}, {"start": 1162.281, "end": 1164.304, "text": "But yeah, everything sounds great.", "speaker": "SPEAKER_03"}, {"start": 1165.747, "end": 1177.605, "text": "It'd be nice to basically supply users with all of the fundamental sort of building blocks for anything related to active inference and", "speaker": "SPEAKER_03"}, {"start": 1178.311, "end": 1192.946, "text": " All sorts of other things like hybrid models, where you have, like, a continuous model at a lower level and then discrete state space model at a higher level with, like, different kinds of link functions between, like, there's been work in active inference on those things.", "speaker": "SPEAKER_03"}, {"start": 1193.006, "end": 1193.788, "text": "They exist.", "speaker": "SPEAKER_03"}, {"start": 1194.14, "end": 1195.782, "text": " but there's not enough of them.", "speaker": "SPEAKER_03"}, {"start": 1196.522, "end": 1208.194, "text": "And they could be really fascinating when thinking about like modeling an agent who takes in a lot of continuous value data, but then has to make discrete decisions and do planning and things like that.", "speaker": "SPEAKER_03"}, {"start": 1208.214, "end": 1214.8, "text": "I mean, it's a very interesting case and even aligns really well with empirical reality in some ways.", "speaker": "SPEAKER_03"}, {"start": 1215.401, "end": 1216.442, "text": "So there needs to be more work.", "speaker": "SPEAKER_03"}, {"start": 1216.782, "end": 1222.287, "text": "Being able to have a repo where people can give that a shot as they please would be fantastic.", "speaker": "SPEAKER_03"}, {"start": 1222.407, "end": 1222.948, "text": "Yeah, Kovac.", "speaker": "SPEAKER_03"}, {"start": 1223.975, "end": 1243.817, "text": " Yeah, and my personal preference for this code base should be to lean over to the side of making it very understandable, almost aligned with the mathematics precisely rather than making it performant.", "speaker": "SPEAKER_01"}, {"start": 1245.418, "end": 1253.307, "text": "You know, there's a saying that says it's easier to make a correct program fast than a fast program correct.", "speaker": "SPEAKER_01"}, {"start": 1253.793, "end": 1277.751, "text": " So, starting off, I think there's virtue in, you know, leaning to the side of, you know, understanding it as much as possible, and even the mathematics and that kind of thing, and forget about making it really fast, like, you know, the Jack's approach and things like that.", "speaker": "SPEAKER_01"}, {"start": 1278.051, "end": 1279.834, "text": "Maybe that can follow...", "speaker": "SPEAKER_01"}, {"start": 1280.0, "end": 1307.053, "text": " in the future as a as a different version or a different development path so that would be my personal personal preference yeah sanjeev himself actually in his notebooks has taken great pains to implement the code in such a way that it is essentially one-to-one with the mathematics yeah you know eschewing all issues of efficiency and such like because that's not the point you know the point is pedagogical in terms of demonstrating", "speaker": "SPEAKER_04"}, {"start": 1307.523, "end": 1309.508, "text": " The idea is, I completely agree.", "speaker": "SPEAKER_04"}, {"start": 1309.849, "end": 1310.109, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 1311.332, "end": 1311.633, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 1312.234, "end": 1312.796, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 1312.816, "end": 1320.013, "text": "It's, I mean, well, I think ideally what should happen for the repo is, you know,", "speaker": "SPEAKER_03"}, {"start": 1320.888, "end": 1326.915, "text": " I mean, first, we're trying to keep it tethered to Sanjeev's textbook for sure.", "speaker": "SPEAKER_03"}, {"start": 1327.075, "end": 1334.444, "text": "So, like, making sure that the repo, like, more or less fully illustrates the mathematics in the textbook.", "speaker": "SPEAKER_03"}, {"start": 1335.204, "end": 1347.098, "text": "Because, yeah, I agree, like, active inference is not purely about achieving, like, you know, high scale, full performance at the expense of sort of, you know,", "speaker": "SPEAKER_03"}, {"start": 1348.411, "end": 1350.754, "text": " mathematical integrity, we'll say.", "speaker": "SPEAKER_03"}, {"start": 1351.715, "end": 1356.622, "text": "You know, no cheap shortcuts, unless it's part of what you're going for.", "speaker": "SPEAKER_03"}, {"start": 1356.642, "end": 1360.767, "text": "But I think it should be focused upon, you know, contributions to research.", "speaker": "SPEAKER_03"}, {"start": 1361.428, "end": 1370.32, "text": "Of course, if we build out all of these things in a repo, then anyone who did want to start optimizing or using JAX or things like that would be able to then build on that further.", "speaker": "SPEAKER_03"}, {"start": 1370.76, "end": 1373.524, "text": "And that just becomes another section of the repository.", "speaker": "SPEAKER_03"}, {"start": 1373.808, "end": 1379.096, "text": " without sort of harming or undoing previous work that's been done.", "speaker": "SPEAKER_03"}, {"start": 1379.156, "end": 1380.018, "text": "But yeah, much agreed.", "speaker": "SPEAKER_03"}, {"start": 1380.278, "end": 1394.5, "text": "Because with, I mean, with JAX alone, because even I've like messed with JAX recently for some of my work that I've been doing, but some of the things I've been doing involve like structure learning, where you're like potentially adding or removing", "speaker": "SPEAKER_03"}, {"start": 1394.48, "end": 1400.632, "text": " different hidden state factors or the discrete levels of the hidden state factors.", "speaker": "SPEAKER_03"}, {"start": 1400.973, "end": 1408.367, "text": "And so what happens is that you end up with these matrices that are changing in dimensionality, literally from iteration to iteration.", "speaker": "SPEAKER_03"}, {"start": 1408.407, "end": 1410.872, "text": "And so Jax doesn't really like that, right?", "speaker": "SPEAKER_03"}, {"start": 1410.892, "end": 1412.235, "text": "Because you're, you, you,", "speaker": "SPEAKER_03"}, {"start": 1412.215, "end": 1419.844, "text": " Ideally, with JAX, you would have a lot of your tensor shapes and the rest like fixed in advance so that you could do all the just-in-time compilation.", "speaker": "SPEAKER_03"}, {"start": 1419.864, "end": 1433.98, "text": "But if the whole structure is changing bit by bit, while structure learning is a very interesting and kind of, you know, topical thing that a lot of people are working on in active inference, it means that JAX just isn't really amenable to it.", "speaker": "SPEAKER_03"}, {"start": 1434.32, "end": 1438.645, "text": "And I've even gotten some insight from Sanjeev on that point because he's been working on similar things.", "speaker": "SPEAKER_03"}, {"start": 1438.625, "end": 1444.437, "text": " So basically just recognizing like, yeah, there will be places that, you know, JAX won't make sense.", "speaker": "SPEAKER_03"}, {"start": 1444.958, "end": 1448.705, "text": "But maybe there is an alternative method that could be used.", "speaker": "SPEAKER_03"}, {"start": 1448.725, "end": 1460.869, "text": "So, yes, making sure we kind of meet the mathematics with fidelity prior to any kind of, you know, fun optimization tricks would be the best priority, I think.", "speaker": "SPEAKER_03"}, {"start": 1464.038, "end": 1465.52, "text": " That's always how you want to do it.", "speaker": "SPEAKER_04"}, {"start": 1465.601, "end": 1469.787, "text": "You want to understand what you're doing first and then move to making it efficient.", "speaker": "SPEAKER_04"}, {"start": 1469.848, "end": 1470.388, "text": "That's for sure.", "speaker": "SPEAKER_04"}, {"start": 1472.712, "end": 1473.033, "text": "All right.", "speaker": "SPEAKER_04"}, {"start": 1473.073, "end": 1476.979, "text": "There was a pre-written question that I think would be interesting to go through.", "speaker": "SPEAKER_04"}, {"start": 1477.08, "end": 1483.49, "text": "But if there are live questions or if you want to put them in the chat, I'll obviously give them preference.", "speaker": "SPEAKER_04"}, {"start": 1484.452, "end": 1486.475, "text": "Have I turned my video off?", "speaker": "SPEAKER_04"}, {"start": 1486.495, "end": 1487.076, "text": "There we go.", "speaker": "SPEAKER_04"}, {"start": 1487.309, "end": 1489.892, "text": " So if there are live questions, please do fire away.", "speaker": "SPEAKER_04"}, {"start": 1489.952, "end": 1496.878, "text": "But other than that, I will maybe share this question here because I think it's a relevant one.", "speaker": "SPEAKER_04"}, {"start": 1498.82, "end": 1500.262, "text": "So this comes from chapter three.", "speaker": "SPEAKER_04"}, {"start": 1500.302, "end": 1503.084, "text": "Hopefully people can see my screen.", "speaker": "SPEAKER_04"}, {"start": 1503.104, "end": 1503.585, "text": "Come down here.", "speaker": "SPEAKER_04"}, {"start": 1504.646, "end": 1505.447, "text": "Zoom in a bit more.", "speaker": "SPEAKER_04"}, {"start": 1508.53, "end": 1509.831, "text": "Chapter four, chapter three.", "speaker": "SPEAKER_04"}, {"start": 1510.832, "end": 1512.173, "text": "So this is on page 72.", "speaker": "SPEAKER_04"}, {"start": 1512.233, "end": 1513.294, "text": "So this question here.", "speaker": "SPEAKER_04"}, {"start": 1514.275, "end": 1515.156, "text": "Oh, my lights went off.", "speaker": "SPEAKER_04"}, {"start": 1517.937, "end": 1524.286, "text": " This is on page 72 from chapter three with respect to the expectation maximization algorithm.", "speaker": "SPEAKER_04"}, {"start": 1524.306, "end": 1526.769, "text": "And this was a big linchpin in part one.", "speaker": "SPEAKER_04"}, {"start": 1527.33, "end": 1537.764, "text": "Really it tied together everything we saw in chapters one and two and most of three with respect to the question of learning and inference.", "speaker": "SPEAKER_04"}, {"start": 1537.784, "end": 1539.627, "text": "How do we do both of these things at the same time?", "speaker": "SPEAKER_04"}, {"start": 1540.568, "end": 1541.63, "text": "Because we typically have to.", "speaker": "SPEAKER_04"}, {"start": 1542.371, "end": 1545.615, "text": "And the expectation maximization, this is how we do it basically.", "speaker": "SPEAKER_04"}, {"start": 1546.574, "end": 1554.827, "text": " This, you know, page 72, so equation 3.44A and B, those are the two steps of the EM algorithm.", "speaker": "SPEAKER_04"}, {"start": 1555.208, "end": 1565.264, "text": "The question is, you know, with respect to the expectation step, when Sanjeev says fill in the unknown parameters with their means, expected value, can someone break it down in this sentence?", "speaker": "SPEAKER_04"}, {"start": 1565.565, "end": 1571.755, "text": "I cannot relate it to equation 3.44A where an expected value of the log likelihood is calculated.", "speaker": "SPEAKER_04"}, {"start": 1572.798, "end": 1581.107, "text": " So if I just go to that equation so that people see it, 3.44a, this is, so we have 3.44a and b.", "speaker": "SPEAKER_04"}, {"start": 1583.95, "end": 1588.775, "text": "So this is the expectation step of the expectation maximization algorithm.", "speaker": "SPEAKER_04"}, {"start": 1589.415, "end": 1594.16, "text": "So I guess the question in some sense is like, what is an expectation at all?", "speaker": "SPEAKER_04"}, {"start": 1594.2, "end": 1601.708, "text": "The idea is that with the expectation maximization algorithm, we don't know what the parameters are.", "speaker": "SPEAKER_04"}, {"start": 1602.43, "end": 1632.032, "text": " so we need to take their expectation i'll define what that is in a second with respect to the current belief about the hidden state and then we can maximize our evidence for the hidden state by doing the maximization step but basically what what this is saying here what this equation is saying is we need to so like this e is an expectation we want to we want to say what we think the", "speaker": "SPEAKER_04"}, {"start": 1632.755, "end": 1645.812, "text": " The probability is of this thing in here, this log probability, where we're taking our values of x, the hidden state, from this distribution down here.", "speaker": "SPEAKER_04"}, {"start": 1646.513, "end": 1647.594, "text": "I know that's a little weird.", "speaker": "SPEAKER_04"}, {"start": 1647.634, "end": 1652.08, "text": "There's better ways of showing this visually, and that's one thing I would like to show.", "speaker": "SPEAKER_04"}, {"start": 1652.92, "end": 1657.947, "text": " But the core idea is that the parameters, they're unknown.", "speaker": "SPEAKER_04"}, {"start": 1658.007, "end": 1664.276, "text": "So we have some sort of distribution about what they could be.", "speaker": "SPEAKER_04"}, {"start": 1665.799, "end": 1670.105, "text": "And we're going to use that distribution to inform a guess about what they are.", "speaker": "SPEAKER_04"}, {"start": 1670.926, "end": 1676.975, "text": "And then we're going to use that guess in our evaluation of the log likelihood here.", "speaker": "SPEAKER_04"}, {"start": 1677.716, "end": 1682.623, "text": "So this Vi theta, this is the parameters.", "speaker": "SPEAKER_04"}, {"start": 1683.211, "end": 1687.957, "text": " So this is saying, you know, what, what do we think, what is our belief about what the parameters should be?", "speaker": "SPEAKER_04"}, {"start": 1687.977, "end": 1691.501, "text": "Uh, given our belief about them.", "speaker": "SPEAKER_04"}, {"start": 1691.541, "end": 1705.118, "text": "Now we take the expectation of the joint, uh, the general model basically under, uh, these, these, uh, the belief about the hidden state given observations.", "speaker": "SPEAKER_04"}, {"start": 1705.138, "end": 1708.342, "text": "So that's, that's the tricky bit.", "speaker": "SPEAKER_04"}, {"start": 1708.362, "end": 1709.903, "text": "This is a belief about parameters.", "speaker": "SPEAKER_04"}, {"start": 1709.964, "end": 1711.986, "text": "It's not a belief about hidden states here.", "speaker": "SPEAKER_04"}, {"start": 1712.438, "end": 1715.822, "text": " So we have to kind of average over our beliefs about hidden states.", "speaker": "SPEAKER_04"}, {"start": 1715.922, "end": 1716.983, "text": "And that's what this is.", "speaker": "SPEAKER_04"}, {"start": 1717.684, "end": 1719.006, "text": "I probably haven't explained that very well.", "speaker": "SPEAKER_04"}, {"start": 1719.066, "end": 1722.209, "text": "And there's actually a lot of nice visual intuition for this here.", "speaker": "SPEAKER_04"}, {"start": 1722.73, "end": 1728.297, "text": "But this is an important thing to get your head around, this idea of expectation.", "speaker": "SPEAKER_04"}, {"start": 1728.997, "end": 1730.379, "text": "This comes up again and again and again.", "speaker": "SPEAKER_04"}, {"start": 1730.399, "end": 1733.002, "text": "So this is an important question.", "speaker": "SPEAKER_04"}, {"start": 1733.523, "end": 1738.108, "text": "I'll go through and I will actually give an answer.", "speaker": "SPEAKER_04"}, {"start": 1738.426, "end": 1742.632, "text": " when there's more time, not in the group itself, but this question is important.", "speaker": "SPEAKER_04"}, {"start": 1742.652, "end": 1750.724, "text": "So it's basically saying, I don't know what the parameter should be, but I can use my belief about the hidden states because I have uncertainty about them.", "speaker": "SPEAKER_04"}, {"start": 1750.744, "end": 1762.12, "text": "I have this distribution of the hidden states to predict the hidden states most likely value in such a way that I can maintain a belief about the parameters.", "speaker": "SPEAKER_04"}, {"start": 1762.881, "end": 1765.345, "text": "So it's kind of a, it's a guess, the expectation,", "speaker": "SPEAKER_04"}, {"start": 1772.935, "end": 1774.717, "text": " I'll just stop sharing temporarily.", "speaker": "SPEAKER_04"}, {"start": 1774.757, "end": 1780.382, "text": "There are a lot of other, well, there are other questions as well.", "speaker": "SPEAKER_04"}, {"start": 1781.683, "end": 1785.407, "text": "Some of them are quite specific, so I don't necessarily want to go over them here.", "speaker": "SPEAKER_04"}, {"start": 1786.928, "end": 1790.832, "text": "Some questions related to general things would be good.", "speaker": "SPEAKER_04"}, {"start": 1790.912, "end": 1792.353, "text": "Andrew, did you have a comment?", "speaker": "SPEAKER_04"}, {"start": 1794.696, "end": 1797.638, "text": "I think Cobus had his hand up first.", "speaker": "SPEAKER_03"}, {"start": 1797.979, "end": 1798.96, "text": "So if Cobus, you want to.", "speaker": "SPEAKER_03"}, {"start": 1799.0, "end": 1799.38, "text": "Oh, I'm sorry.", "speaker": "SPEAKER_03"}, {"start": 1799.4, "end": 1801.542, "text": "I thought that was from before.", "speaker": "SPEAKER_04"}, {"start": 1801.859, "end": 1818.224, "text": " Just, you know, on the previous question about that expectation, I think what might help is to, in parallel with this expression 344a, just put down a very basic expectation expression.", "speaker": "SPEAKER_01"}, {"start": 1818.265, "end": 1826.297, "text": "For example, the expectation of, let's say, f of x.", "speaker": "SPEAKER_01"}, {"start": 1827.121, "end": 1836.074, "text": " And then let's make that f of x maybe 2x.", "speaker": "SPEAKER_01"}, {"start": 1836.094, "end": 1850.614, "text": "And then the subscript of the expectation operator is simply the distribution, the density from which that function is associated.", "speaker": "SPEAKER_01"}, {"start": 1851.235, "end": 1860.925, "text": " So here it looks very intimidating, but if you come up with a trivial example, you start to see what the role of each of these symbols are.", "speaker": "SPEAKER_01"}, {"start": 1862.12, "end": 1863.502, "text": " Yeah, no, I completely agree.", "speaker": "SPEAKER_04"}, {"start": 1863.562, "end": 1867.887, "text": "It is a bit of a hairy thing to get your head around initially.", "speaker": "SPEAKER_04"}, {"start": 1868.408, "end": 1879.722, "text": "And what's confusing is usually the expectation operator does not have a subscript in a trivial case, but all of a sudden we encounter them with subscripts.", "speaker": "SPEAKER_01"}, {"start": 1879.742, "end": 1882.145, "text": "And I think that's part of the confusion maybe.", "speaker": "SPEAKER_01"}, {"start": 1883.207, "end": 1885.89, "text": "Yeah, I mean, so appendix, what is it here?", "speaker": "SPEAKER_04"}, {"start": 1885.91, "end": 1887.372, "text": "Appendix C, I think.", "speaker": "SPEAKER_04"}, {"start": 1887.793, "end": 1888.353, "text": "If I go down here.", "speaker": "SPEAKER_04"}, {"start": 1888.594, "end": 1890.456, "text": "Yeah, mathematical fundamentals.", "speaker": "SPEAKER_04"}, {"start": 1891.027, "end": 1897.634, "text": " that Appendix C takes you through the notation for things like subscripts, and they're used everywhere in the book.", "speaker": "SPEAKER_04"}, {"start": 1898.295, "end": 1913.312, "text": "So this would be a good opportunity if you're unfamiliar with some of this subscript notation to maybe go over, sprees over Appendix C again with respect to that notation, because that will continue to show up throughout the rest of the book for sure, yeah.", "speaker": "SPEAKER_04"}, {"start": 1915.634, "end": 1916.355, "text": "No, thanks for that, Cobus.", "speaker": "SPEAKER_04"}, {"start": 1918.658, "end": 1920.78, "text": "I guess I'd also add,", "speaker": "SPEAKER_03"}, {"start": 1922.178, "end": 1945.452, "text": " I'm not sure who asked the question that you were looking at in the coda, and maybe they do understand this, but I guess it can bear repeating of like, in chapter three, we have an algorithm that's given to us, just like we're given these other nice sort of algorithm charts.", "speaker": "SPEAKER_03"}, {"start": 1946.039, "end": 1967.406, "text": " in other chapters and so with algorithm three in the textbook which is expectation for linear factor analysis uh it just it bears remembering that um or excuse it's worthwhile to remember that this this expectation maximization process is like iterative and we're not literally", "speaker": "SPEAKER_03"}, {"start": 1967.842, "end": 1975.296, "text": " at precisely the same moment in the same computation, updating our hidden state estimates and parameters.", "speaker": "SPEAKER_03"}, {"start": 1975.416, "end": 1984.193, "text": "It's like just the general notion of we're updating hidden states, then we're updating parameters, then hidden states, then parameters.", "speaker": "SPEAKER_03"}, {"start": 1984.994, "end": 1990.204, "text": "And in both cases, whenever we move up to variational inference and we're looking at free energy, then", "speaker": "SPEAKER_03"}, {"start": 1990.437, "end": 2010.64, "text": " sort of the golden rule is that ideally you'd be able to update anything, whether it be a hidden state estimate or a particular parameter or otherwise in this iterative fashion with respect to sort of a gradient on free energy with respect to the thing that you're updating.", "speaker": "SPEAKER_03"}, {"start": 2010.62, "end": 2035.53, "text": " so um I I don't mean that as like an answer to the question uh but it's just like because I think that it it gets a little mixed up and like whenever you actually implement these things and computational code or otherwise it's like well yeah we have all of these different components in a model and we're talking about updating these different things but like how does that happen it's like well it's very procedural it's very iterative and it's very", "speaker": "SPEAKER_03"}, {"start": 2035.831, "end": 2065.551, "text": " we update this we update that we update this we update that so um yeah just adding that conversation now yeah fair enough that's that's for sure all right well uh yes I I have been going slowly through the uh the coder and answering questions putting my attempted answers not to uh denote well not to to connote the idea that that's final but I have been going through I do plan on um", "speaker": "SPEAKER_04"}, {"start": 2065.885, "end": 2066.626, "text": " finishing them all.", "speaker": "SPEAKER_04"}, {"start": 2066.686, "end": 2070.671, "text": "If I come to here, share my screen again.", "speaker": "SPEAKER_04"}, {"start": 2070.691, "end": 2091.399, "text": "One thing that's currently missing, as some of you might note, is in chapter five, where we've left off, typically I have this overall set of notes, but the more detailed notes for sections 5.1 and 5.2 are still not there.", "speaker": "SPEAKER_04"}, {"start": 2091.999, "end": 2092.6, "text": "I will put them in.", "speaker": "SPEAKER_04"}, {"start": 2092.7, "end": 2094.002, "text": "I'll also, for chapter", "speaker": "SPEAKER_04"}, {"start": 2094.151, "end": 2107.536, "text": " six next week we're going to be picking up uh produce those those same things for chapter six that's quite bare just now and that will help us maybe going forward as well so", "speaker": "SPEAKER_04"}, {"start": 2112.562, "end": 2137.81, "text": " in a bit of a rhythm is this idea of the the the code that people can interact with just you know in a very friction-free manner quickly and easily without having to really interface with the the the details you know just so you can get an intuition that's very much something that i i would like to bring into the group and have people have another interface for people to to play with the ideas", "speaker": "SPEAKER_04"}, {"start": 2138.093, "end": 2153.952, "text": " Because we haven't yet, especially for things like in chapter three, where we do talk a lot about the expectation maximization algorithm, it can be quite a visual thing.", "speaker": "SPEAKER_04"}, {"start": 2153.992, "end": 2157.837, "text": "And so it can be very useful to have visualizations and such like.", "speaker": "SPEAKER_04"}, {"start": 2157.897, "end": 2159.158, "text": "And the figures do a good job.", "speaker": "SPEAKER_04"}, {"start": 2160.64, "end": 2163.023, "text": "But it would be nice to have additional material.", "speaker": "SPEAKER_04"}, {"start": 2163.043, "end": 2167.408, "text": "So that's something I'm keen on providing going forward for sure.", "speaker": "SPEAKER_04"}, {"start": 2172.265, "end": 2173.947, "text": " All right, let me just check the chat real quick.", "speaker": "SPEAKER_04"}, {"start": 2178.691, "end": 2191.743, "text": "If there are no additional questions or anything like that, or if no one's got any burning concerns, we might potentially have an early session today, unless people are interested in asking some questions.", "speaker": "SPEAKER_04"}, {"start": 2194.526, "end": 2197.108, "text": "Do remember the Discord.", "speaker": "SPEAKER_04"}, {"start": 2197.128, "end": 2197.929, "text": "So that is in the chat.", "speaker": "SPEAKER_04"}, {"start": 2198.009, "end": 2199.15, "text": "I'll just put it there again.", "speaker": "SPEAKER_04"}, {"start": 2202.235, "end": 2204.78, "text": " There's the link to the actual Discord chat.", "speaker": "SPEAKER_04"}, {"start": 2204.8, "end": 2209.57, "text": "So that would be, as I said, a good way to stay in contact with us more asynchronously.", "speaker": "SPEAKER_04"}, {"start": 2210.692, "end": 2216.824, "text": "I can put my email address in the chat as well, although I think we've actually got that shared elsewhere.", "speaker": "SPEAKER_04"}, {"start": 2217.305, "end": 2218.708, "text": "Hey, Fraser, I'll jump in.", "speaker": "SPEAKER_00"}, {"start": 2219.296, "end": 2231.71, "text": " And I'm following up on the same line of communication from the last class, which is about Andrew's chapter two demo.", "speaker": "SPEAKER_00"}, {"start": 2232.651, "end": 2238.398, "text": "And this is more sort of just a sort of a broader approach question.", "speaker": "SPEAKER_00"}, {"start": 2239.139, "end": 2244.865, "text": "But it seems to me that that demo or not the demo, but the but the", "speaker": "SPEAKER_00"}, {"start": 2246.077, "end": 2263.208, "text": " um there's an opportunity to sort of train our intuition about the relationship between the priors the likelihood and and the posteriors and especially with us casting sort of um like", "speaker": "SPEAKER_00"}, {"start": 2264.133, "end": 2270.085, "text": " anthropo, you know, like that we think, like, I think, I think you just, you called one of it the stubborn scenario.", "speaker": "SPEAKER_00"}, {"start": 2271.488, "end": 2284.774, "text": "So just sort of running through some of these intuitions about like behavior that we, you know, and also breaking down some of the misunderstandings, especially as, as,", "speaker": "SPEAKER_00"}, {"start": 2284.754, "end": 2292.013, "text": " active inference is used by engineers that may not have as deep of an experience in the social sciences.", "speaker": "SPEAKER_00"}, {"start": 2292.093, "end": 2298.51, "text": "And then the other way around, social scientists who don't have such a deep understanding of", "speaker": "SPEAKER_00"}, {"start": 2299.216, "end": 2305.184, "text": " who's driving the car, you know, like what thing is contributing to what response.", "speaker": "SPEAKER_00"}, {"start": 2305.264, "end": 2317.181, "text": "And so I just, you know, again, like I'm in two different worlds where, you know, I want to be able to share active inference with people that are like physios.", "speaker": "SPEAKER_00"}, {"start": 2317.301, "end": 2321.647, "text": "And so people who may not necessarily have much math behind them.", "speaker": "SPEAKER_00"}, {"start": 2322.328, "end": 2327.856, "text": "And then the other extreme that I might be working with engineers that I want to be able to bring", "speaker": "SPEAKER_00"}, {"start": 2327.836, "end": 2356.813, "text": " um the social sciences intuitions and language from that world um and i and so i think to me this is like you know playing with it coming up with some demos you know creating some um ways to just talk about um what we perceive of in our models you know that and the behaviors um because again and i really do and again it's been so valuable to just focus on", "speaker": "SPEAKER_00"}, {"start": 2358.076, "end": 2359.038, "text": " Just perception.", "speaker": "SPEAKER_00"}, {"start": 2359.278, "end": 2362.644, "text": "That's all we've been breaking down.", "speaker": "SPEAKER_00"}, {"start": 2363.485, "end": 2373.322, "text": "While some of these intuitions will support us tremendously as we move into policy and action and everything else, but this has been a really valuable learning for me.", "speaker": "SPEAKER_00"}, {"start": 2373.582, "end": 2374.944, "text": "I just wanted to put that out there.", "speaker": "SPEAKER_00"}, {"start": 2374.985, "end": 2377.389, "text": "I don't know if there's anything more you'd want to share.", "speaker": "SPEAKER_00"}, {"start": 2377.469, "end": 2380.033, "text": "Maybe Andrew wants to share about his experience with", "speaker": "SPEAKER_00"}, {"start": 2380.013, "end": 2397.516, "text": " this sort of working within domains, because you've got two different worlds that converge and kind of sussing out maybe some misunderstandings that may be just sort of encoded.", "speaker": "SPEAKER_00"}, {"start": 2397.556, "end": 2402.403, "text": "Yeah.", "speaker": "SPEAKER_00"}, {"start": 2403.965, "end": 2406.308, "text": "Anyway, I'll just leave it in that.", "speaker": "SPEAKER_00"}, {"start": 2406.963, "end": 2407.484, "text": " It's tough.", "speaker": "SPEAKER_04"}, {"start": 2407.985, "end": 2411.894, "text": "Obviously, Andrew, I mean, you are you explicitly work in the social sciences and I don't.", "speaker": "SPEAKER_04"}, {"start": 2411.914, "end": 2416.644, "text": "I explicitly work in the, you know, hardcore computational side of things.", "speaker": "SPEAKER_04"}, {"start": 2416.664, "end": 2418.207, "text": "But I interface with those people.", "speaker": "SPEAKER_04"}, {"start": 2418.267, "end": 2425.984, "text": "So that's exactly why I think it's valuable to have a sort of gradated", "speaker": "SPEAKER_04"}, {"start": 2426.622, "end": 2451.178, "text": " know learning route let's say from more exploratory all the way to you know you're directly interfacing with the source code and that kind of thing um because it services those two different um necessities you know you don't as a physio let's say you know it probably doesn't behoove you to know the details about the latest variational optimization algorithm blah blah but having an appreciation of", "speaker": "SPEAKER_04"}, {"start": 2451.361, "end": 2456.227, "text": " the underlying ideas in abstract would very much.", "speaker": "SPEAKER_04"}, {"start": 2457.429, "end": 2460.853, "text": "So that's what we need to service.", "speaker": "SPEAKER_04"}, {"start": 2460.873, "end": 2472.287, "text": "One of my mentors states often like loyalty to the priors, like the problems that come up in different movement environments.", "speaker": "SPEAKER_00"}, {"start": 2472.387, "end": 2476.052, "text": "If we're too loyal to our priors, the problems we might encounter.", "speaker": "SPEAKER_00"}, {"start": 2476.032, "end": 2492.792, "text": " And so I feel like that graph and playing with those variables really do invite a deeper understanding of how we might want to be creating movement situations.", "speaker": "SPEAKER_00"}, {"start": 2492.772, "end": 2495.176, "text": " Yeah, yeah, I completely agree.", "speaker": "SPEAKER_04"}, {"start": 2495.196, "end": 2497.2, "text": "Well, thank you for that feedback.", "speaker": "SPEAKER_04"}, {"start": 2497.36, "end": 2502.81, "text": "I think that is, as I just said, it is a very central example, I think maybe of all of part one, definitely part two.", "speaker": "SPEAKER_04"}, {"start": 2504.152, "end": 2504.573, "text": "So it's good.", "speaker": "SPEAKER_04"}, {"start": 2504.693, "end": 2514.831, "text": "Yeah, it's the more we can give you guys the ability to play with that idea or similar ideas practically, I think the better.", "speaker": "SPEAKER_04"}, {"start": 2515.165, "end": 2516.526, "text": " I will see what we can do going forward.", "speaker": "SPEAKER_04"}, {"start": 2516.606, "end": 2518.208, "text": "Andrew, you had your, you had your hand up.", "speaker": "SPEAKER_04"}, {"start": 2518.308, "end": 2520.45, "text": "I didn't want to let you do.", "speaker": "SPEAKER_04"}, {"start": 2521.191, "end": 2521.632, "text": "Sure.", "speaker": "SPEAKER_03"}, {"start": 2521.652, "end": 2521.752, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 2521.772, "end": 2529.54, "text": "I just want to give a response to some of the things that Carrie said, because yeah, it's all, all of these things are what's interesting to me.", "speaker": "SPEAKER_03"}, {"start": 2529.62, "end": 2537.028, "text": "Like I'm someone who has this like deeper background in the humanities and social sciences.", "speaker": "SPEAKER_03"}, {"start": 2537.768, "end": 2542.253, "text": "And then, you know, I had an interest in psychology, but also like,", "speaker": "SPEAKER_03"}, {"start": 2542.79, "end": 2567.667, "text": " sociology and all the rest uh economics and so i was very convinced for a long time like oh like computer programming like like variational inference like what is any of that i'm never going to understand that that's what other people do uh and then maybe i collaborate with them someday but maybe not because it's kind of like scary it's a whole world of things and um", "speaker": "SPEAKER_03"}, {"start": 2567.647, "end": 2574.501, "text": " I guess it's just sort of I mean, not to wax philosophical or too broadly here, but we're in a", "speaker": "SPEAKER_03"}, {"start": 2574.734, "end": 2588.387, "text": " I mean, my perspective on anything academic or research oriented, or even just, you know, in an institutional sense or industry sense, like, we're in a highly interdisciplinary world today.", "speaker": "SPEAKER_03"}, {"start": 2588.407, "end": 2601.079, "text": "I think that there have been a lot of, like, sort of siloing of different disciplines as if they sort of pertain to their own respective material and don't really borrow from or dialogue with others.", "speaker": "SPEAKER_03"}, {"start": 2601.059, "end": 2607.813, "text": " People have been making that same criticism that I just did for decades, so I'm not saying anything particularly new.", "speaker": "SPEAKER_03"}, {"start": 2609.536, "end": 2618.595, "text": "Back in World War II, people were saying like, oh, we're all divided, not just as people, but also like in a disciplinary sense and academic sense.", "speaker": "SPEAKER_03"}, {"start": 2618.575, "end": 2623.405, "text": " So there were people trying to find ways of integrating all these different things.", "speaker": "SPEAKER_03"}, {"start": 2623.485, "end": 2626.671, "text": "And in some ways, active inference itself is this interesting thing.", "speaker": "SPEAKER_03"}, {"start": 2626.691, "end": 2639.858, "text": "We have people talking about statistical and quantum mechanics over here, and then people who are talking about emotion and affect and development of a human over here.", "speaker": "SPEAKER_03"}, {"start": 2640.04, "end": 2659.821, "text": " and um so yeah it's there's so much to wrap one's head around i think a lot of the work to be done very much relates to kind of what you seem to be pointing at carrie which i would uh one way one take on it is you know some of what we're doing is just trying to map cognitive", "speaker": "SPEAKER_03"}, {"start": 2659.801, "end": 2670.012, "text": " and not cognitive, subcognitive or subconscious processes to these mathematical variables, ultimately.", "speaker": "SPEAKER_03"}, {"start": 2670.152, "end": 2671.774, "text": "So, like, what is a prior?", "speaker": "SPEAKER_03"}, {"start": 2671.834, "end": 2673.275, "text": "What is a likelihood?", "speaker": "SPEAKER_03"}, {"start": 2673.836, "end": 2675.518, "text": "What is precision?", "speaker": "SPEAKER_03"}, {"start": 2676.098, "end": 2682.165, "text": "And, you know, we see in mathematics, actually, a lot of these things are kind of simple or basic, right?", "speaker": "SPEAKER_03"}, {"start": 2682.185, "end": 2686.129, "text": "Like a distribution, a prior distribution, or", "speaker": "SPEAKER_03"}, {"start": 2686.109, "end": 2710.075, "text": " uh precision is just like this this sort of modulator term in an equation um and then those in social science you know like like a previous an earlier version of myself are like you know intimidated by by looking at things like variables and all of that but even in the social sciences we're kind of making these different sorts of mental models like what does it mean for someone to develop in a particular", "speaker": "SPEAKER_03"}, {"start": 2710.055, "end": 2720.457, "text": " socioeconomic environment, like you're already thinking about different kinds of variables, like, oh, you're in this environment, or it could be this kind of environment that you're in and that can change its trajectory.", "speaker": "SPEAKER_03"}, {"start": 2720.578, "end": 2727.372, "text": "It doesn't look terribly different from, you know, what is the particular realization of a variable in this equation?", "speaker": "SPEAKER_03"}, {"start": 2727.352, "end": 2734.005, "text": " So, yeah, that's sort of the things that I would like to further.", "speaker": "SPEAKER_03"}, {"start": 2734.365, "end": 2736.249, "text": "I shared a couple things in the chat.", "speaker": "SPEAKER_03"}, {"start": 2736.83, "end": 2745.046, "text": "Mainly, I repeatedly share this epistemic communities under active inference paper, which is a few years old now.", "speaker": "SPEAKER_03"}, {"start": 2746.348, "end": 2747.731, "text": "But I just like it very much.", "speaker": "SPEAKER_03"}, {"start": 2747.751, "end": 2750.035, "text": "I think it's in the direction of what", "speaker": "SPEAKER_03"}, {"start": 2750.015, "end": 2752.759, "text": " social sciences and active inference could be.", "speaker": "SPEAKER_03"}, {"start": 2752.779, "end": 2764.315, "text": "It makes a lot of references to things like confirmation bias and other kinds of like psychological concepts or constructs and relates them directly to active inference.", "speaker": "SPEAKER_03"}, {"start": 2764.916, "end": 2767.72, "text": "Confirmation bias heavily relating to the idea of", "speaker": "SPEAKER_03"}, {"start": 2767.902, "end": 2770.925, "text": " self-evidencing that we see all the time in active inference.", "speaker": "SPEAKER_03"}, {"start": 2770.945, "end": 2772.827, "text": "We're trying to realize our predictions.", "speaker": "SPEAKER_03"}, {"start": 2773.989, "end": 2779.294, "text": "We're perhaps biased in a particular direction with our priors at the outset.", "speaker": "SPEAKER_03"}, {"start": 2779.795, "end": 2791.007, "text": "And so whenever we engage in the world, we very well may be trying to realize those prior predictions, whether it be something as simple as, oh, I believe that if I eat this food, I will", "speaker": "SPEAKER_03"}, {"start": 2790.987, "end": 2802.651, "text": " Then be sated very everyday sorts of things all the way down to like, oh, I will seek out this particular social group who has similar ideas to me because, you know.", "speaker": "SPEAKER_03"}, {"start": 2803.103, "end": 2807.527, "text": " they seem to have it right, just as I'm right, and we'll get along really well.", "speaker": "SPEAKER_03"}, {"start": 2807.707, "end": 2813.413, "text": "And none of this is to say that this is precisely the way that things should be worded all the time.", "speaker": "SPEAKER_03"}, {"start": 2813.433, "end": 2819.418, "text": "It's just a lot of these common principles between social science and active inferring sort of already exist.", "speaker": "SPEAKER_03"}, {"start": 2819.919, "end": 2830.849, "text": "And so some of the work to be done is simply to help with sort of that interdisciplinary dialogue, like being able to find the relationships", "speaker": "SPEAKER_03"}, {"start": 2830.829, "end": 2833.894, "text": " Between different nomenclatures and.", "speaker": "SPEAKER_03"}, {"start": 2833.934, "end": 2847.859, "text": "Engineering or mathematics or physics talking about flow we're going to talk about flow in chapter 6 and part 2 of the textbook pretty soon and then.", "speaker": "SPEAKER_03"}, {"start": 2848.193, "end": 2853.846, "text": " In psychology or social science, we can think people have a perception of time.", "speaker": "SPEAKER_03"}, {"start": 2854.708, "end": 2858.437, "text": "That's no simple thing and it's not something we can ignore.", "speaker": "SPEAKER_03"}, {"start": 2858.537, "end": 2864.852, "text": "We make predictions and assumptions about what happened in the past, what will happen in future.", "speaker": "SPEAKER_03"}, {"start": 2865.305, "end": 2868.571, "text": " You know, memory and how we store memory and all those sorts of things.", "speaker": "SPEAKER_03"}, {"start": 2868.631, "end": 2873.219, "text": "Like, there are different kinds of alignments between all of that with with active inference.", "speaker": "SPEAKER_03"}, {"start": 2873.239, "end": 2873.58, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 2873.62, "end": 2883.718, "text": "So so it's a great comment that you make and I think it's worthwhile to further those kinds of conversations at the end of the day.", "speaker": "SPEAKER_03"}, {"start": 2883.698, "end": 2889.43, "text": " You know, there will be, we could come up with our own singular universal language of how to talk about these things.", "speaker": "SPEAKER_03"}, {"start": 2889.991, "end": 2896.183, "text": "There'll be another team of researchers somewhere else who say, no, actually, we have the right universal language for talking about these things.", "speaker": "SPEAKER_03"}, {"start": 2896.203, "end": 2902.636, "text": "So we're always going to be having these kinds of like disputes and things like that probably to some degree.", "speaker": "SPEAKER_03"}, {"start": 2903.638, "end": 2904.981, "text": "But I.", "speaker": "SPEAKER_03"}, {"start": 2905.4, "end": 2913.71, "text": " I think there is zero harm in trying because I think it would be helpful to everyone to at least, you know, learn a few of these different languages.", "speaker": "SPEAKER_03"}, {"start": 2914.191, "end": 2922.641, "text": "I feel like I'm a little over speaking at some point, but I hope that this kind of subject matter comes up for like more in future in the textbook group.", "speaker": "SPEAKER_03"}, {"start": 2922.661, "end": 2927.327, "text": "It'd be nice to hear more voices and just talk more about kind of what that looks like.", "speaker": "SPEAKER_03"}, {"start": 2927.548, "end": 2935.137, "text": "Daniel is also highly interested in like trying to figure out like what is the active inference sort of ontology", "speaker": "SPEAKER_03"}, {"start": 2935.403, "end": 2960.061, "text": " and how can we you know translate he looks at it as like a translation problem like how do we translate that to uh you know psychology how do we translate that to um you know economics or whatever else um so yeah uh there's no very good though i see so yeah i was gonna say uh son did you wanna did you wanna ask a question really quick", "speaker": "SPEAKER_03"}, {"start": 2963.163, "end": 2967.79, "text": " Okay, just two quick questions.", "speaker": "SPEAKER_02"}, {"start": 2967.81, "end": 2979.608, "text": "In the book, The Free Energy Principle in Mind, Brain and Behavior, I noticed that mu is used to denote external states, while x is used for internal states.", "speaker": "SPEAKER_02"}, {"start": 2979.868, "end": 2985.977, "text": "However, in our current textbook group, the convention seems to be different from that.", "speaker": "SPEAKER_02"}, {"start": 2986.177, "end": 2989.322, "text": "Could this difference in notation cause", "speaker": "SPEAKER_02"}, {"start": 2989.302, "end": 2992.966, "text": " confusion later when applying the framework.", "speaker": "SPEAKER_02"}, {"start": 2992.986, "end": 3007.681, "text": "And the second question is, I have noticed that many papers simulating real-world psychiatry cases seem quite far from the mathematical formulas presented in fundamental sessions like this one.", "speaker": "SPEAKER_02"}, {"start": 3008.382, "end": 3011.525, "text": "Could you help me understand the relationship between them?", "speaker": "SPEAKER_02"}, {"start": 3011.825, "end": 3014.348, "text": "Is there a bridge I'm missing?", "speaker": "SPEAKER_02"}, {"start": 3016.741, "end": 3017.082, "text": " Right.", "speaker": "SPEAKER_04"}, {"start": 3017.122, "end": 3017.905, "text": "Well, OK. Yeah.", "speaker": "SPEAKER_04"}, {"start": 3017.925, "end": 3025.367, "text": "I mean, on your first question about so this is I've got the original par at all textbook up.", "speaker": "SPEAKER_04"}, {"start": 3025.702, "end": 3030.507, "text": " Yes, I mean, they I don't actually remember specifically their notation.", "speaker": "SPEAKER_04"}, {"start": 3030.567, "end": 3031.007, "text": "So here we go.", "speaker": "SPEAKER_04"}, {"start": 3031.027, "end": 3032.829, "text": "And this is maybe helpful.", "speaker": "SPEAKER_04"}, {"start": 3032.849, "end": 3046.303, "text": "So, yeah, they use X for hidden states and X star was an X for hidden states in the in the model and the agent and X star for hidden states in the environment.", "speaker": "SPEAKER_04"}, {"start": 3046.663, "end": 3050.507, "text": "So that's that's the way that Sanjeev has chosen to do things in his book as well.", "speaker": "SPEAKER_04"}, {"start": 3051.328, "end": 3053.43, "text": "And observations, why?", "speaker": "SPEAKER_04"}, {"start": 3053.765, "end": 3058.155, "text": " and actions you, but you're talking about when we look inside the model, I think,", "speaker": "SPEAKER_04"}, {"start": 3058.827, "end": 3067.559, "text": " I don't know if I'm going to get it up quickly enough, where we're looking at beliefs inside the model in terms of mu and such like.", "speaker": "SPEAKER_04"}, {"start": 3067.599, "end": 3085.001, "text": "So suffice it to say that, yes, and especially in the literature, in the papers and so on, there's lots and lots and lots of conflicting notation about exactly what's mu or eta means.", "speaker": "SPEAKER_04"}, {"start": 3085.042, "end": 3085.823, "text": "There's a nice picture.", "speaker": "SPEAKER_04"}, {"start": 3085.843, "end": 3087.124, "text": "There's a nice figure I'm trying to find.", "speaker": "SPEAKER_04"}, {"start": 3087.144, "end": 3088.446, "text": "I'm not sure where it is here.", "speaker": "SPEAKER_04"}, {"start": 3089.455, "end": 3095.603, "text": " So I'm not sure, I can't recall off the top of my head if, here we go, maybe we're talking about blanket states now.", "speaker": "SPEAKER_04"}, {"start": 3096.404, "end": 3107.057, "text": "So yes, sometimes the means of certain distributions are meant to encode internal states, you know, that stands in for x, let's say.", "speaker": "SPEAKER_04"}, {"start": 3107.398, "end": 3108.179, "text": "Sometimes it doesn't.", "speaker": "SPEAKER_04"}, {"start": 3109.6, "end": 3116.509, "text": "So we just have to be very careful about our interpretation of certain variables, and we can't", "speaker": "SPEAKER_04"}, {"start": 3116.489, "end": 3127.33, "text": " assume that they're going to hold, you know, okay, so mu here may or may not mean mu for the internal states in the Sanjeev textbook or not.", "speaker": "SPEAKER_04"}, {"start": 3127.43, "end": 3133.823, "text": "I will say that Sanjeev has made a great effort to be consistent throughout his book about how he uses variables, obviously, but", "speaker": "SPEAKER_04"}, {"start": 3134.073, "end": 3141.631, "text": " There are differences between this book and the Fundamentals book, although hopefully they're pretty minimal.", "speaker": "SPEAKER_04"}, {"start": 3142.012, "end": 3146.423, "text": "Now, your second question, I'm afraid I don't recall.", "speaker": "SPEAKER_04"}, {"start": 3146.443, "end": 3152.217, "text": "Do you mind just quickly running me through it again?", "speaker": "SPEAKER_04"}, {"start": 3153.462, "end": 3174.576, "text": " I have noticed that many papers simulating real-world psychiatric cases seem far from the mathematical formalism presented in foundations like this one.", "speaker": "SPEAKER_02"}, {"start": 3174.843, "end": 3182.176, "text": " I don't know whether if I miss some bridges between them or.", "speaker": "SPEAKER_02"}, {"start": 3182.637, "end": 3183.538, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 3184.941, "end": 3185.782, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 3186.424, "end": 3191.312, "text": "I know you've I know you've talked about this specifically.", "speaker": "SPEAKER_04"}, {"start": 3191.332, "end": 3192.454, "text": "I'll see if I can find some examples.", "speaker": "SPEAKER_04"}, {"start": 3192.474, "end": 3192.655, "text": "Yeah.", "speaker": "SPEAKER_04"}, {"start": 3193.296, "end": 3193.636, "text": "Oh, sure.", "speaker": "SPEAKER_03"}, {"start": 3194.278, "end": 3197.323, "text": "So I will say quickly, I added", "speaker": "SPEAKER_03"}, {"start": 3198.468, "end": 3209.391, "text": " this link to the chat just now on the TNU Computational Psychiatry course, which has been given for a fair number of years in a row.", "speaker": "SPEAKER_03"}, {"start": 3209.411, "end": 3211.596, "text": "And I've personally enjoyed it.", "speaker": "SPEAKER_03"}, {"start": 3211.636, "end": 3213.179, "text": "It's rather cheap.", "speaker": "SPEAKER_03"}, {"start": 3213.32, "end": 3215.785, "text": "You can attend online as well.", "speaker": "SPEAKER_03"}, {"start": 3215.765, "end": 3229.13, "text": " But what they do is that they give a very, basically for an entire week, they give many lectures in computational psychiatry from a lot of different perspectives.", "speaker": "SPEAKER_03"}, {"start": 3229.752, "end": 3233.098, "text": "And you can see that the way that they organize this course is that", "speaker": "SPEAKER_03"}, {"start": 3233.078, "end": 3239.651, "text": " They make the point that all of the researchers, all of these different directions are actually trying to do a lot of the same things.", "speaker": "SPEAKER_03"}, {"start": 3240.994, "end": 3246.966, "text": "So just kind of in that same conversation we had earlier, it's like, they're trying to seek something like a universal language.", "speaker": "SPEAKER_03"}, {"start": 3247.788, "end": 3251.475, "text": "What's really important to recognize is that as far as.", "speaker": "SPEAKER_03"}, {"start": 3251.607, "end": 3262.321, "text": " computational and what we could call precision psychiatry goes, it is very much moving into a Bayesian paradigm, Bayesian framework.", "speaker": "SPEAKER_03"}, {"start": 3262.341, "end": 3271.573, "text": "It's being used more and more as opposed to like more kind of classical statistics, like frequentist kind of paradigms.", "speaker": "SPEAKER_03"}, {"start": 3271.553, "end": 3276.039, "text": " So that's already one relationship that active inference shares with many others.", "speaker": "SPEAKER_03"}, {"start": 3277.14, "end": 3294.284, "text": "Then the questions, further questions become about action and perception because most of psychiatry's like simulations and the rest that have been done traditionally have been more focused on perception without paying attention to action.", "speaker": "SPEAKER_03"}, {"start": 3294.684, "end": 3296.647, "text": "So that's its own kind of paradigm shift.", "speaker": "SPEAKER_03"}, {"start": 3296.707, "end": 3300.632, "text": "Now we're starting to look at like trying to simulate like, okay, well not just", "speaker": "SPEAKER_03"}, {"start": 3300.612, "end": 3310.882, "text": " what does a person or, you know, experiencing a psychiatric disorder or disturbance or otherwise, like, what do they perceive, but what do they do in response?", "speaker": "SPEAKER_03"}, {"start": 3311.382, "end": 3322.753, "text": "You know, you could even relate that to sort of like a Skinnerian paradigm of like, here's the stimulus and here's the response, except in active inference, we're not just doing stimulus response.", "speaker": "SPEAKER_03"}, {"start": 3323.334, "end": 3328.999, "text": "We're doing like stimulus, like, think of that as the observations we're seeing your why.", "speaker": "SPEAKER_03"}, {"start": 3328.979, "end": 3336.206, "text": " And then at the end, we see the response, the action that they do, the behavior that they exhibit, which we haven't seen quite yet.", "speaker": "SPEAKER_03"}, {"start": 3336.226, "end": 3338.635, "text": "We're going to see that in part two of the textbook.", "speaker": "SPEAKER_03"}, {"start": 3338.8, "end": 3347.356, "text": " But then the way we break away from the Skinnerian tradition is that we also have our agent model where we can see the inside of it.", "speaker": "SPEAKER_03"}, {"start": 3348.217, "end": 3352.084, "text": "The behavioralist Skinnerian stuff ignores that.", "speaker": "SPEAKER_03"}, {"start": 3352.205, "end": 3353.748, "text": "It says, we don't know what's in the middle.", "speaker": "SPEAKER_03"}, {"start": 3353.788, "end": 3356.833, "text": "We just treat the organism, the person as a black box.", "speaker": "SPEAKER_03"}, {"start": 3357.174, "end": 3359.458, "text": "Here, we're looking at the inside.", "speaker": "SPEAKER_03"}, {"start": 3359.438, "end": 3387.808, "text": " and seeing like well if we come up with a relatively accurate model which is what all of computational psychiatry is trying to do they're trying to develop models that they can then fit to empirical Behavioral Behavior and experiments to try and get to a good model if you say you have a good model then perhaps it's the case that you can inspect the insides of it so to speak in order to better understand like how can we further develop a certain kind of treatment whether it be pharmacological", "speaker": "SPEAKER_03"}, {"start": 3388.042, "end": 3394.693, "text": " or be more related to traditional like talk therapy or otherwise.", "speaker": "SPEAKER_03"}, {"start": 3394.943, "end": 3398.267, "text": " psychodynamic therapy or something like that.", "speaker": "SPEAKER_03"}, {"start": 3398.708, "end": 3404.795, "text": "So there is a lot of strong interrelationships between the things in reference to your question.", "speaker": "SPEAKER_03"}, {"start": 3405.316, "end": 3410.242, "text": "I think the first place to look would be looking at the computational psychiatry course.", "speaker": "SPEAKER_03"}, {"start": 3410.642, "end": 3416.65, "text": "And I will say that they have made a lot of their previous year's lectures entirely free.", "speaker": "SPEAKER_03"}, {"start": 3416.73, "end": 3423.178, "text": "If you just kind of look through that website a little bit, you'll find a lot of resources and you'll see that they're actually very similar.", "speaker": "SPEAKER_03"}, {"start": 3423.158, "end": 3450.212, "text": " i think that the most distinct things about active inference is that it's it's saying like okay well everyone's doing a bayesian framework what's the right way of doing that and what are the right things we want to include active inference prioritizes keeping action in the picture rather than only assuming that we're perceptive and that everything about psychiatry is just a perception issue it also relates to what you end up doing in response to", "speaker": "SPEAKER_03"}, {"start": 3450.192, "end": 3454.879, "text": " the hallucination you're experiencing in the case of schizophrenia or something like that.", "speaker": "SPEAKER_03"}, {"start": 3455.5, "end": 3460.889, "text": "And then furthermore, mathematically, the variational methods, right?", "speaker": "SPEAKER_03"}, {"start": 3460.989, "end": 3471.525, "text": "Which, I mean, there's empirical work that's been done in relating free energy minimization specifically to different kinds of organisms' behavior and what they do.", "speaker": "SPEAKER_03"}, {"start": 3471.505, "end": 3484.914, "text": " So, yeah, there are many open questions, but I just want to make the point that active inference is not dramatically different from a lot of what's going on in computational psychiatry.", "speaker": "SPEAKER_03"}, {"start": 3485.175, "end": 3487.219, "text": "However, it will look incredibly different", "speaker": "SPEAKER_03"}, {"start": 3487.199, "end": 3495.356, "text": " from historical work in psychiatry, which is almost entirely focused on observation of people.", "speaker": "SPEAKER_03"}, {"start": 3495.416, "end": 3498.984, "text": "It's very much in the realm of like the way we're talking about social sciences earlier.", "speaker": "SPEAKER_03"}, {"start": 3499.004, "end": 3500.748, "text": "We're not quite looking at mathematics.", "speaker": "SPEAKER_03"}, {"start": 3500.768, "end": 3501.85, "text": "We're not quite looking at", "speaker": "SPEAKER_03"}, {"start": 3502.302, "end": 3508.293, "text": " Data collection, in the sense of, like, this continuous data doing neuro imaging scans and the rest.", "speaker": "SPEAKER_03"}, {"start": 3508.553, "end": 3511.459, "text": "So I hope that that's helpful as an explanation.", "speaker": "SPEAKER_03"}, {"start": 3511.479, "end": 3515.686, "text": "But the main thing is probably check out that computational psychiatry course.", "speaker": "SPEAKER_03"}, {"start": 3515.706, "end": 3523.18, "text": "They're really, they're really putting in the work to try and standardize and make a lot of things accessible on that front.", "speaker": "SPEAKER_03"}, {"start": 3524.341, "end": 3537.425, "text": " Yeah, maybe just as a quick last addendum to Sun's question, I would say that active inference, as far as I conceive of it and many other people conceive of it, is a theory of agency.", "speaker": "SPEAKER_04"}, {"start": 3537.585, "end": 3540.19, "text": "It's a theory about what an agent is.", "speaker": "SPEAKER_04"}, {"start": 3540.17, "end": 3548.121, "text": " If you make an active inference agent, you've made a thing that can perceive and act and it carves out its own little niche in some environment.", "speaker": "SPEAKER_04"}, {"start": 3548.742, "end": 3557.393, "text": "But if you're interested in maybe some narrow question about decision making under some context, you don't necessarily need to build a whole agent.", "speaker": "SPEAKER_04"}, {"start": 3557.413, "end": 3564.102, "text": "You just need to build a little bit of the procedure that corresponds to the particular decision making question at issue.", "speaker": "SPEAKER_04"}, {"start": 3564.182, "end": 3566.245, "text": "So if you're dealing with some kind of", "speaker": "SPEAKER_04"}, {"start": 3566.225, "end": 3587.676, "text": " question about okay what you know can i model a delusion as a delusion of inference on hidden states you can do that you can implement the mathematics with respect to that that could be used in an active differentiation but you don't need to build up the whole thing necessarily so i think maybe some that can be a reason why in some of the literature you see some", "speaker": "SPEAKER_04"}, {"start": 3588.196, "end": 3592.583, "text": " differing standards of implementation with respect to what is implemented.", "speaker": "SPEAKER_04"}, {"start": 3593.965, "end": 3599.213, "text": "Like you don't necessarily need to always build a full active inference agent or indeed maybe a very sophisticated one.", "speaker": "SPEAKER_04"}, {"start": 3599.233, "end": 3601.617, "text": "You can build a very simple sort of thing.", "speaker": "SPEAKER_04"}, {"start": 3601.677, "end": 3607.706, "text": "So that's another reason maybe why you don't see as much standardization as you might think in the field.", "speaker": "SPEAKER_04"}, {"start": 3607.746, "end": 3609.509, "text": "So I hope that helps as well.", "speaker": "SPEAKER_04"}, {"start": 3610.971, "end": 3611.693, "text": "Thank you.", "speaker": "SPEAKER_02"}, {"start": 3611.733, "end": 3612.634, "text": "Thank you very much.", "speaker": "SPEAKER_02"}, {"start": 3613.458, "end": 3613.899, "text": " Cool.", "speaker": "SPEAKER_04"}, {"start": 3613.919, "end": 3614.139, "text": "All right.", "speaker": "SPEAKER_04"}, {"start": 3614.159, "end": 3615.662, "text": "Well, we're out of time today, guys.", "speaker": "SPEAKER_04"}, {"start": 3615.742, "end": 3618.127, "text": "But please do add questions on the Coda.", "speaker": "SPEAKER_04"}, {"start": 3618.889, "end": 3620.832, "text": "I'll monitor questions there.", "speaker": "SPEAKER_04"}, {"start": 3620.853, "end": 3622.876, "text": "I'll monitor questions on Discord and such like.", "speaker": "SPEAKER_04"}, {"start": 3623.418, "end": 3629.169, "text": "Other than that, we'll see you all on Tuesday for the first resuming of part two for chapter six.", "speaker": "SPEAKER_04"}, {"start": 3630.091, "end": 3630.632, "text": "Thank you very much.", "speaker": "SPEAKER_04"}, {"start": 3630.672, "end": 3631.554, "text": "I'll stop the recording now.", "speaker": "SPEAKER_04"}, {"start": 3632.856, "end": 3633.357, "text": "See you later.", "speaker": "SPEAKER_04"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.txt new file mode 100644 index 000000000..43900084d --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_026/transcript.txt @@ -0,0 +1,1172 @@ +SPEAKER_04: +all right hello everyone so it's uh july 10th 2026 we're here at the active inference institute going through fundamentals of active inference so this is our second week of review we've we've gone through part one of the textbook we're moving into part two we're going to be kicking up with chapter six next week uh and part two is all about kind of active inference proper when we actually begin to put action into the mix + +So this is the last session of chapter one review, maybe more general questions and things like that before we do move on to the more regular pace of things again, where we go chapter by chapter. + +So this session today, this is an opportunity to ask questions, live questions, or maybe we can go through some of the prerecorded ones as well relating to really anything to do with part one. + +So that's what we're all here to do. + +I will say before we do that, there's interest. + +And indeed, we have been big because we have gone through part one. + +There's three parts of the book. + +We've gotten somewhat established in terms of the general lay of the land and things like this. + +And we now want to think about maybe ways in which you guys and participants at all levels would like to participate in experimenting with things practically. + +So I just want to let you all know that we are actually, so Daniel, myself, Andrew, and even Sanjeev are actively thinking about the best way to sort of create those affordances. + +As we saw with Daniel's session last time, I'll just share my screen here. + +Probably the entire screen is best. + +Let me see the big nested things. + +Hopefully everyone can see my screen. + +I'll just maximize this here. + +So we're seeing the coder just now. + +As we saw last time, Daniel went over the existing, there is a place where we host the code for the existing code such as it is in GitHub. + +And this was really cool. + +There's a lot of stuff here. + +It's very rich. + +Daniel's putting quite a bit of work about + +You know how it is that you can palpate and go through this at your own pace. + +It is a bit overwhelming at first, maybe for people who are new. + +So it's very powerful. + +So pretty much all of the code is here for all of the chapters. + +Very cool stuff. + +But one thing that we're quite enthusiastic about is surfacing maybe different levels of participation. + +So we saw in, if I go to chapter two here, I tend to think that it's probably a good idea for, you know, there's a lot of people in this group, all different backgrounds and so on, software developers, there's even a professor or two, and then there's + +variety of backgrounds and perspectives. + +I am thinking it would be quite nice to + +have a sort of opportunity for you to just explore the code very much along the lines of what Andrew did for chapter two. + +So if we come over here, Andrew created this very wonderful notebook where you don't have, I mean, there is code, you have to run some stuff for some setup, but fundamentally what you can do is you can just sort of look at an example. + +This is the canonical example in chapter two. + +I think figure 2.11, or at least that's what this instantiates, where you can have a look at the general relationship that we've got here and actually play around interactively with, okay, well, what if the observation Y was 2.6? + +How does this affect the likelihood and then the posterior distribution? + +You can sort of do this interactively. + +And we shared this for chapter two, but we haven't yet implemented these kinds of notebooks for subsequent chapters. + +I think for someone really new and who doesn't have much experience, this is probably a very nice way to begin to have some ingress into being able to play around with these ideas interactively. + +um so that'd be great so we have you know this is just looking at the effects of changing correct while changing the parameters of various distributions when it comes to bayesian inference there's a wonderful example down here as well on gradient descent some of this can be cleaned up and maybe we can hide some things that don't necessarily need to be in here or whatever but this style of interaction i think is a very + +excellent way to begin interacting with practical examples. + +So you don't have to set up anything. + +You don't have to download anything. + +You don't have to worry about GitHub repositories or anything like that. + +You just click on this link here and you'll go to this web page and you can just directly interact in the browser, which is very, very cool. + +Now, for those of you who do want more of an interaction side of thing, + +If we go here, Sanjeev has actually prepared, these are quite old, but he's prepared notebooks, code notebooks that are not yet available on the book. + +So the idea is that these would be very similar to what you're seeing over here. + +This is a notebook style thing where you've got these cells and you can run some code. + +also some notes you know this is marked out notes whatever fundamentally he's done a similar kind of thing where you know chapter one all the way through to I think chapter 10 but he hasn't had time really to to build this out in a way that would be professional enough to release yes to the public + +so this style of thing is very much what he's interested in in affording and the idea would be okay maybe you can go into chapter one uh well let me see here i'm not going to actually render anything particularly interesting to look at but if we take a look at a particular notebook you know these are these are hosted here on on github so the idea would be someone who's a little bit more interested in messing around with the code itself + +you would have the ability to have this notebook here, maybe hosted on GitHub. + +We could very easily have our own version of this in the fundamentals repo, where they're version controlled, they're on GitHub, so you clone the repo and then kind of play around with stuff in the notebook. + +And then there would be hooks for you to directly interface with certain parts of the code. + +So rather than, let's say, okay, we've got a generative model here, + +And this is our observation likelihood. + +And oh, look, we have the function that generates our observation likelihood is a linear model. + +And then what you could do is you could come down here and say, all right, well, my parameters for my linear model, they're beta 1 and beta 0. + +They're here. + +What if I change beta 1 to be + +something else you know you directly manipulate the code and then you could sort of see by changing the structure of the model and the structure of the environment how things change so that would be a little bit more of a direct use of you know python so the idea with the explorer you know the explorer type thing you wouldn't have to know any code to be able to interact with any of this stuff okay + +and you would still be able to have the benefit of interactivity with the example. + +But with this sort of style of thing, you would be able to directly play around with the code. + +And then for someone who's maybe a little bit more advanced, someone who wants to build stuff, there is actually, if I go back here into the main repo, + +Sanjeev himself has built out all of these examples in such a way that there is abstract, there's an abstract specification of what an agent is, right? + +So there's a class that says, here's an agent, and it's got a bunch of things it can do. + +It can learn, and it can infer, blah, blah, blah. + +Daniel's actually done a very similar thing in the existing repo. + +And, oh, you know, okay, there's a class for an environment, and you could hook up an agent to an environment, you know. + +This is empty just now. + +Yeah, so there's a few things missing from here. + +And that's a very yet more abstract way of doing things. + +And typically what you do is you work with these source files. + +You don't work with notebooks anymore. + +If I just look at the repo here, go to Active Inference. + +I guess it would be good to look at, what would it be good to look at? + +Maybe go to Core. + +There's a lot of stuff here from Daniel. + +We haven't looked at POMDPs yet, but I just want to get the idea. + +This is a style of problem, POMDP. + +We're going to look at that later. + +There's all these different methods for talking about a POMDP. + +You would have the ability to wire this up very much like a gym. + +Some people here are probably very familiar with gym. + +This is a well-known, go to gymnasium website, you know, set of environments where you can say, all right, maybe I've got a lunar lander, you know, and I can control this lunar lander and at certain times I can get rewards and observations and I can progress in a state space. + +very much like this here. + +So this kind of thing where you can hook up an abstract agent to an abstract problem and do active inference, that kind of thing would be very, very possible to do in the sort of source code based version of things. + +And that's very similar to what Daniel showed on Tuesday. + +So I think it's very important to be able to service this kind of spectrum of interest. + +Obviously, some people have more time than others. + +Some people have more interest than others. + +Um, but, uh, you know, there's active influence itself is so heterogeneous in terms of what it covers. + +We've got people from psychology, people from non-equilibrium statistical physics, and they all kind of speak different languages and they all have different backgrounds. + +So it's not a matter of, you know, Oh, you don't know enough code really. + +You should be spending 10 hours a day to learn code, blah, blah, blah. + +Um, but we do have to be mindful of servicing people's, um, + +degree of prior ability. + +So that's kind of what I'm thinking going forward. + +We're going to hopefully build out initially some of these notebooks here that you're seeing, the interactive style of things. + +So the fundamental example of Bayesian inference, this is really, really important. + +So that's very central. + +And then other things like gradient descent. + +This one here is excellent, but we maybe would like a visualization of the actual gradient landscape and that kind of thing. + +That would be another very canonical implementation. + +Maybe, you know, expectation maximization and then things like variational free energy. + +So that's kind of what we're thinking going forward. + +I personally would be very interested in people's thoughts as to what they would like in terms of the examples that they want to see covered. + +Obviously there's lots of examples in the book and eventually we'd like to cover them all, but it'd be good to cover the sort of canonical core of the examples so far. + +So I'm interested in people's thoughts with respect to, you know, A, the level of, or if that sort of makes sense, the kind of explorer and, + +well, you know, experimenter and then maybe builder levels of participation and experimentation. + +To that end, I myself have been sort of negligent. + +We do actually have a Discord chat here for the textbook group. + +This is actually, as Andrew said earlier, the chat that has been used for previous textbook sessions on the PAR et al textbook. + +So this is the + +Although Sanjeev's recent one really does expand on a lot of stuff in detail. + +Previous textbook sessions have been going on and they've been happening here in this chat. + +Now, this is a space for us to be able to talk about things with respect to the book. + +Myself and Andrew and Daniel are on here pretty regularly. + +So it'd be an excellent way to stay in asynchronous contact with us if you're interested in that. + +Obviously, our emails are another affordance for that. + +So if you don't have a Discord account, you'd have to make one, obviously. + +But for those who do have them, you can immediately hop on over here. + +And that would be a great way to share your thoughts on things related to examples. + +So that's that. + +I don't know if there's any questions or anything people would like to ask with respect to the examples. + +We are thinking about how to build them and so on. + +But I think now we can perhaps take any questions that are related to part one. + +I'd be very interested in assisting where I can there. + +The floor is open for live questions. + +If there aren't live questions that people don't want to ask them, we can go over some ones in the chat as well, in the page. + +Just taking a moment to read the chat here. + +Andrew, yes. + + +SPEAKER_03: +Yeah, just in lieu of + +They're not being any other immediate questions right now. + +This is a bit more of a brainstorming point that we don't have to spend too much time on. + +But I was thinking for the code examples, you know, given that right now we're in, we're going to start part two. + +So we've not got there yet. + +We ended chapter five with predictive coding. + +Um, 1 direction we could go. + +Is that we could have a code example. + +That sort of consolidates a lot of the things that we saw in chapters 2 through 5. + +Not everything, obviously, there's some methods that were given that, like, kind of contradict each other. + +They're sort of in their own respect paradigm. + +Um, but still, um. + +You know, in almost every chapter we see, it's like, well, here's the univariate case and then here's the multivariate case. + +A coding example could just be geared around the multivariate case and then there's some option to only have one hidden state and one observation modality if you did want it to be univariate or something. + +But, yeah, I'm just, you know, these are the kind of questions that we're sort of thinking about with regards to developing any code further. + +Obviously, we'd want to exploit sort of Sanjeev's code that he shared with us to make sure that we're really kind of serving what he was going for as far as developing code goes. + +But, yeah, like we could potentially do like a full, you know, + +Here's code for a hierarchical predictive coding model. + +It has all of your precision weighting and updates. + +It has parameter learning. + +It has all of the things. + +And then having some kinds of interactive options for users that maybe initially lets you specify how many hidden states are there and all the rest. + +It all has to do with how much time everyone, you know, kind of has available and how much we're able to build this out. + +And then this is also why I've, like, in the past repeatedly mentioned, like, for anyone who wants to see something in particular, that could be really helpful. + +Because given we only have so much time, it would be great to make sure we're spending our time on things that learners would most like to sort of focus upon. + +But, yeah, I'll stop there. + +I noticed Kobus has his hand up. + + +SPEAKER_01: +I could go for it. + +Yeah, Andrew, that sounds like an exciting idea. + +I like the idea of having a generic code base where you can say, I want to have so many components in my state vector, so many components in my observation vector. + +I want to have so many layers. + +But the code works for every possible case. + +Now, I realize there's some downside to that. + +It makes it more complicated. + +less easy to understand. + +But I think there's value in a generic code base. + +And maybe this can even form the beginnings of what you mentioned previously, that you want to put together a library that works for the continuous case. + +And eventually, who knows, we can make it even more generic to work for both the continuous formulation as well as the discrete formulation. + +So I like that idea. + + +SPEAKER_04: +No, I completely agree. + +I think that is a sensible way to go long term. + + +SPEAKER_03: +Yeah. + +Yeah. + +Thanks for the feedback, Kobus. + +Because, yeah, it's like to some degree, it would be great to make sure that we're not sort of starting over with the code for each, you know, chapter related example, because a lot of this is like we're building up over time. + +Like if we were to sort of do that, that more generalized approach, + +I hesitate to use the word generalized because we're seeing it so much in the textbook. + +But yeah, but a more general library, like that hierarchical predictive coding model set up that I mentioned, + +You know, in Chapters 6 and 7 and 8, we're going to be looking at generalized filtering and then active generalized filtering, which also have an immense amount of overlap with the hierarchical predictive coding models. + +We're going to start seeing similar diagrams of multiple layers, all the rest, and precision weighting. + +So then, you know, with the hierarchical predictive coding, we could then continue building off of that as a general. + +Library yeah to to like, further serve, being able to move further through the chapters. + +No. + +know, starting all over, maybe having a more consistent singular API for users to be able to call these things. + +Because, yeah, I mean, that's sort of, that's sort of where Active Inference is at on the programming side, more generally, everyone's seeking some kind of, in the same way of like what Keras, the library Keras did for deep learning, right? + +It'd be nice to have sort of like a universal + +library that kind of lets us do anything we'd like. + +Other questions that come up are like, yeah, as far as continuous and discrete state space problems go in a singular library, it's like, well, PIMDP is currently fully open source. + +I mean, we even know some of the developers, so it might be nice to, you know, we could potentially integrate with PIMTP or we could do something else entirely. + +I've been working with others who are actually trying to do the discrete state space stuff from scratch. + +PyMDP does have its limitations, so maybe it is worthwhile to think that through. + +But yeah, everything sounds great. + +It'd be nice to basically supply users with all of the fundamental sort of building blocks for anything related to active inference and + +All sorts of other things like hybrid models, where you have, like, a continuous model at a lower level and then discrete state space model at a higher level with, like, different kinds of link functions between, like, there's been work in active inference on those things. + +They exist. + +but there's not enough of them. + +And they could be really fascinating when thinking about like modeling an agent who takes in a lot of continuous value data, but then has to make discrete decisions and do planning and things like that. + +I mean, it's a very interesting case and even aligns really well with empirical reality in some ways. + +So there needs to be more work. + +Being able to have a repo where people can give that a shot as they please would be fantastic. + +Yeah, Kovac. + + +SPEAKER_01: +Yeah, and my personal preference for this code base should be to lean over to the side of making it very understandable, almost aligned with the mathematics precisely rather than making it performant. + +You know, there's a saying that says it's easier to make a correct program fast than a fast program correct. + +So, starting off, I think there's virtue in, you know, leaning to the side of, you know, understanding it as much as possible, and even the mathematics and that kind of thing, and forget about making it really fast, like, you know, the Jack's approach and things like that. + +Maybe that can follow... + + +SPEAKER_04: +in the future as a as a different version or a different development path so that would be my personal personal preference yeah sanjeev himself actually in his notebooks has taken great pains to implement the code in such a way that it is essentially one-to-one with the mathematics yeah you know eschewing all issues of efficiency and such like because that's not the point you know the point is pedagogical in terms of demonstrating + +The idea is, I completely agree. + +Yeah. + + +SPEAKER_03: +Yeah. + +Yeah. + +It's, I mean, well, I think ideally what should happen for the repo is, you know, + +I mean, first, we're trying to keep it tethered to Sanjeev's textbook for sure. + +So, like, making sure that the repo, like, more or less fully illustrates the mathematics in the textbook. + +Because, yeah, I agree, like, active inference is not purely about achieving, like, you know, high scale, full performance at the expense of sort of, you know, + +mathematical integrity, we'll say. + +You know, no cheap shortcuts, unless it's part of what you're going for. + +But I think it should be focused upon, you know, contributions to research. + +Of course, if we build out all of these things in a repo, then anyone who did want to start optimizing or using JAX or things like that would be able to then build on that further. + +And that just becomes another section of the repository. + +without sort of harming or undoing previous work that's been done. + +But yeah, much agreed. + +Because with, I mean, with JAX alone, because even I've like messed with JAX recently for some of my work that I've been doing, but some of the things I've been doing involve like structure learning, where you're like potentially adding or removing + +different hidden state factors or the discrete levels of the hidden state factors. + +And so what happens is that you end up with these matrices that are changing in dimensionality, literally from iteration to iteration. + +And so Jax doesn't really like that, right? + +Because you're, you, you, + +Ideally, with JAX, you would have a lot of your tensor shapes and the rest like fixed in advance so that you could do all the just-in-time compilation. + +But if the whole structure is changing bit by bit, while structure learning is a very interesting and kind of, you know, topical thing that a lot of people are working on in active inference, it means that JAX just isn't really amenable to it. + +And I've even gotten some insight from Sanjeev on that point because he's been working on similar things. + +So basically just recognizing like, yeah, there will be places that, you know, JAX won't make sense. + +But maybe there is an alternative method that could be used. + +So, yes, making sure we kind of meet the mathematics with fidelity prior to any kind of, you know, fun optimization tricks would be the best priority, I think. + + +SPEAKER_04: +That's always how you want to do it. + +You want to understand what you're doing first and then move to making it efficient. + +That's for sure. + +All right. + +There was a pre-written question that I think would be interesting to go through. + +But if there are live questions or if you want to put them in the chat, I'll obviously give them preference. + +Have I turned my video off? + +There we go. + +So if there are live questions, please do fire away. + +But other than that, I will maybe share this question here because I think it's a relevant one. + +So this comes from chapter three. + +Hopefully people can see my screen. + +Come down here. + +Zoom in a bit more. + +Chapter four, chapter three. + +So this is on page 72. + +So this question here. + +Oh, my lights went off. + +This is on page 72 from chapter three with respect to the expectation maximization algorithm. + +And this was a big linchpin in part one. + +Really it tied together everything we saw in chapters one and two and most of three with respect to the question of learning and inference. + +How do we do both of these things at the same time? + +Because we typically have to. + +And the expectation maximization, this is how we do it basically. + +This, you know, page 72, so equation 3.44A and B, those are the two steps of the EM algorithm. + +The question is, you know, with respect to the expectation step, when Sanjeev says fill in the unknown parameters with their means, expected value, can someone break it down in this sentence? + +I cannot relate it to equation 3.44A where an expected value of the log likelihood is calculated. + +So if I just go to that equation so that people see it, 3.44a, this is, so we have 3.44a and b. + +So this is the expectation step of the expectation maximization algorithm. + +So I guess the question in some sense is like, what is an expectation at all? + +The idea is that with the expectation maximization algorithm, we don't know what the parameters are. + +so we need to take their expectation i'll define what that is in a second with respect to the current belief about the hidden state and then we can maximize our evidence for the hidden state by doing the maximization step but basically what what this is saying here what this equation is saying is we need to so like this e is an expectation we want to we want to say what we think the + +The probability is of this thing in here, this log probability, where we're taking our values of x, the hidden state, from this distribution down here. + +I know that's a little weird. + +There's better ways of showing this visually, and that's one thing I would like to show. + +But the core idea is that the parameters, they're unknown. + +So we have some sort of distribution about what they could be. + +And we're going to use that distribution to inform a guess about what they are. + +And then we're going to use that guess in our evaluation of the log likelihood here. + +So this Vi theta, this is the parameters. + +So this is saying, you know, what, what do we think, what is our belief about what the parameters should be? + +Uh, given our belief about them. + +Now we take the expectation of the joint, uh, the general model basically under, uh, these, these, uh, the belief about the hidden state given observations. + +So that's, that's the tricky bit. + +This is a belief about parameters. + +It's not a belief about hidden states here. + +So we have to kind of average over our beliefs about hidden states. + +And that's what this is. + +I probably haven't explained that very well. + +And there's actually a lot of nice visual intuition for this here. + +But this is an important thing to get your head around, this idea of expectation. + +This comes up again and again and again. + +So this is an important question. + +I'll go through and I will actually give an answer. + +when there's more time, not in the group itself, but this question is important. + +So it's basically saying, I don't know what the parameter should be, but I can use my belief about the hidden states because I have uncertainty about them. + +I have this distribution of the hidden states to predict the hidden states most likely value in such a way that I can maintain a belief about the parameters. + +So it's kind of a, it's a guess, the expectation, + +I'll just stop sharing temporarily. + +There are a lot of other, well, there are other questions as well. + +Some of them are quite specific, so I don't necessarily want to go over them here. + +Some questions related to general things would be good. + +Andrew, did you have a comment? + + +SPEAKER_03: +I think Cobus had his hand up first. + +So if Cobus, you want to. + +Oh, I'm sorry. + + +SPEAKER_04: +I thought that was from before. + + +SPEAKER_01: +Just, you know, on the previous question about that expectation, I think what might help is to, in parallel with this expression 344a, just put down a very basic expectation expression. + +For example, the expectation of, let's say, f of x. + +And then let's make that f of x maybe 2x. + +And then the subscript of the expectation operator is simply the distribution, the density from which that function is associated. + +So here it looks very intimidating, but if you come up with a trivial example, you start to see what the role of each of these symbols are. + + +SPEAKER_04: +Yeah, no, I completely agree. + +It is a bit of a hairy thing to get your head around initially. + + +SPEAKER_01: +And what's confusing is usually the expectation operator does not have a subscript in a trivial case, but all of a sudden we encounter them with subscripts. + +And I think that's part of the confusion maybe. + + +SPEAKER_04: +Yeah, I mean, so appendix, what is it here? + +Appendix C, I think. + +If I go down here. + +Yeah, mathematical fundamentals. + +that Appendix C takes you through the notation for things like subscripts, and they're used everywhere in the book. + +So this would be a good opportunity if you're unfamiliar with some of this subscript notation to maybe go over, sprees over Appendix C again with respect to that notation, because that will continue to show up throughout the rest of the book for sure, yeah. + +No, thanks for that, Cobus. + + +SPEAKER_03: +I guess I'd also add, + +I'm not sure who asked the question that you were looking at in the coda, and maybe they do understand this, but I guess it can bear repeating of like, in chapter three, we have an algorithm that's given to us, just like we're given these other nice sort of algorithm charts. + +in other chapters and so with algorithm three in the textbook which is expectation for linear factor analysis uh it just it bears remembering that um or excuse it's worthwhile to remember that this this expectation maximization process is like iterative and we're not literally + +at precisely the same moment in the same computation, updating our hidden state estimates and parameters. + +It's like just the general notion of we're updating hidden states, then we're updating parameters, then hidden states, then parameters. + +And in both cases, whenever we move up to variational inference and we're looking at free energy, then + +sort of the golden rule is that ideally you'd be able to update anything, whether it be a hidden state estimate or a particular parameter or otherwise in this iterative fashion with respect to sort of a gradient on free energy with respect to the thing that you're updating. + +so um I I don't mean that as like an answer to the question uh but it's just like because I think that it it gets a little mixed up and like whenever you actually implement these things and computational code or otherwise it's like well yeah we have all of these different components in a model and we're talking about updating these different things but like how does that happen it's like well it's very procedural it's very iterative and it's very + + +SPEAKER_04: +we update this we update that we update this we update that so um yeah just adding that conversation now yeah fair enough that's that's for sure all right well uh yes I I have been going slowly through the uh the coder and answering questions putting my attempted answers not to uh denote well not to to connote the idea that that's final but I have been going through I do plan on um + +finishing them all. + +If I come to here, share my screen again. + +One thing that's currently missing, as some of you might note, is in chapter five, where we've left off, typically I have this overall set of notes, but the more detailed notes for sections 5.1 and 5.2 are still not there. + +I will put them in. + +I'll also, for chapter + +six next week we're going to be picking up uh produce those those same things for chapter six that's quite bare just now and that will help us maybe going forward as well so + +in a bit of a rhythm is this idea of the the the code that people can interact with just you know in a very friction-free manner quickly and easily without having to really interface with the the the details you know just so you can get an intuition that's very much something that i i would like to bring into the group and have people have another interface for people to to play with the ideas + +Because we haven't yet, especially for things like in chapter three, where we do talk a lot about the expectation maximization algorithm, it can be quite a visual thing. + +And so it can be very useful to have visualizations and such like. + +And the figures do a good job. + +But it would be nice to have additional material. + +So that's something I'm keen on providing going forward for sure. + +All right, let me just check the chat real quick. + +If there are no additional questions or anything like that, or if no one's got any burning concerns, we might potentially have an early session today, unless people are interested in asking some questions. + +Do remember the Discord. + +So that is in the chat. + +I'll just put it there again. + +There's the link to the actual Discord chat. + +So that would be, as I said, a good way to stay in contact with us more asynchronously. + +I can put my email address in the chat as well, although I think we've actually got that shared elsewhere. + + +SPEAKER_00: +Hey, Fraser, I'll jump in. + +And I'm following up on the same line of communication from the last class, which is about Andrew's chapter two demo. + +And this is more sort of just a sort of a broader approach question. + +But it seems to me that that demo or not the demo, but the but the + +um there's an opportunity to sort of train our intuition about the relationship between the priors the likelihood and and the posteriors and especially with us casting sort of um like + +anthropo, you know, like that we think, like, I think, I think you just, you called one of it the stubborn scenario. + +So just sort of running through some of these intuitions about like behavior that we, you know, and also breaking down some of the misunderstandings, especially as, as, + +active inference is used by engineers that may not have as deep of an experience in the social sciences. + +And then the other way around, social scientists who don't have such a deep understanding of + +who's driving the car, you know, like what thing is contributing to what response. + +And so I just, you know, again, like I'm in two different worlds where, you know, I want to be able to share active inference with people that are like physios. + +And so people who may not necessarily have much math behind them. + +And then the other extreme that I might be working with engineers that I want to be able to bring + +um the social sciences intuitions and language from that world um and i and so i think to me this is like you know playing with it coming up with some demos you know creating some um ways to just talk about um what we perceive of in our models you know that and the behaviors um because again and i really do and again it's been so valuable to just focus on + +Just perception. + +That's all we've been breaking down. + +While some of these intuitions will support us tremendously as we move into policy and action and everything else, but this has been a really valuable learning for me. + +I just wanted to put that out there. + +I don't know if there's anything more you'd want to share. + +Maybe Andrew wants to share about his experience with + +this sort of working within domains, because you've got two different worlds that converge and kind of sussing out maybe some misunderstandings that may be just sort of encoded. + +Yeah. + +Anyway, I'll just leave it in that. + + +SPEAKER_04: +It's tough. + +Obviously, Andrew, I mean, you are you explicitly work in the social sciences and I don't. + +I explicitly work in the, you know, hardcore computational side of things. + +But I interface with those people. + +So that's exactly why I think it's valuable to have a sort of gradated + +know learning route let's say from more exploratory all the way to you know you're directly interfacing with the source code and that kind of thing um because it services those two different um necessities you know you don't as a physio let's say you know it probably doesn't behoove you to know the details about the latest variational optimization algorithm blah blah but having an appreciation of + +the underlying ideas in abstract would very much. + +So that's what we need to service. + + +SPEAKER_00: +One of my mentors states often like loyalty to the priors, like the problems that come up in different movement environments. + +If we're too loyal to our priors, the problems we might encounter. + +And so I feel like that graph and playing with those variables really do invite a deeper understanding of how we might want to be creating movement situations. + + +SPEAKER_04: +Yeah, yeah, I completely agree. + +Well, thank you for that feedback. + +I think that is, as I just said, it is a very central example, I think maybe of all of part one, definitely part two. + +So it's good. + +Yeah, it's the more we can give you guys the ability to play with that idea or similar ideas practically, I think the better. + +I will see what we can do going forward. + +Andrew, you had your, you had your hand up. + +I didn't want to let you do. + + +SPEAKER_03: +Sure. + +Yeah. + +I just want to give a response to some of the things that Carrie said, because yeah, it's all, all of these things are what's interesting to me. + +Like I'm someone who has this like deeper background in the humanities and social sciences. + +And then, you know, I had an interest in psychology, but also like, + +sociology and all the rest uh economics and so i was very convinced for a long time like oh like computer programming like like variational inference like what is any of that i'm never going to understand that that's what other people do uh and then maybe i collaborate with them someday but maybe not because it's kind of like scary it's a whole world of things and um + +I guess it's just sort of I mean, not to wax philosophical or too broadly here, but we're in a + +I mean, my perspective on anything academic or research oriented, or even just, you know, in an institutional sense or industry sense, like, we're in a highly interdisciplinary world today. + +I think that there have been a lot of, like, sort of siloing of different disciplines as if they sort of pertain to their own respective material and don't really borrow from or dialogue with others. + +People have been making that same criticism that I just did for decades, so I'm not saying anything particularly new. + +Back in World War II, people were saying like, oh, we're all divided, not just as people, but also like in a disciplinary sense and academic sense. + +So there were people trying to find ways of integrating all these different things. + +And in some ways, active inference itself is this interesting thing. + +We have people talking about statistical and quantum mechanics over here, and then people who are talking about emotion and affect and development of a human over here. + +and um so yeah it's there's so much to wrap one's head around i think a lot of the work to be done very much relates to kind of what you seem to be pointing at carrie which i would uh one way one take on it is you know some of what we're doing is just trying to map cognitive + +and not cognitive, subcognitive or subconscious processes to these mathematical variables, ultimately. + +So, like, what is a prior? + +What is a likelihood? + +What is precision? + +And, you know, we see in mathematics, actually, a lot of these things are kind of simple or basic, right? + +Like a distribution, a prior distribution, or + +uh precision is just like this this sort of modulator term in an equation um and then those in social science you know like like a previous an earlier version of myself are like you know intimidated by by looking at things like variables and all of that but even in the social sciences we're kind of making these different sorts of mental models like what does it mean for someone to develop in a particular + +socioeconomic environment, like you're already thinking about different kinds of variables, like, oh, you're in this environment, or it could be this kind of environment that you're in and that can change its trajectory. + +It doesn't look terribly different from, you know, what is the particular realization of a variable in this equation? + +So, yeah, that's sort of the things that I would like to further. + +I shared a couple things in the chat. + +Mainly, I repeatedly share this epistemic communities under active inference paper, which is a few years old now. + +But I just like it very much. + +I think it's in the direction of what + +social sciences and active inference could be. + +It makes a lot of references to things like confirmation bias and other kinds of like psychological concepts or constructs and relates them directly to active inference. + +Confirmation bias heavily relating to the idea of + +self-evidencing that we see all the time in active inference. + +We're trying to realize our predictions. + +We're perhaps biased in a particular direction with our priors at the outset. + +And so whenever we engage in the world, we very well may be trying to realize those prior predictions, whether it be something as simple as, oh, I believe that if I eat this food, I will + +Then be sated very everyday sorts of things all the way down to like, oh, I will seek out this particular social group who has similar ideas to me because, you know. + +they seem to have it right, just as I'm right, and we'll get along really well. + +And none of this is to say that this is precisely the way that things should be worded all the time. + +It's just a lot of these common principles between social science and active inferring sort of already exist. + +And so some of the work to be done is simply to help with sort of that interdisciplinary dialogue, like being able to find the relationships + +Between different nomenclatures and. + +Engineering or mathematics or physics talking about flow we're going to talk about flow in chapter 6 and part 2 of the textbook pretty soon and then. + +In psychology or social science, we can think people have a perception of time. + +That's no simple thing and it's not something we can ignore. + +We make predictions and assumptions about what happened in the past, what will happen in future. + +You know, memory and how we store memory and all those sorts of things. + +Like, there are different kinds of alignments between all of that with with active inference. + +Yeah. + +So so it's a great comment that you make and I think it's worthwhile to further those kinds of conversations at the end of the day. + +You know, there will be, we could come up with our own singular universal language of how to talk about these things. + +There'll be another team of researchers somewhere else who say, no, actually, we have the right universal language for talking about these things. + +So we're always going to be having these kinds of like disputes and things like that probably to some degree. + +But I. + +I think there is zero harm in trying because I think it would be helpful to everyone to at least, you know, learn a few of these different languages. + +I feel like I'm a little over speaking at some point, but I hope that this kind of subject matter comes up for like more in future in the textbook group. + +It'd be nice to hear more voices and just talk more about kind of what that looks like. + +Daniel is also highly interested in like trying to figure out like what is the active inference sort of ontology + +and how can we you know translate he looks at it as like a translation problem like how do we translate that to uh you know psychology how do we translate that to um you know economics or whatever else um so yeah uh there's no very good though i see so yeah i was gonna say uh son did you wanna did you wanna ask a question really quick + + +SPEAKER_02: +Okay, just two quick questions. + +In the book, The Free Energy Principle in Mind, Brain and Behavior, I noticed that mu is used to denote external states, while x is used for internal states. + +However, in our current textbook group, the convention seems to be different from that. + +Could this difference in notation cause + +confusion later when applying the framework. + +And the second question is, I have noticed that many papers simulating real-world psychiatry cases seem quite far from the mathematical formulas presented in fundamental sessions like this one. + +Could you help me understand the relationship between them? + +Is there a bridge I'm missing? + + +SPEAKER_04: +Right. + +Well, OK. Yeah. + +I mean, on your first question about so this is I've got the original par at all textbook up. + +Yes, I mean, they I don't actually remember specifically their notation. + +So here we go. + +And this is maybe helpful. + +So, yeah, they use X for hidden states and X star was an X for hidden states in the in the model and the agent and X star for hidden states in the environment. + +So that's that's the way that Sanjeev has chosen to do things in his book as well. + +And observations, why? + +and actions you, but you're talking about when we look inside the model, I think, + +I don't know if I'm going to get it up quickly enough, where we're looking at beliefs inside the model in terms of mu and such like. + +So suffice it to say that, yes, and especially in the literature, in the papers and so on, there's lots and lots and lots of conflicting notation about exactly what's mu or eta means. + +There's a nice picture. + +There's a nice figure I'm trying to find. + +I'm not sure where it is here. + +So I'm not sure, I can't recall off the top of my head if, here we go, maybe we're talking about blanket states now. + +So yes, sometimes the means of certain distributions are meant to encode internal states, you know, that stands in for x, let's say. + +Sometimes it doesn't. + +So we just have to be very careful about our interpretation of certain variables, and we can't + +assume that they're going to hold, you know, okay, so mu here may or may not mean mu for the internal states in the Sanjeev textbook or not. + +I will say that Sanjeev has made a great effort to be consistent throughout his book about how he uses variables, obviously, but + +There are differences between this book and the Fundamentals book, although hopefully they're pretty minimal. + +Now, your second question, I'm afraid I don't recall. + +Do you mind just quickly running me through it again? + + +SPEAKER_02: +I have noticed that many papers simulating real-world psychiatric cases seem far from the mathematical formalism presented in foundations like this one. + +I don't know whether if I miss some bridges between them or. + +Yeah. + + +SPEAKER_04: +Yeah. + +I know you've I know you've talked about this specifically. + +I'll see if I can find some examples. + +Yeah. + + +SPEAKER_03: +Oh, sure. + +So I will say quickly, I added + +this link to the chat just now on the TNU Computational Psychiatry course, which has been given for a fair number of years in a row. + +And I've personally enjoyed it. + +It's rather cheap. + +You can attend online as well. + +But what they do is that they give a very, basically for an entire week, they give many lectures in computational psychiatry from a lot of different perspectives. + +And you can see that the way that they organize this course is that + +They make the point that all of the researchers, all of these different directions are actually trying to do a lot of the same things. + +So just kind of in that same conversation we had earlier, it's like, they're trying to seek something like a universal language. + +What's really important to recognize is that as far as. + +computational and what we could call precision psychiatry goes, it is very much moving into a Bayesian paradigm, Bayesian framework. + +It's being used more and more as opposed to like more kind of classical statistics, like frequentist kind of paradigms. + +So that's already one relationship that active inference shares with many others. + +Then the questions, further questions become about action and perception because most of psychiatry's like simulations and the rest that have been done traditionally have been more focused on perception without paying attention to action. + +So that's its own kind of paradigm shift. + +Now we're starting to look at like trying to simulate like, okay, well not just + +what does a person or, you know, experiencing a psychiatric disorder or disturbance or otherwise, like, what do they perceive, but what do they do in response? + +You know, you could even relate that to sort of like a Skinnerian paradigm of like, here's the stimulus and here's the response, except in active inference, we're not just doing stimulus response. + +We're doing like stimulus, like, think of that as the observations we're seeing your why. + +And then at the end, we see the response, the action that they do, the behavior that they exhibit, which we haven't seen quite yet. + +We're going to see that in part two of the textbook. + +But then the way we break away from the Skinnerian tradition is that we also have our agent model where we can see the inside of it. + +The behavioralist Skinnerian stuff ignores that. + +It says, we don't know what's in the middle. + +We just treat the organism, the person as a black box. + +Here, we're looking at the inside. + +and seeing like well if we come up with a relatively accurate model which is what all of computational psychiatry is trying to do they're trying to develop models that they can then fit to empirical Behavioral Behavior and experiments to try and get to a good model if you say you have a good model then perhaps it's the case that you can inspect the insides of it so to speak in order to better understand like how can we further develop a certain kind of treatment whether it be pharmacological + +or be more related to traditional like talk therapy or otherwise. + +psychodynamic therapy or something like that. + +So there is a lot of strong interrelationships between the things in reference to your question. + +I think the first place to look would be looking at the computational psychiatry course. + +And I will say that they have made a lot of their previous year's lectures entirely free. + +If you just kind of look through that website a little bit, you'll find a lot of resources and you'll see that they're actually very similar. + +i think that the most distinct things about active inference is that it's it's saying like okay well everyone's doing a bayesian framework what's the right way of doing that and what are the right things we want to include active inference prioritizes keeping action in the picture rather than only assuming that we're perceptive and that everything about psychiatry is just a perception issue it also relates to what you end up doing in response to + +the hallucination you're experiencing in the case of schizophrenia or something like that. + +And then furthermore, mathematically, the variational methods, right? + +Which, I mean, there's empirical work that's been done in relating free energy minimization specifically to different kinds of organisms' behavior and what they do. + +So, yeah, there are many open questions, but I just want to make the point that active inference is not dramatically different from a lot of what's going on in computational psychiatry. + +However, it will look incredibly different + +from historical work in psychiatry, which is almost entirely focused on observation of people. + +It's very much in the realm of like the way we're talking about social sciences earlier. + +We're not quite looking at mathematics. + +We're not quite looking at + +Data collection, in the sense of, like, this continuous data doing neuro imaging scans and the rest. + +So I hope that that's helpful as an explanation. + +But the main thing is probably check out that computational psychiatry course. + +They're really, they're really putting in the work to try and standardize and make a lot of things accessible on that front. + + +SPEAKER_04: +Yeah, maybe just as a quick last addendum to Sun's question, I would say that active inference, as far as I conceive of it and many other people conceive of it, is a theory of agency. + +It's a theory about what an agent is. + +If you make an active inference agent, you've made a thing that can perceive and act and it carves out its own little niche in some environment. + +But if you're interested in maybe some narrow question about decision making under some context, you don't necessarily need to build a whole agent. + +You just need to build a little bit of the procedure that corresponds to the particular decision making question at issue. + +So if you're dealing with some kind of + +question about okay what you know can i model a delusion as a delusion of inference on hidden states you can do that you can implement the mathematics with respect to that that could be used in an active differentiation but you don't need to build up the whole thing necessarily so i think maybe some that can be a reason why in some of the literature you see some + +differing standards of implementation with respect to what is implemented. + +Like you don't necessarily need to always build a full active inference agent or indeed maybe a very sophisticated one. + +You can build a very simple sort of thing. + +So that's another reason maybe why you don't see as much standardization as you might think in the field. + +So I hope that helps as well. + + +SPEAKER_02: +Thank you. + +Thank you very much. + + +SPEAKER_04: +Cool. + +All right. + +Well, we're out of time today, guys. + +But please do add questions on the Coda. + +I'll monitor questions there. + +I'll monitor questions on Discord and such like. + +Other than that, we'll see you all on Tuesday for the first resuming of part two for chapter six. + +Thank you very much. + +I'll stop the recording now. + +See you later. diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.json new file mode 100644 index 000000000..89132e10c --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.json @@ -0,0 +1 @@ +[{"video_id": "SBCl8DeaJYU", "segments": [{"start": 3.018, "end": 20.689, "text": " okay we are now going into chapter six to part two of the book uh Frasier or Andrew or anyone want to give an over uh a opening statement or look towards part two of the book", "speaker": "SPEAKER_02"}, {"start": 23.538, "end": 29.143, "text": " I guess to say that this is really part two is where we're putting the active in active inference.", "speaker": "SPEAKER_05"}, {"start": 29.444, "end": 33.027, "text": "So we're actually going to be doing active inference now, which is going to be very exciting.", "speaker": "SPEAKER_05"}, {"start": 34.008, "end": 46.08, "text": "And just to say for part, well, this chapter we're doing now, this is the first time we're looking at the hidden states being something that's dynamic, the environment being something that can change.", "speaker": "SPEAKER_05"}, {"start": 46.12, "end": 48.542, "text": "That's quite a big change from what we've seen in part one.", "speaker": "SPEAKER_03"}, {"start": 48.683, "end": 53.207, "text": "So that's what I'll say.", "speaker": "SPEAKER_03"}, {"start": 53.643, "end": 75.947, "text": " Yeah, something else that's really important right now for Chapter 6 is we're going to really be looking at temporality, which I don't think gets necessarily its own long exposition on what is meant by that, because it's more about we jump right into how we're implementing this.", "speaker": "SPEAKER_01"}, {"start": 76.948, "end": 82.013, "text": "But essentially, one, we're looking at the idea of a flow.", "speaker": "SPEAKER_01"}, {"start": 82.33, "end": 96.593, "text": " in hidden states as well as a variety of other things how do they change over time we're looking at things like temporal derivatives these sorts of things come up a lot in very interesting experiments on motor control", "speaker": "SPEAKER_01"}, {"start": 96.978, "end": 101.766, "text": " But many other things as well, and so, and then furthermore, we're going to be.", "speaker": "SPEAKER_01"}, {"start": 102.827, "end": 114.987, "text": "Sort of positioning the model in in time such that it kind of has a look back and look forward and it's going to be operating more iteratively that is receiving 1 observation at a time.", "speaker": "SPEAKER_01"}, {"start": 114.967, "end": 140.417, "text": " uh rather than some examples in previous chapters we saw had to do with like batch uh sampling and the rest so i just want to highlight that there are going to be a lot of interesting things that have to do with with time uh being sort of embedded in the in the model yeah so observations coming in one by one", "speaker": "SPEAKER_01"}, {"start": 142.473, "end": 170.1, "text": " and let's this is um from the part two pages which are before chapter six it's just on page 124 and 125. so that's a good short entry point to kind of read it and uh what is going to come up in this main section of the book if we look back to the overall part of the book", "speaker": "SPEAKER_02"}, {"start": 172.51, "end": 183.101, "text": " that this middle section is, this is like the active inference section of the book and the fundamentals is kind of the dot zero or the background for it.", "speaker": "SPEAKER_02"}, {"start": 183.801, "end": 186.044, "text": "And then part three is applications and extensions.", "speaker": "SPEAKER_02"}, {"start": 187.625, "end": 201.319, "text": "So this is kind of going into this section of the book where Namjoshi has organized the actual core that goes beyond something that might be covered in some other textbook.", "speaker": "SPEAKER_02"}, {"start": 202.733, "end": 203.034, "text": " All right.", "speaker": "SPEAKER_02"}, {"start": 206.04, "end": 208.626, "text": "Can I just say really briefly one thing?", "speaker": "SPEAKER_05"}, {"start": 208.646, "end": 210.911, "text": "I mean, I second all of that.", "speaker": "SPEAKER_05"}, {"start": 212.915, "end": 222.396, "text": "Sanjeev does say in close to the beginning of chapter six that appendix C8 on dynamical systems is going to be relevant.", "speaker": "SPEAKER_05"}, {"start": 222.376, "end": 228.285, "text": " I highly, highly recommend that for those who don't have any background in dynamical systems to have a look at Appendix C8.", "speaker": "SPEAKER_05"}, {"start": 228.426, "end": 229.087, "text": "It's not very long.", "speaker": "SPEAKER_05"}, {"start": 229.107, "end": 230.108, "text": "It's only about two pages.", "speaker": "SPEAKER_05"}, {"start": 232.091, "end": 237.179, "text": "Summing up all of dynamical systems in two pages is quite a mean feat.", "speaker": "SPEAKER_05"}, {"start": 237.259, "end": 245.392, "text": "But this is going to be the bedrock of everything that we do going forward, is we're going to have this state dynamics.", "speaker": "SPEAKER_05"}, {"start": 245.532, "end": 247.776, "text": "The hidden state is going to change throughout time.", "speaker": "SPEAKER_05"}, {"start": 247.756, "end": 254.383, "text": " So we've already seen that we have an observation generating function that takes hidden states to observations.", "speaker": "SPEAKER_05"}, {"start": 255.324, "end": 260.71, "text": "Going forward, we're also going to have the state dynamics function, and that's going to cause us a lot of pain.", "speaker": "SPEAKER_05"}, {"start": 262.111, "end": 267.617, "text": "So it's going to be very, very important going forward, this idea of the state dynamics.", "speaker": "SPEAKER_05"}, {"start": 267.637, "end": 271.401, "text": "So appendix C8, very, very relevant for this section.", "speaker": "SPEAKER_05"}, {"start": 273.22, "end": 301.32, "text": " yeah these are all really related important functions of the the ability to take in observations one by one rather than just batch the uh ability of the model to account for latent states that are changing in time and to model observables and unobservables where causes are one type of unobservable", "speaker": "SPEAKER_02"}, {"start": 302.532, "end": 326.633, "text": " and uh also that that connects to kind of the broader discussion around dynamical systems and that kind of comes up in the inactivist and philosophy and thinking about what is uh what is object and process and um order and change in what what time scale should different parts of systems be modeled in", "speaker": "SPEAKER_02"}, {"start": 326.883, "end": 334.833, "text": " which are models that reify things that are fixed, which are models that actually account for change.", "speaker": "SPEAKER_02"}, {"start": 336.775, "end": 341.0, "text": "What kinds of change can a given statistical model account for?", "speaker": "SPEAKER_02"}, {"start": 342.402, "end": 355.978, "text": "Like a linear model could have all of its parameters changing for slope and for intercept, but unless otherwise set up, it's not going to change from being that kind of linear model.", "speaker": "SPEAKER_02"}, {"start": 356.903, "end": 358.184, "text": " And then filtering.", "speaker": "SPEAKER_02"}, {"start": 359.886, "end": 360.947, "text": "Frasier put.", "speaker": "SPEAKER_02"}, {"start": 363.95, "end": 366.653, "text": "We're going to be doing a lot of filtering in the next two or so chapters.", "speaker": "SPEAKER_02"}, {"start": 367.514, "end": 367.814, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 367.834, "end": 370.537, "text": "What is the current hidden state given my current observations?", "speaker": "SPEAKER_02"}, {"start": 371.218, "end": 381.809, "text": "That's in a sense the same question that even the first thought experiment was about the brain in the box and the first chapters on latent state estimation.", "speaker": "SPEAKER_02"}, {"start": 382.99, "end": 385.493, "text": "And that was setting up a sort of", "speaker": "SPEAKER_02"}, {"start": 385.844, "end": 413.792, "text": " uh related and antecedent and other relevant latent state estimation strategies in statistics like doing a linear regression and then fitting a slope based upon a linear regression which which may even come um into play again for a generative model but it kind of just sets you up if nothing else with this question what is the current hidden state given my current observations", "speaker": "SPEAKER_02"}, {"start": 413.941, "end": 432.526, "text": " And filtering is going to take that into a more capable, more sort of nuanced setting than just fitting a batch of data to a linear regression parameter.", "speaker": "SPEAKER_02"}, {"start": 433.187, "end": 434.869, "text": "But in principle, very similar.", "speaker": "SPEAKER_02"}, {"start": 436.132, "end": 463.067, "text": " to a kind of model that has this ability to do online real-time streaming, like lifelong, and then also this relates to the live stream from earlier today on test time scaling laws, which is kind of like the temporal horizon of the generalizability", "speaker": "SPEAKER_02"}, {"start": 464.144, "end": 467.098, "text": " of a given model training.", "speaker": "SPEAKER_02"}, {"start": 468.927, "end": 472.585, "text": "So this was a pretty long presentation.", "speaker": "SPEAKER_02"}, {"start": 473.746, "end": 502.538, "text": " a really interesting idea and kind of relates to some some topics um that are coming up here but not not to get too off the track here just to say that it's the same question that's being addressed in the first chapter with a thought experiment the intro chapters with estimating like a latent state um how big is the food given the light intensity or what is the relationship of food and light intensity given the data that i've seen", "speaker": "SPEAKER_02"}, {"start": 503.142, "end": 522.228, "text": " In what was more like a batch regression setting and then certain features of that kind of pipeline and strategy is going to get changed into the setting of filtering, which has very helpful name because it's, it is kind of like an audio filtering something.", "speaker": "SPEAKER_02"}, {"start": 523.249, "end": 530.078, "text": "So like some audio signal coming in, this is okay.", "speaker": "SPEAKER_02"}, {"start": 531.543, "end": 549.49, "text": " So it's like an audio stream coming in from a microphone or like a EEG sensor or a thermometer, some kind of microphone or thermometer or sensor-like sensor.", "speaker": "SPEAKER_02"}, {"start": 552.234, "end": 553.936, "text": "And then there's a filtering process.", "speaker": "SPEAKER_02"}, {"start": 555.198, "end": 557.061, "text": "And there could be sort of", "speaker": "SPEAKER_02"}, {"start": 558.475, "end": 577.007, "text": " simple filtering processes like moving average type and the Bayesian Kalman filter on like just a moving average, something you might see from like technical analysis of financial data sets.", "speaker": "SPEAKER_02"}, {"start": 578.101, "end": 602.522, "text": " um on through more advanced filtering like for um audio filtering you could filter it based upon knowledge about the hierarchical structure of songs or speech and basically have a more structured model that would like use different knowledge about what words are likely to be in lyrics", "speaker": "SPEAKER_02"}, {"start": 602.637, "end": 614.391, "text": " to, for example, reduce uncertainty about lyrics or other features of the audio, which wouldn't really be accessible from just doing like a noise filtering, like a smoothing alone.", "speaker": "SPEAKER_02"}, {"start": 615.332, "end": 616.173, "text": "But that's very simple.", "speaker": "SPEAKER_02"}, {"start": 616.213, "end": 618.997, "text": "You could apply that to any kind of time series.", "speaker": "SPEAKER_02"}, {"start": 619.738, "end": 625.945, "text": "So that's the chapter six kind of perceptual filtering.", "speaker": "SPEAKER_02"}, {"start": 626.212, "end": 652.658, "text": " focus which is analogous to like a markov process or a partially observable markov process without the action and then chapter six goes into generalized um active generalized filtering which is just where you go from the setting of perceptual filtering where there could be varying causes externally like in that audio track", "speaker": "SPEAKER_02"}, {"start": 652.925, "end": 661.417, "text": " There could be two different parts of the song and then they switch between each other with some transitions.", "speaker": "SPEAKER_02"}, {"start": 661.958, "end": 680.805, "text": "So like latent causal states that give rise in some causal networked way to that audio stream or using frequency or wavelet deconvolution of EEG data", "speaker": "SPEAKER_02"}, {"start": 681.663, "end": 697.187, "text": " or using information across sensors to look at sensor error patterns in like an EEG dataset or an fMRI dataset, which is what's very core to the SPM work.", "speaker": "SPEAKER_02"}, {"start": 698.849, "end": 701.393, "text": "That's all on the perceptual filtering.", "speaker": "SPEAKER_02"}, {"start": 702.402, "end": 720.09, "text": " Chapter six is then gonna, or chapter seven, the active generalized filtering is going to include action in that setup as another cause that is also being modeled as action as inference.", "speaker": "SPEAKER_02"}, {"start": 720.762, "end": 738.443, "text": " It's sometimes called the autonomous states, with the cognitive and active states only separating out, whereas the agent is the perception, cognition, and action, all to the exclusion of the environment, three components of the particular partition.", "speaker": "SPEAKER_02"}, {"start": 739.024, "end": 748.896, "text": "But the autonomous states, once we get into the active setting, are what usually get directly targeted", "speaker": "SPEAKER_02"}, {"start": 748.977, "end": 753.527, "text": " for variational free energy minimization optimization.", "speaker": "SPEAKER_02"}, {"start": 754.449, "end": 768.741, "text": "Because given the, we're not trying to smooth out the audio incoming to the microphone to be unsurprising, which can be done trivially by just flattening it to zero.", "speaker": "SPEAKER_02"}, {"start": 769.902, "end": 792.737, "text": " the inference for the training or for the test time for the perception and cognition and action is really you only want to control the cognitive and active states in how that they they relate to um some overall like loss or training function", "speaker": "SPEAKER_02"}, {"start": 793.105, "end": 809.002, "text": " Because directly including the observational free energy into the training can just lead to transforming the data to be uninformative rather than addressing the data as it actually comes in.", "speaker": "SPEAKER_02"}, {"start": 809.963, "end": 821.155, "text": "Or like, again, I don't want to go too far off before we go into six, but action has all of these other secondary elements", "speaker": "SPEAKER_02"}, {"start": 822.417, "end": 851.007, "text": " questions that are going to come up six is going to focus on just the filtering in the perception setting so anyone want to add anything or ask anything um yeah this is", "speaker": "SPEAKER_02"}, {"start": 852.303, "end": 856.567, "text": " Certainly related, but not one-to-one with what you mentioned.", "speaker": "SPEAKER_01"}, {"start": 856.587, "end": 867.036, "text": "But with chapter six, as we get going, there's something that comes up with respect to flow and how hidden states change over time.", "speaker": "SPEAKER_01"}, {"start": 867.056, "end": 879.728, "text": "And as we said, this sort of like online observation by observation process, as opposed to sort of doing everything at once in a large batch or otherwise.", "speaker": "SPEAKER_01"}, {"start": 879.708, "end": 889.84, "text": " And this sort of hinges on the idea of an agent that's dynamic and able to potentially adapt in dynamic environments.", "speaker": "SPEAKER_01"}, {"start": 890.42, "end": 899.971, "text": "Thus, the very idea that the hidden state itself could change is sort of embedded in the model, which is something we haven't expressly seen before.", "speaker": "SPEAKER_01"}, {"start": 899.992, "end": 904.737, "text": "And so that leads to this notion of states having flow.", "speaker": "SPEAKER_01"}, {"start": 905.257, "end": 930.724, "text": " there being kind of like transitions between states that brings up the idea that um there are like these state transition you know functions or or sort of a state transition model within our model just like this whole time we've had likelihood models that relate observations to states we're now seeing how hidden states sort of relates to itself temporally and that's sort of where the", "speaker": "SPEAKER_01"}, {"start": 930.704, "end": 940.275, "text": " The derivatives come in, and this idea of state transitions is essentially going to be very important for the remainder of the book.", "speaker": "SPEAKER_01"}, {"start": 940.896, "end": 954.872, "text": "It's going to come up in both the continuous state and time settings that you've been looking at and will continue to look at for the next couple of chapters up to the end of active generalized filtering.", "speaker": "SPEAKER_01"}, {"start": 955.51, "end": 971.37, "text": " And then it will also be very important whenever we shift into discrete time and space models in the instance of POMDPs, which a lot of folks will have seen in active inference literature as well.", "speaker": "SPEAKER_01"}, {"start": 971.891, "end": 980.762, "text": "I think it's worthwhile to kind of hone in on some of these really important topics that right now we're just getting introduced to.", "speaker": "SPEAKER_01"}, {"start": 980.802, "end": 985.348, "text": "And it's worthwhile to sort of prioritize", "speaker": "SPEAKER_01"}, {"start": 985.328, "end": 990.337, "text": " Paying attention to those as we move through the chapters given they'll come up again and again.", "speaker": "SPEAKER_01"}, {"start": 990.737, "end": 990.978, "text": "So.", "speaker": "SPEAKER_01"}, {"start": 992.02, "end": 993.061, "text": "Yeah, that's it.", "speaker": "SPEAKER_01"}, {"start": 993.502, "end": 997.108, "text": "Yeah, and that brings up this point, which is.", "speaker": "SPEAKER_02"}, {"start": 998.31, "end": 1005.463, "text": "We have talked about how chapter 6 and 7 are going to differ with 6 focus on perception as filtering.", "speaker": "SPEAKER_02"}, {"start": 1006.064, "end": 1007.927, "text": "And 7, bringing in action.", "speaker": "SPEAKER_02"}, {"start": 1009.274, "end": 1014.081, "text": " chapters six through eight are in a continuous random variable setting.", "speaker": "SPEAKER_02"}, {"start": 1014.802, "end": 1025.878, "text": "So yes, it gets discretized like in some small way by a digital computer when it's stored, but all the math is gonna be like smooth, differentiable settings.", "speaker": "SPEAKER_02"}, {"start": 1026.92, "end": 1034.691, "text": "And then chapters nine and 10 is going to look at state spaces where there are discrete numbers of options.", "speaker": "SPEAKER_02"}, {"start": 1038.653, "end": 1056.537, "text": " just looking at this part two overview, focusing on chapter six through eight, continuous random variables get split into two approximation strategies.", "speaker": "SPEAKER_02"}, {"start": 1058.039, "end": 1068.593, "text": "Laplace quadratic approximation, which is essentially going to fit a downward fitting quadratic regression.", "speaker": "SPEAKER_02"}, {"start": 1069.501, "end": 1088.409, "text": " over a point in a distribution and just kind of move those two variables, the variables that matter for the quadratic, which are kind of like the mean and the variance of a Gaussian, but it's kind of curves down the other way and it has a few different parameters to fit.", "speaker": "SPEAKER_02"}, {"start": 1089.27, "end": 1097.382, "text": "But that's one classical approximation to some varying continuous surface.", "speaker": "SPEAKER_02"}, {"start": 1098.357, "end": 1115.792, "text": " And then there's another approximation strategy that's going to be more amenable to Bayesian causal graphical model types, the mean field approximation and the physics types.", "speaker": "SPEAKER_02"}, {"start": 1116.093, "end": 1118.096, "text": " has a variational inference.", "speaker": "SPEAKER_02"}, {"start": 1120.098, "end": 1123.022, "text": "I don't know if there could be like a Laplace variational.", "speaker": "SPEAKER_02"}, {"start": 1123.363, "end": 1143.13, "text": "So I'm not sure how we read exactly every arrow or absence of arrow, but it's definitely one that's used in the variational inference setting, as is the expectation maximization, the EM, which came up earlier in the chapters.", "speaker": "SPEAKER_02"}, {"start": 1143.835, "end": 1155.192, "text": " Looking out, I don't really read this as just as only as like a total idea map, but just how the textbook is going to order these topics.", "speaker": "SPEAKER_02"}, {"start": 1156.834, "end": 1160.52, "text": "I'm not sure if others feel similarly.", "speaker": "SPEAKER_02"}, {"start": 1161.08, "end": 1163.564, "text": "But just like they're roadmaps for the chapter.", "speaker": "SPEAKER_02"}, {"start": 1164.566, "end": 1169.613, "text": "They're not a knowledge network or like the only ordering.", "speaker": "SPEAKER_02"}, {"start": 1171.516, "end": 1171.716, "text": "All right.", "speaker": "SPEAKER_02"}, {"start": 1172.084, "end": 1174.669, "text": " Now looking to chapter 6.", "speaker": "SPEAKER_02"}, {"start": 1178.296, "end": 1186.332, "text": "First, just equation 6.1 to the points from what Fraser and Andrew said about time and about state dynamics.", "speaker": "SPEAKER_02"}, {"start": 1187.133, "end": 1192.303, "text": "It's just with the dot is one notation for a change.", "speaker": "SPEAKER_02"}, {"start": 1193.043, "end": 1222.088, "text": " um through time and also sometimes derivatives are with a prime i don't think in this book and other times with a delta lowercase d in front so kind of multiple um derivative notations and um how the integrals and derivatives are notated is going to matter in the generalized coordinates but for now it's just a single variable with its derivative and this looks almost like the equation", "speaker": "SPEAKER_02"}, {"start": 1224.7, "end": 1251.248, "text": " from some earlier chapter where we had like the latent generating process and then we had i don't know what the equation number but the latent generating process uh like the the true regression of how does food size relate to light intensity and then the observation um the observation generating from uh that", "speaker": "SPEAKER_02"}, {"start": 1251.802, "end": 1274.94, "text": " here the the minor notation difference but puts it into the dynamic hidden state setting is basically um modeling the rate of change of x as latent state that can change and then some sequence of observations associated with that so that's um the environment and the generative model", "speaker": "SPEAKER_02"}, {"start": 1275.545, "end": 1293.412, "text": " of how the environment changes, the star being like the true environmental state dynamic change, the X dot from the generative model side being like the agents representation of the state transition model.", "speaker": "SPEAKER_02"}, {"start": 1294.714, "end": 1304.269, "text": "And it's getting observations in this case, just to match that environment.", "speaker": "SPEAKER_02"}, {"start": 1305.818, "end": 1331.185, "text": " uh so chapter four kind of came to which was the variational basing inference chapter but brought up the background and introduced several formats of the variational free energy", "speaker": "SPEAKER_02"}, {"start": 1332.447, "end": 1343.303, "text": " free energy, functional, some number that has some properties where you might want to target it for approximation from a dataset.", "speaker": "SPEAKER_02"}, {"start": 1346.147, "end": 1361.29, "text": "Section 6.1 is going to go into Laplace encoded energy and some Laplace approximations, so kind of quadratic placed over a moving or over dynamical continuous landscape.", "speaker": "SPEAKER_02"}, {"start": 1361.422, "end": 1371.514, "text": " some quadratic placed over that, where there's a flow and a filtering on those hidden states and observations.", "speaker": "SPEAKER_02"}, {"start": 1373.756, "end": 1387.632, "text": "And as the last name sort of suggests, these are classical statistical methods, not doing, not doing", "speaker": "SPEAKER_02"}, {"start": 1389.012, "end": 1398.222, "text": " free energy minimization necessarily themself, but using it to fit on this free energy landscape.", "speaker": "SPEAKER_02"}, {"start": 1401.525, "end": 1410.795, "text": "On this other side, part one is provided us with this question of the static hidden state.", "speaker": "SPEAKER_02"}, {"start": 1412.257, "end": 1415.02, "text": "And this is kind of a parallel path.", "speaker": "SPEAKER_02"}, {"start": 1415.236, "end": 1423.045, "text": " getting to that same sort of point about wanting to model situation where the hidden state can change through time.", "speaker": "SPEAKER_02"}, {"start": 1424.487, "end": 1444.77, "text": "And sort of analogously to what we discussed earlier, there could be like the hidden state changing, like different food particles, the closest food particle could change through time in size so that the observations of the light intensity change through time.", "speaker": "SPEAKER_02"}, {"start": 1445.375, "end": 1456.629, "text": " So it'd be like within a fixed relationship between food particle size and the light intensity, you could have quantitative variation in that.", "speaker": "SPEAKER_02"}, {"start": 1458.612, "end": 1467.784, "text": "You could also say, well, also maybe the relationship between the food particle size and the light intensity, it can change.", "speaker": "SPEAKER_02"}, {"start": 1468.685, "end": 1471.028, "text": "And that's kind of like a derivative of a model.", "speaker": "SPEAKER_02"}, {"start": 1471.868, "end": 1475.773, "text": " which is sort of how it's gonna be addressed through time in the generalized coordinates.", "speaker": "SPEAKER_02"}, {"start": 1477.155, "end": 1483.603, "text": "But hidden state change brings up these different ways to model dynamic systems.", "speaker": "SPEAKER_02"}, {"start": 1485.685, "end": 1488.989, "text": "And that's right on page 128 in Volt.", "speaker": "SPEAKER_02"}, {"start": 1490.01, "end": 1495.277, "text": "It's like, this is how dynamical systems are addressed.", "speaker": "SPEAKER_02"}, {"start": 1496.084, "end": 1525.905, "text": " And then that kind of brings the two prior streams that are laid out there together to this hidden state update rule or just update process that combines beliefs about how true latent states and observations and all that flow through time", "speaker": "SPEAKER_02"}, {"start": 1526.307, "end": 1540.857, "text": " beliefs about those like a landscape and solving the initial value problem to kind of drop into that landscape in a place where the optimization is, um,", "speaker": "SPEAKER_02"}, {"start": 1542.676, "end": 1558.592, "text": " able to start somewhere um that's like useful to start optimizing from I think there's a lot of subtlety there so just just to say that just to start the algorithm like from on its first iteration before", "speaker": "SPEAKER_02"}, {"start": 1559.348, "end": 1587.347, "text": " those two ideas can be used to make some iterative process that's amenable to that one by one observation real-time learning because it's not just a rule or a process for going from the batch of data to minimizing like the overall least sum of squares it's a method for going from the current beliefs", "speaker": "SPEAKER_02"}, {"start": 1587.563, "end": 1590.808, "text": " and then the next observation to the updated model.", "speaker": "SPEAKER_02"}, {"start": 1592.431, "end": 1602.827, "text": "And that brings in learnt beliefs about how latent and observed variables are and change.", "speaker": "SPEAKER_02"}, {"start": 1605.732, "end": 1607.935, "text": "Any questions or anything related to that?", "speaker": "SPEAKER_02"}, {"start": 1614.074, "end": 1624.426, "text": " Just to say this, you know, this is a massive change to what we've done so far, and it's going to stay with us for the rest of the book and beyond this book and forever.", "speaker": "SPEAKER_05"}, {"start": 1624.446, "end": 1627.45, "text": "So this is definitely a big juncture.", "speaker": "SPEAKER_05"}, {"start": 1628.131, "end": 1631.455, "text": "I just want to keep highlighting that basically.", "speaker": "SPEAKER_03"}, {"start": 1634.158, "end": 1635.84, "text": "Okay, then going to the second part here.", "speaker": "SPEAKER_02"}, {"start": 1636.761, "end": 1638.503, "text": "So this was all section 6162.", "speaker": "SPEAKER_02"}, {"start": 1638.643, "end": 1642.528, "text": "Now, that", "speaker": "SPEAKER_02"}, {"start": 1643.082, "end": 1670.288, "text": " question of well how to model hidden state changes over time like all we have is this eeg trace or some sensor data and it uh how to model hidden state changes from sort of like a uh unknown signal", "speaker": "SPEAKER_02"}, {"start": 1670.504, "end": 1683.467, "text": " setting on through a setting with where you have a lot of knowledge about the domain a lot of beliefs about what it might relate to or other like sensors of the similar or different type so a lot of secondary questions but", "speaker": "SPEAKER_02"}, {"start": 1684.072, "end": 1696.868, "text": " Andrew Bickford- The real heart of how all those different data models are specified or different ways that you might use the same statistical model to like generate output data like synthetic synthetic data regenerative Ai.", "speaker": "SPEAKER_02"}, {"start": 1696.888, "end": 1710.045, "text": "Andrew Bickford- or to use it to update beliefs given observations that you know a little bit or a lot about all that kind of converges on this core question of modeling hidden state change over time Andrew.", "speaker": "SPEAKER_02"}, {"start": 1712.894, "end": 1739.34, "text": " yeah thanks um yeah i think so with this kind of like these temporal aspects and we're looking at generalized states and generalized noise i also wanted to comment on um this aspect of colored noise um it's it's very interesting because a lot of the things we've been looking at so far could uh in some respects be treated as like these kind of", "speaker": "SPEAKER_01"}, {"start": 1739.32, "end": 1769.277, "text": " markovian processes of like step by step there's one update that largely just depends on what's going on currently and not on the past or history or any kind of assumed correlations between particular things and so with noise uh yeah i just wanted to comment that so far whenever we've added noise to to functions and kind of consider them to be you know probabilistic rather than purely deterministic these are all things we've already seen uh in part one of the book", "speaker": "SPEAKER_01"}, {"start": 1769.257, "end": 1784.983, "text": " But whenever we were doing that with noise, we were assuming that the noise was itself Gaussian or like this kind of white noise, something like put in a sentence sort of like an ideally like, you know, random noise.", "speaker": "SPEAKER_01"}, {"start": 1785.083, "end": 1786.185, "text": "There's no.", "speaker": "SPEAKER_01"}, {"start": 1786.165, "end": 1789.43, "text": " patterns in the noise that are assumed to be there.", "speaker": "SPEAKER_01"}, {"start": 1789.47, "end": 1806.115, "text": "And that's why we end up with these sort of like, you know, whenever we plot the linear line of how observations are generated from states, then the observation noise is just sort of scattered around the line, no particular pattern.", "speaker": "SPEAKER_01"}, {"start": 1806.095, "end": 1809.039, "text": " But now we're moving into colored noise.", "speaker": "SPEAKER_01"}, {"start": 1809.6, "end": 1818.092, "text": "And so what happens is that we start assuming that actually there might be some sort of temporal correlation in the noise, that there's some kind of pattern within the noise itself.", "speaker": "SPEAKER_01"}, {"start": 1818.492, "end": 1827.124, "text": "So this kind of renders noise to be a potentially useful aspect for the model to sort of exploit for making predictions.", "speaker": "SPEAKER_01"}, {"start": 1827.104, "end": 1843.122, "text": " As opposed to just assuming, oh, this is just something that is always going to minutely throw off probably what would otherwise be a perfect relationship between my predictions in my internal model versus what I think is going on in the world.", "speaker": "SPEAKER_01"}, {"start": 1843.102, "end": 1854.684, "text": " So, yeah, it's all to say that it's a very interesting thing, and that's where we start seeing these sort of covariance matrices and the rest and looking at how noise relates to itself.", "speaker": "SPEAKER_01"}, {"start": 1854.884, "end": 1863.54, "text": "And then finally, some aspects of this are not purely just, oh, let's add another thing to the model to make it better at performance.", "speaker": "SPEAKER_01"}, {"start": 1863.52, "end": 1878.576, "text": " We do see in a lot of areas of neuroscience and other forms of time series analysis, different kinds of data, that whenever there's temporal correlation in the noise, that maybe helps us better understand what's really going on.", "speaker": "SPEAKER_01"}, {"start": 1878.86, "end": 1902.488, "text": " um in in reality and uh it's super fascinating how the the brain itself like noise is one of the things that neuroscientists have attempted the most to try and understand over time because it's within noise that we start seeing these patterns and like uh you know the the kind of frequencies of neural activity that that are generated that correspond or correlate to different kinds of", "speaker": "SPEAKER_01"}, {"start": 1902.468, "end": 1908.139, "text": " behavioral decision-making that an organism shows or otherwise.", "speaker": "SPEAKER_01"}, {"start": 1908.399, "end": 1919.16, "text": "So noise itself, previously we've been treating noise basically as this thing that kind of throws our model off a little bit, but as long as we can get a good enough handle on that, then it's fine.", "speaker": "SPEAKER_01"}, {"start": 1919.14, "end": 1948.017, "text": " but now suddenly noise is this kind of colored temporally rich and and covariating thing so just that's another aspect of this chapter is like like taking noise seriously rather than it's just some messy stuff that throws things off yeah great comment and fraser wrote i think section 6.1 6.2 6.3 are the theoretical core of the chapter i would agree six four and six five", "speaker": "SPEAKER_01"}, {"start": 1948.233, "end": 1975.59, "text": " are going to take that state space generative model that was of a single variable and put it into generalized coordinates of motion, which are the first derivative, second derivative, and higher order derivatives, leading to some generalizations about how different kinds of temporal dynamics in signal and noise can be brought into this model.", "speaker": "SPEAKER_02"}, {"start": 1976.566, "end": 2002.202, "text": " Fraser's right though that in the first three sections it's a focus on a variable at the sort of most core filtering layer um without going into the generalized coordinates which gets sort of layered into the same method but just as a different way to encode temporal information", "speaker": "SPEAKER_02"}, {"start": 2003.6, "end": 2014.57, "text": " Yeah, just really quickly, I think the takeaways for this chapter, the core ideas are this idea of state dynamics, number one, that you can say is going to be changing.", "speaker": "SPEAKER_05"}, {"start": 2015.13, "end": 2022.997, "text": "Number two is this idea that we're going to have a very specific kind of approximation, namely the Laplace approximation.", "speaker": "SPEAKER_05"}, {"start": 2023.057, "end": 2024.018, "text": "I think that's number two.", "speaker": "SPEAKER_05"}, {"start": 2024.798, "end": 2027.521, "text": "And that is used throughout this chapter.", "speaker": "SPEAKER_05"}, {"start": 2028.281, "end": 2032.425, "text": "But I do think it needs to be stressed that that is actually quite an important", "speaker": "SPEAKER_05"}, {"start": 2032.878, "end": 2038.031, "text": " core sort of thing that we're going to, that shows up in this chapter and will show up in other chapters as well.", "speaker": "SPEAKER_05"}, {"start": 2038.633, "end": 2041.34, "text": "That's an assumption about the nature of the variational posterior.", "speaker": "SPEAKER_05"}, {"start": 2041.66, "end": 2044.568, "text": "And then number three would be this idea of generalized coordinates of motion.", "speaker": "SPEAKER_05"}, {"start": 2044.969, "end": 2048.237, "text": "I think I suspect those are the three kind of big", "speaker": "SPEAKER_05"}, {"start": 2048.977, "end": 2071.069, "text": " ideas that you would do well to try and grok um throughout this this chapter and then there's a whole bunch of other discussion about implications of this and various you know multivariate this is univariate things but i think that those are the core takeaways to my mind at least yeah that's a great comment", "speaker": "SPEAKER_05"}, {"start": 2077.006, "end": 2079.652, "text": " State dynamics, hidden states can be changing.", "speaker": "SPEAKER_02"}, {"start": 2082.278, "end": 2088.012, "text": "There are different approximation methods, like the Laplace approximation, which is long known.", "speaker": "SPEAKER_02"}, {"start": 2088.43, "end": 2090.594, "text": " And then there's this different approximation approach.", "speaker": "SPEAKER_02"}, {"start": 2091.275, "end": 2102.014, "text": "The mean field that has some different functions or some different analytical relationships or software implementations or abilities, like, to be run incrementally.", "speaker": "SPEAKER_02"}, {"start": 2102.775, "end": 2107.263, "text": "Using a update learning.", "speaker": "SPEAKER_02"}, {"start": 2107.824, "end": 2112.953, "text": "Function then there's the generalized coordinates of motion.", "speaker": "SPEAKER_02"}, {"start": 2113.54, "end": 2129.608, "text": " which is one particular way to, a kind of physics informed way, like position, speed, acceleration, that sort of physics derivatives.", "speaker": "SPEAKER_02"}, {"start": 2134.988, "end": 2143.778, "text": " I was just going to say, for those interested in, there are a lot of people interested in implementation and experimentation with the code.", "speaker": "SPEAKER_05"}, {"start": 2144.499, "end": 2147.522, "text": "Obviously, we're going to do our best to implement some things ourselves.", "speaker": "SPEAKER_05"}, {"start": 2147.562, "end": 2162.599, "text": "But example 6.1, univariate Gaussian filtering, and then 6.2, multivariate Gaussian filtering, are extremely fundamental examples that will show up again and again and again throughout the entire book, actually.", "speaker": "SPEAKER_05"}, {"start": 2164.064, "end": 2166.526, "text": " stuff to them and things like that.", "speaker": "SPEAKER_05"}, {"start": 2166.546, "end": 2170.37, "text": "It's not quite that simple, but they are very canonical.", "speaker": "SPEAKER_05"}, {"start": 2171.811, "end": 2187.686, "text": "So if you were looking for like bang for buck examples to study and even to implement, I mean, Sanjeev himself says you would do well to maybe play around with trying to implement, you know, at least 6.1, you know, the univariate case, that's a bit easier in code.", "speaker": "SPEAKER_05"}, {"start": 2187.706, "end": 2193.531, "text": "So I think that's something that I'll endeavor to create maybe in a Google Colab or something like that.", "speaker": "SPEAKER_05"}, {"start": 2193.73, "end": 2194.952, "text": " put it in the examples repo.", "speaker": "SPEAKER_05"}, {"start": 2195.913, "end": 2202.604, "text": "Because you could just spend forever playing around with various parameters and things like that and learn by the bit.", "speaker": "SPEAKER_05"}, {"start": 2204.367, "end": 2207.792, "text": "Because that's going to be the foundation of chapter 7, really, active generalized filtering.", "speaker": "SPEAKER_05"}, {"start": 2207.812, "end": 2211.438, "text": "We're going to do this whole story again, but just by putting action into the mix.", "speaker": "SPEAKER_05"}, {"start": 2211.458, "end": 2214.763, "text": "And then we're going to have a whole bunch of different things we can look at once we do that.", "speaker": "SPEAKER_05"}, {"start": 2214.863, "end": 2218.489, "text": "So 6.1, 6.2, they're very, very, very fundamental.", "speaker": "SPEAKER_05"}, {"start": 2219.498, "end": 2220.339, "text": " Great comment.", "speaker": "SPEAKER_02"}, {"start": 2220.98, "end": 2236.326, "text": "And the multivariate is in this case, like, and in many related kind of statistical cases or algorithmic cases, is like a vectorized or parallel version of the univariate case.", "speaker": "SPEAKER_02"}, {"start": 2236.778, "end": 2244.549, "text": " And then there's sort of a simpler multivariate, which is just doing all the operations in parallel, but they don't interact with each other.", "speaker": "SPEAKER_02"}, {"start": 2244.589, "end": 2252.119, "text": "And then there's more of like a joint multivariate where you also are looking at relationships amongst different variables.", "speaker": "SPEAKER_02"}, {"start": 2254.763, "end": 2265.638, "text": "But whatever you do with multivariate, however much you use information across the different variables, it comes down to what is the one variable doing?", "speaker": "SPEAKER_02"}, {"start": 2266.36, "end": 2289.032, "text": " And figure 6.4 shows basically what is happening, is there's some X star true latent variable and a belief about that latent variable, the mean centeredness of the belief about that variable.", "speaker": "SPEAKER_02"}, {"start": 2290.875, "end": 2295.321, "text": "Then there's an observation sequence in the blue.", "speaker": "SPEAKER_02"}, {"start": 2295.503, "end": 2324.987, "text": " and a belief about an observation sequence in blue and uh the x-axis is happening through time and the initial values of the beliefs are set to 15 and 13 which are too high on both um but they really quickly drop in to to the area and then kind of follow along with this lag it's correlated", "speaker": "SPEAKER_02"}, {"start": 2325.372, "end": 2328.131, "text": " but clearly informational tracking.", "speaker": "SPEAKER_02"}, {"start": 2330.931, "end": 2331.636, "text": "And,", "speaker": "SPEAKER_02"}, {"start": 2332.358, "end": 2345.129, "text": " You could assess whether this performance is good enough or what other kind of cousin models or transformations of data or other improvements could improve the error rate, like if you were refining the model.", "speaker": "SPEAKER_02"}, {"start": 2345.649, "end": 2361.943, "text": "But this is just showing, this is kind of in character what these algorithms do, which is from a given starting point, which is the initial value problem, it can track onto and then continue to follow", "speaker": "SPEAKER_02"}, {"start": 2361.923, "end": 2386.21, "text": " with some secondary questions about like time scales of learning and how to bring complex causal models into picture a loss functional driven observable model output setting where initially when it was way out of the range of observations the loss function was high", "speaker": "SPEAKER_02"}, {"start": 2386.578, "end": 2411.86, "text": " And then a gradient was followed to lead to a number that, while it doesn't really have a value like in itself, like the fact that it's 21 or 2,600 isn't really what's relevant, but you can look at this loss training trace like you would look at the loss training trace of another kind of Bayesian model or a neural network.", "speaker": "SPEAKER_02"}, {"start": 2412.43, "end": 2434.303, "text": " and use it to model or relate to other model like absolute fit parameters or model adequacy or comparative model performance in relation to these free energy functionals to add more like context because it is a low dimensional representation of what the model is doing.", "speaker": "SPEAKER_02"}, {"start": 2434.823, "end": 2444.272, "text": " And it tells you that even if you're in a flat, you might be still way off base or totally working in the wrong distribution family, et cetera.", "speaker": "SPEAKER_02"}, {"start": 2445.593, "end": 2463.75, "text": "But in this pedagogical example, it drops in and then stays like doing smoothing, filtering with this lagged, smoothed belief updating process that happens time point by time point.", "speaker": "SPEAKER_02"}, {"start": 2464.152, "end": 2469.261, "text": " And you could imagine other sort of non-base ways to do this.", "speaker": "SPEAKER_02"}, {"start": 2469.582, "end": 2478.739, "text": "Like you could do a sliding window and then just take the average of the last end data points and then divide by the new data point.", "speaker": "SPEAKER_02"}, {"start": 2479.58, "end": 2481.604, "text": "Or you could do a linear regression", "speaker": "SPEAKER_02"}, {"start": 2482.09, "end": 2511.473, "text": " against the most recent data points or you could fit a classical time series model with like autocorrelation or time series and recalculate it at some interval so there's a total and you could do the laplace approximation there are many ways to approach the the setting that this is described by in the equations", "speaker": "SPEAKER_02"}, {"start": 2512.077, "end": 2532.488, "text": " And in this chapter, the focus is on with this, I wonder if it would be clearer if this branch were more, well, here's variational free energy, but Laplace encoded energy is only on this side.", "speaker": "SPEAKER_02"}, {"start": 2532.508, "end": 2540.38, "text": "So I'm just wondering whether,", "speaker": "SPEAKER_02"}, {"start": 2541.473, "end": 2549.48, "text": " Here it's the Laplace approximation, but when is the mean field approximation come in?", "speaker": "SPEAKER_02"}, {"start": 2553.424, "end": 2569.499, "text": "Yeah, I guess maybe the difference is that this is emphasizing the approximations we're making about the VFE itself, as opposed to our approximate posterior, which would be your mean field approximation there.", "speaker": "SPEAKER_05"}, {"start": 2570.14, "end": 2570.62, "text": "Yes.", "speaker": "SPEAKER_05"}, {"start": 2571.612, "end": 2579.21, "text": " At least that's kind of how I read it, that we're really concerned about the form of the VFE here specifically.", "speaker": "SPEAKER_05"}, {"start": 2582.959, "end": 2583.26, "text": "Yeah.", "speaker": "SPEAKER_02"}, {"start": 2585.806, "end": 2589.334, "text": "Setting up the domain generative model", "speaker": "SPEAKER_02"}, {"start": 2590.394, "end": 2601.568, "text": " so that the VFE is trainable representing what it's supposed to mean, like putting labels on images.", "speaker": "SPEAKER_02"}, {"start": 2602.429, "end": 2617.667, "text": "It's like, then, let's see how Laplace encoded energy is used in 6.1, and if that's the same or related to the Laplace approximation.", "speaker": "SPEAKER_02"}, {"start": 2622.963, "end": 2627.589, "text": " Or let's look at an example of 6.1.", "speaker": "SPEAKER_02"}, {"start": 2627.609, "end": 2627.97, "text": "Good call.", "speaker": "SPEAKER_02"}, {"start": 2636.881, "end": 2645.813, "text": "And 6.1 is expressed in terms of the language of chapter 5, where we've got prediction errors and precision weighted prediction errors.", "speaker": "SPEAKER_05"}, {"start": 2646.454, "end": 2648.998, "text": "We could equally well have not represented it that way.", "speaker": "SPEAKER_05"}, {"start": 2649.018, "end": 2650.86, "text": "It just happens to be", "speaker": "SPEAKER_05"}, {"start": 2651.33, "end": 2653.072, "text": " how we've chosen to represent things here.", "speaker": "SPEAKER_05"}, {"start": 2653.132, "end": 2659.099, "text": "I think that that emphasizes that we do have two different kinds of errors that we can, that are apparent now.", "speaker": "SPEAKER_05"}, {"start": 2659.139, "end": 2664.344, "text": "We obviously have our observation errors and state dynamic beliefs errors as well.", "speaker": "SPEAKER_05"}, {"start": 2664.364, "end": 2668.95, "text": "So it does kind of help highlight that we now have this additional thing.", "speaker": "SPEAKER_05"}, {"start": 2669.69, "end": 2670.832, "text": "Andrew, I think you've got to add up.", "speaker": "SPEAKER_05"}, {"start": 2670.872, "end": 2671.112, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 2673.094, "end": 2673.655, "text": "Yeah, sorry.", "speaker": "SPEAKER_01"}, {"start": 2673.695, "end": 2678.4, "text": "I don't want to detract from what was being said, but just like,", "speaker": "SPEAKER_01"}, {"start": 2679.292, "end": 2693.773, "text": " I mean, the big thing here is that by incorporating this notion of state transitions, or there being a flow to states, there are different ways that we can sort of model that relationship.", "speaker": "SPEAKER_01"}, {"start": 2693.99, "end": 2714.495, "text": " So, one way that we can do that, and it's frequently done in active inference and other things leading up to active inference, including this perception only sort of example that we're looking at in Chapter 6, is that the Laplace is essentially related to taking Taylor expansions.", "speaker": "SPEAKER_01"}, {"start": 2715.476, "end": 2717.398, "text": "And so, you know,", "speaker": "SPEAKER_01"}, {"start": 2717.53, "end": 2720.596, "text": " With the Laplace, it's specifically up to the second derivative.", "speaker": "SPEAKER_01"}, {"start": 2721.318, "end": 2725.226, "text": "So you can sort of think of that because the equation is not precisely the same.", "speaker": "SPEAKER_01"}, {"start": 2725.647, "end": 2733.683, "text": "But you can think of each state having its own position as well as its own speed, if it's moving up or down.", "speaker": "SPEAKER_01"}, {"start": 2734.0, "end": 2760.683, "text": " over time right that's a temporal like we're actually thinking of states changing as opposed to just potentially being static and then finally the second derivative which is something like an acceleration we can liken it to that and so we're actually doing this for for many things we're not just doing it for states we're also doing it for observations we're doing it for noise um so so all of these things kind of come together", "speaker": "SPEAKER_01"}, {"start": 2760.663, "end": 2762.928, "text": " But that's the big distinction here.", "speaker": "SPEAKER_01"}, {"start": 2762.948, "end": 2774.233, "text": "It's just that the Laplace approximation is a particular way of modeling these sort of temporal derivatives and how something relates to itself in time.", "speaker": "SPEAKER_01"}, {"start": 2774.821, "end": 2777.404, "text": " Yeah, there are other ways you could do it.", "speaker": "SPEAKER_01"}, {"start": 2777.624, "end": 2779.026, "text": "It wouldn't have to be Laplace.", "speaker": "SPEAKER_01"}, {"start": 2779.086, "end": 2794.303, "text": "You could go with a single derivative or like there are many other instances and like other writings on generalized coordinates of motion where we go up to like six different temporal derivatives, but the Laplace generally revolves around using two.", "speaker": "SPEAKER_01"}, {"start": 2794.984, "end": 2802.853, "text": "And then you could kind of potentially modify that in different areas, depending on what you're trying to do or how you think your agent should be modeled.", "speaker": "SPEAKER_01"}, {"start": 2804.926, "end": 2808.309, "text": " That's a great point.", "speaker": "SPEAKER_02"}, {"start": 2808.329, "end": 2832.773, "text": "The Laplace approximation can be run iteratively classically because it makes certain constrained assumptions, like by only considering up to the quadratic, not exploring optimization over", "speaker": "SPEAKER_02"}, {"start": 2833.006, "end": 2847.032, "text": " what higher orders of motion, like higher derivatives of what the Taylor series describes are parameterized, which is like the general polynomial setting.", "speaker": "SPEAKER_02"}, {"start": 2847.451, "end": 2862.771, "text": " But the Laplace approximation can just quickly scoot around a quadratic pretty similarly to how the variational distribution can scoot around a Gaussian.", "speaker": "SPEAKER_02"}, {"start": 2863.392, "end": 2874.387, "text": "But I believe, and let's nuance this if this is not accurate, but the Gaussian, the variational Gaussian that we move around is a true statistical distribution.", "speaker": "SPEAKER_02"}, {"start": 2874.367, "end": 2883.16, "text": " Whereas the quadratic is a heuristic, but it pierces down, and it's not necessarily a probability distribution itself.", "speaker": "SPEAKER_02"}, {"start": 2884.042, "end": 2888.669, "text": "It's just like a central tendency and curvature estimate, very locally.", "speaker": "SPEAKER_02"}, {"start": 2889.57, "end": 2900.086, "text": "Whereas the variational setting is more composable, and we're talking about real statistical variables, not just a local curvature.", "speaker": "SPEAKER_02"}, {"start": 2901.127, "end": 2901.488, "text": "Verena?", "speaker": "SPEAKER_02"}, {"start": 2906.108, "end": 2911.895, "text": " I have a question about those smooth trajectories.", "speaker": "SPEAKER_00"}, {"start": 2912.796, "end": 2935.643, "text": "So assuming that maybe we have to deal with discontinuous events or degraded information, how is this interpreted as a prediction error or does the generative model", "speaker": "SPEAKER_00"}, {"start": 2936.568, "end": 2963.655, "text": " need something like explicit change point dynamics or something like that i think a good example that explores that is um i'll put it in the chat too this is the rx infer example of a recurrent switching linear dynamical system so um there are", "speaker": "SPEAKER_02"}, {"start": 2963.838, "end": 2992.478, "text": " two different discreetly switching settings and then there's noisy data coming in that are um being on a first order fit by a smoothing filtering and then at the same time the the latent state estimation is like am i in state one or state two with low very low volatility and a high volatility uh regime", "speaker": "SPEAKER_02"}, {"start": 2993.116, "end": 2998.683, "text": " And then, uh, this is the estimates through time for that latent state.", "speaker": "SPEAKER_02"}, {"start": 3000.425, "end": 3022.212, "text": "So that's, that's kind of one way to model that, like to include what was the discontinuity of, of one type or of, of a causal type into the same sensor by, by just, um, testing, like, should there be, um,", "speaker": "SPEAKER_02"}, {"start": 3023.035, "end": 3029.609, "text": " one, two, three latent variables.", "speaker": "SPEAKER_02"}, {"start": 3030.671, "end": 3034.8, "text": "So it kind of goes, okay, now we're going to have a prior over how many latent states there are.", "speaker": "SPEAKER_02"}, {"start": 3035.624, "end": 3055.405, "text": " So it's kind of the process of going from building these hierarchical models that at the end of the day, like the thermometer or the microphone or the EEG sensor, it is going to be some continuous one by one stream of discrete data points.", "speaker": "SPEAKER_02"}, {"start": 3056.313, "end": 3072.904, "text": " And then there are like well now we can take into account sequences of that's the whole building statistical models that that explore state spaces related to that primary filter in question pressure.", "speaker": "SPEAKER_02"}, {"start": 3074.268, "end": 3077.252, "text": " Yeah, just really quickly, maybe an example from my own work.", "speaker": "SPEAKER_05"}, {"start": 3077.352, "end": 3082.88, "text": "I've got this robotic arm that has, you know, the states are the angles of the segments.", "speaker": "SPEAKER_05"}, {"start": 3083.081, "end": 3083.301, "text": "Right.", "speaker": "SPEAKER_05"}, {"start": 3083.401, "end": 3087.166, "text": "So and we have a belief about all of those angles.", "speaker": "SPEAKER_05"}, {"start": 3087.226, "end": 3095.398, "text": "But I'm approximating that belief by taking a couple of samples and then I stick those samples through my state dynamics.", "speaker": "SPEAKER_05"}, {"start": 3095.716, "end": 3116.543, "text": " but if you if if a link wraps around 180 degrees then some of those samples can wrap around and you don't know if you're at you know where you are or 180 degrees removed from now and that would be an example of a discontinuity so at the moment i just have to be careful to not ever bend", "speaker": "SPEAKER_05"}, {"start": 3116.708, "end": 3117.809, "text": " all the way around.", "speaker": "SPEAKER_05"}, {"start": 3117.829, "end": 3123.877, "text": "That would be one example of a discontinuity in my state dynamics that would cause me a lot of problems.", "speaker": "SPEAKER_05"}, {"start": 3125.299, "end": 3129.665, "text": "It just so happens that I don't actually have to worry about it in practice for whatever reason most of the time.", "speaker": "SPEAKER_05"}, {"start": 3130.326, "end": 3133.63, "text": "But those are big problems that do come up now.", "speaker": "SPEAKER_05"}, {"start": 3134.491, "end": 3136.814, "text": "You have to worry about, okay, well, is there a discontinuity here?", "speaker": "SPEAKER_05"}, {"start": 3136.854, "end": 3139.658, "text": "Is there some region where something isn't defined?", "speaker": "SPEAKER_05"}, {"start": 3141.24, "end": 3145.686, "text": "So that is definitely going to be the thing that's added to our list of concerns, basically.", "speaker": "SPEAKER_03"}, {"start": 3149.463, "end": 3157.17, "text": " Yeah, I just want to maybe provide another intuitive example, even though both of these are nice.", "speaker": "SPEAKER_01"}, {"start": 3159.072, "end": 3173.987, "text": "So imagine, you know, someone who is a catcher in baseball, or basically someone who has to catch a ball for whatever reason.", "speaker": "SPEAKER_01"}, {"start": 3174.087, "end": 3177.63, "text": "You know, maybe they're just having fun playing catch.", "speaker": "SPEAKER_01"}, {"start": 3178.049, "end": 3195.231, "text": " So in previous sort of settings that we've seen in the textbook prior to Chapter 6, we have models that are probabilistic Bayesian models that would be able to predict potentially where a ball is at in time.", "speaker": "SPEAKER_01"}, {"start": 3195.211, "end": 3214.35, "text": " um but in order to actually be able to make a full prediction it would need to be able to see the entire trajectory of the ball coming towards the the catcher the person who's trying to predict where the ball is who wants to catch it they would have to like see and advance the entire trajectory and then even then", "speaker": "SPEAKER_01"}, {"start": 3214.33, "end": 3238.195, "text": " there's nothing in the model that's saying that that trajectory is specifically a trajectory in time that there's some kind of relationship between where the ball is at in time step one versus time step four so it's just a bunch of these like kind of static observations that the the agent is getting um whenever we provide for temporality in the way that we are", "speaker": "SPEAKER_01"}, {"start": 3238.293, "end": 3243.625, "text": " Uh, here, this is where we're able to have an agent who says, oh, well, the ball is here.", "speaker": "SPEAKER_01"}, {"start": 3243.685, "end": 3257.415, "text": "But it was here just prior and because I'm using these temporal derivatives, I'm able to sort of approximate estimate predict, not just precisely where the ball is, but.", "speaker": "SPEAKER_01"}, {"start": 3257.632, "end": 3264.002, "text": " what is its current speed and direction, sort of acceleration, and potentially more if you include more terms.", "speaker": "SPEAKER_01"}, {"start": 3264.763, "end": 3275.6, "text": "And so I hope that that gives a sense of like, we're able to actually establish sort of a trajectory that allows for an expectation of where the ball will be next.", "speaker": "SPEAKER_01"}, {"start": 3275.96, "end": 3279.606, "text": "Because if we just stay in this static world without a sense of time,", "speaker": "SPEAKER_01"}, {"start": 3279.586, "end": 3286.214, "text": " then it would be very surprising to us to see that the ball suddenly changed position from one time step to the next.", "speaker": "SPEAKER_01"}, {"start": 3286.315, "end": 3293.944, "text": "We have no sense of how that could happen if we don't include some sense of flow or temporal dynamics into our model.", "speaker": "SPEAKER_01"}, {"start": 3296.988, "end": 3297.529, "text": "Thank you.", "speaker": "SPEAKER_00"}, {"start": 3298.19, "end": 3303.096, "text": "I was thinking about cochlear implant recipients.", "speaker": "SPEAKER_00"}, {"start": 3303.582, "end": 3332.239, "text": " who have a very degraded signal input over average 22 electrodes and to actually become the full you know voice or the full information over the words takes time through learning and this is not a very linear process so that was the reason why i was asking", "speaker": "SPEAKER_00"}, {"start": 3332.877, "end": 3347.658, "text": " Yeah, just as like one place to explore, but I don't think we'll cover it, is that it's the hearing active inference setting for DeVry, which is related to the RX and for streaming real time.", "speaker": "SPEAKER_02"}, {"start": 3348.138, "end": 3356.81, "text": "So definitely like the audio processing and hearing is related to what people are working on.", "speaker": "SPEAKER_02"}, {"start": 3359.033, "end": 3360.876, "text": "Okay, I'm going to stop the recording there.", "speaker": "SPEAKER_02"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.txt new file mode 100644 index 000000000..16034e36a --- /dev/null +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_027/transcript.txt @@ -0,0 +1,672 @@ +SPEAKER_02: +okay we are now going into chapter six to part two of the book uh Frasier or Andrew or anyone want to give an over uh a opening statement or look towards part two of the book + + +SPEAKER_05: +I guess to say that this is really part two is where we're putting the active in active inference. + +So we're actually going to be doing active inference now, which is going to be very exciting. + +And just to say for part, well, this chapter we're doing now, this is the first time we're looking at the hidden states being something that's dynamic, the environment being something that can change. + + +SPEAKER_03: +That's quite a big change from what we've seen in part one. + +So that's what I'll say. + + +SPEAKER_01: +Yeah, something else that's really important right now for Chapter 6 is we're going to really be looking at temporality, which I don't think gets necessarily its own long exposition on what is meant by that, because it's more about we jump right into how we're implementing this. + +But essentially, one, we're looking at the idea of a flow. + +in hidden states as well as a variety of other things how do they change over time we're looking at things like temporal derivatives these sorts of things come up a lot in very interesting experiments on motor control + +But many other things as well, and so, and then furthermore, we're going to be. + +Sort of positioning the model in in time such that it kind of has a look back and look forward and it's going to be operating more iteratively that is receiving 1 observation at a time. + +uh rather than some examples in previous chapters we saw had to do with like batch uh sampling and the rest so i just want to highlight that there are going to be a lot of interesting things that have to do with with time uh being sort of embedded in the in the model yeah so observations coming in one by one + + +SPEAKER_02: +and let's this is um from the part two pages which are before chapter six it's just on page 124 and 125. so that's a good short entry point to kind of read it and uh what is going to come up in this main section of the book if we look back to the overall part of the book + +that this middle section is, this is like the active inference section of the book and the fundamentals is kind of the dot zero or the background for it. + +And then part three is applications and extensions. + +So this is kind of going into this section of the book where Namjoshi has organized the actual core that goes beyond something that might be covered in some other textbook. + +All right. + + +SPEAKER_05: +Can I just say really briefly one thing? + +I mean, I second all of that. + +Sanjeev does say in close to the beginning of chapter six that appendix C8 on dynamical systems is going to be relevant. + +I highly, highly recommend that for those who don't have any background in dynamical systems to have a look at Appendix C8. + +It's not very long. + +It's only about two pages. + +Summing up all of dynamical systems in two pages is quite a mean feat. + +But this is going to be the bedrock of everything that we do going forward, is we're going to have this state dynamics. + +The hidden state is going to change throughout time. + +So we've already seen that we have an observation generating function that takes hidden states to observations. + +Going forward, we're also going to have the state dynamics function, and that's going to cause us a lot of pain. + +So it's going to be very, very important going forward, this idea of the state dynamics. + +So appendix C8, very, very relevant for this section. + + +SPEAKER_02: +yeah these are all really related important functions of the the ability to take in observations one by one rather than just batch the uh ability of the model to account for latent states that are changing in time and to model observables and unobservables where causes are one type of unobservable + +and uh also that that connects to kind of the broader discussion around dynamical systems and that kind of comes up in the inactivist and philosophy and thinking about what is uh what is object and process and um order and change in what what time scale should different parts of systems be modeled in + +which are models that reify things that are fixed, which are models that actually account for change. + +What kinds of change can a given statistical model account for? + +Like a linear model could have all of its parameters changing for slope and for intercept, but unless otherwise set up, it's not going to change from being that kind of linear model. + +And then filtering. + +Frasier put. + +We're going to be doing a lot of filtering in the next two or so chapters. + +Yeah. + +What is the current hidden state given my current observations? + +That's in a sense the same question that even the first thought experiment was about the brain in the box and the first chapters on latent state estimation. + +And that was setting up a sort of + +uh related and antecedent and other relevant latent state estimation strategies in statistics like doing a linear regression and then fitting a slope based upon a linear regression which which may even come um into play again for a generative model but it kind of just sets you up if nothing else with this question what is the current hidden state given my current observations + +And filtering is going to take that into a more capable, more sort of nuanced setting than just fitting a batch of data to a linear regression parameter. + +But in principle, very similar. + +to a kind of model that has this ability to do online real-time streaming, like lifelong, and then also this relates to the live stream from earlier today on test time scaling laws, which is kind of like the temporal horizon of the generalizability + +of a given model training. + +So this was a pretty long presentation. + +a really interesting idea and kind of relates to some some topics um that are coming up here but not not to get too off the track here just to say that it's the same question that's being addressed in the first chapter with a thought experiment the intro chapters with estimating like a latent state um how big is the food given the light intensity or what is the relationship of food and light intensity given the data that i've seen + +In what was more like a batch regression setting and then certain features of that kind of pipeline and strategy is going to get changed into the setting of filtering, which has very helpful name because it's, it is kind of like an audio filtering something. + +So like some audio signal coming in, this is okay. + +So it's like an audio stream coming in from a microphone or like a EEG sensor or a thermometer, some kind of microphone or thermometer or sensor-like sensor. + +And then there's a filtering process. + +And there could be sort of + +simple filtering processes like moving average type and the Bayesian Kalman filter on like just a moving average, something you might see from like technical analysis of financial data sets. + +um on through more advanced filtering like for um audio filtering you could filter it based upon knowledge about the hierarchical structure of songs or speech and basically have a more structured model that would like use different knowledge about what words are likely to be in lyrics + +to, for example, reduce uncertainty about lyrics or other features of the audio, which wouldn't really be accessible from just doing like a noise filtering, like a smoothing alone. + +But that's very simple. + +You could apply that to any kind of time series. + +So that's the chapter six kind of perceptual filtering. + +focus which is analogous to like a markov process or a partially observable markov process without the action and then chapter six goes into generalized um active generalized filtering which is just where you go from the setting of perceptual filtering where there could be varying causes externally like in that audio track + +There could be two different parts of the song and then they switch between each other with some transitions. + +So like latent causal states that give rise in some causal networked way to that audio stream or using frequency or wavelet deconvolution of EEG data + +or using information across sensors to look at sensor error patterns in like an EEG dataset or an fMRI dataset, which is what's very core to the SPM work. + +That's all on the perceptual filtering. + +Chapter six is then gonna, or chapter seven, the active generalized filtering is going to include action in that setup as another cause that is also being modeled as action as inference. + +It's sometimes called the autonomous states, with the cognitive and active states only separating out, whereas the agent is the perception, cognition, and action, all to the exclusion of the environment, three components of the particular partition. + +But the autonomous states, once we get into the active setting, are what usually get directly targeted + +for variational free energy minimization optimization. + +Because given the, we're not trying to smooth out the audio incoming to the microphone to be unsurprising, which can be done trivially by just flattening it to zero. + +the inference for the training or for the test time for the perception and cognition and action is really you only want to control the cognitive and active states in how that they they relate to um some overall like loss or training function + +Because directly including the observational free energy into the training can just lead to transforming the data to be uninformative rather than addressing the data as it actually comes in. + +Or like, again, I don't want to go too far off before we go into six, but action has all of these other secondary elements + +questions that are going to come up six is going to focus on just the filtering in the perception setting so anyone want to add anything or ask anything um yeah this is + + +SPEAKER_01: +Certainly related, but not one-to-one with what you mentioned. + +But with chapter six, as we get going, there's something that comes up with respect to flow and how hidden states change over time. + +And as we said, this sort of like online observation by observation process, as opposed to sort of doing everything at once in a large batch or otherwise. + +And this sort of hinges on the idea of an agent that's dynamic and able to potentially adapt in dynamic environments. + +Thus, the very idea that the hidden state itself could change is sort of embedded in the model, which is something we haven't expressly seen before. + +And so that leads to this notion of states having flow. + +there being kind of like transitions between states that brings up the idea that um there are like these state transition you know functions or or sort of a state transition model within our model just like this whole time we've had likelihood models that relate observations to states we're now seeing how hidden states sort of relates to itself temporally and that's sort of where the + +The derivatives come in, and this idea of state transitions is essentially going to be very important for the remainder of the book. + +It's going to come up in both the continuous state and time settings that you've been looking at and will continue to look at for the next couple of chapters up to the end of active generalized filtering. + +And then it will also be very important whenever we shift into discrete time and space models in the instance of POMDPs, which a lot of folks will have seen in active inference literature as well. + +I think it's worthwhile to kind of hone in on some of these really important topics that right now we're just getting introduced to. + +And it's worthwhile to sort of prioritize + +Paying attention to those as we move through the chapters given they'll come up again and again. + +So. + +Yeah, that's it. + + +SPEAKER_02: +Yeah, and that brings up this point, which is. + +We have talked about how chapter 6 and 7 are going to differ with 6 focus on perception as filtering. + +And 7, bringing in action. + +chapters six through eight are in a continuous random variable setting. + +So yes, it gets discretized like in some small way by a digital computer when it's stored, but all the math is gonna be like smooth, differentiable settings. + +And then chapters nine and 10 is going to look at state spaces where there are discrete numbers of options. + +just looking at this part two overview, focusing on chapter six through eight, continuous random variables get split into two approximation strategies. + +Laplace quadratic approximation, which is essentially going to fit a downward fitting quadratic regression. + +over a point in a distribution and just kind of move those two variables, the variables that matter for the quadratic, which are kind of like the mean and the variance of a Gaussian, but it's kind of curves down the other way and it has a few different parameters to fit. + +But that's one classical approximation to some varying continuous surface. + +And then there's another approximation strategy that's going to be more amenable to Bayesian causal graphical model types, the mean field approximation and the physics types. + +has a variational inference. + +I don't know if there could be like a Laplace variational. + +So I'm not sure how we read exactly every arrow or absence of arrow, but it's definitely one that's used in the variational inference setting, as is the expectation maximization, the EM, which came up earlier in the chapters. + +Looking out, I don't really read this as just as only as like a total idea map, but just how the textbook is going to order these topics. + +I'm not sure if others feel similarly. + +But just like they're roadmaps for the chapter. + +They're not a knowledge network or like the only ordering. + +All right. + +Now looking to chapter 6. + +First, just equation 6.1 to the points from what Fraser and Andrew said about time and about state dynamics. + +It's just with the dot is one notation for a change. + +um through time and also sometimes derivatives are with a prime i don't think in this book and other times with a delta lowercase d in front so kind of multiple um derivative notations and um how the integrals and derivatives are notated is going to matter in the generalized coordinates but for now it's just a single variable with its derivative and this looks almost like the equation + +from some earlier chapter where we had like the latent generating process and then we had i don't know what the equation number but the latent generating process uh like the the true regression of how does food size relate to light intensity and then the observation um the observation generating from uh that + +here the the minor notation difference but puts it into the dynamic hidden state setting is basically um modeling the rate of change of x as latent state that can change and then some sequence of observations associated with that so that's um the environment and the generative model + +of how the environment changes, the star being like the true environmental state dynamic change, the X dot from the generative model side being like the agents representation of the state transition model. + +And it's getting observations in this case, just to match that environment. + +uh so chapter four kind of came to which was the variational basing inference chapter but brought up the background and introduced several formats of the variational free energy + +free energy, functional, some number that has some properties where you might want to target it for approximation from a dataset. + +Section 6.1 is going to go into Laplace encoded energy and some Laplace approximations, so kind of quadratic placed over a moving or over dynamical continuous landscape. + +some quadratic placed over that, where there's a flow and a filtering on those hidden states and observations. + +And as the last name sort of suggests, these are classical statistical methods, not doing, not doing + +free energy minimization necessarily themself, but using it to fit on this free energy landscape. + +On this other side, part one is provided us with this question of the static hidden state. + +And this is kind of a parallel path. + +getting to that same sort of point about wanting to model situation where the hidden state can change through time. + +And sort of analogously to what we discussed earlier, there could be like the hidden state changing, like different food particles, the closest food particle could change through time in size so that the observations of the light intensity change through time. + +So it'd be like within a fixed relationship between food particle size and the light intensity, you could have quantitative variation in that. + +You could also say, well, also maybe the relationship between the food particle size and the light intensity, it can change. + +And that's kind of like a derivative of a model. + +which is sort of how it's gonna be addressed through time in the generalized coordinates. + +But hidden state change brings up these different ways to model dynamic systems. + +And that's right on page 128 in Volt. + +It's like, this is how dynamical systems are addressed. + +And then that kind of brings the two prior streams that are laid out there together to this hidden state update rule or just update process that combines beliefs about how true latent states and observations and all that flow through time + +beliefs about those like a landscape and solving the initial value problem to kind of drop into that landscape in a place where the optimization is, um, + +able to start somewhere um that's like useful to start optimizing from I think there's a lot of subtlety there so just just to say that just to start the algorithm like from on its first iteration before + +those two ideas can be used to make some iterative process that's amenable to that one by one observation real-time learning because it's not just a rule or a process for going from the batch of data to minimizing like the overall least sum of squares it's a method for going from the current beliefs + +and then the next observation to the updated model. + +And that brings in learnt beliefs about how latent and observed variables are and change. + +Any questions or anything related to that? + + +SPEAKER_05: +Just to say this, you know, this is a massive change to what we've done so far, and it's going to stay with us for the rest of the book and beyond this book and forever. + +So this is definitely a big juncture. + + +SPEAKER_03: +I just want to keep highlighting that basically. + + +SPEAKER_02: +Okay, then going to the second part here. + +So this was all section 6162. + +Now, that + +question of well how to model hidden state changes over time like all we have is this eeg trace or some sensor data and it uh how to model hidden state changes from sort of like a uh unknown signal + +setting on through a setting with where you have a lot of knowledge about the domain a lot of beliefs about what it might relate to or other like sensors of the similar or different type so a lot of secondary questions but + +Andrew Bickford- The real heart of how all those different data models are specified or different ways that you might use the same statistical model to like generate output data like synthetic synthetic data regenerative Ai. + +Andrew Bickford- or to use it to update beliefs given observations that you know a little bit or a lot about all that kind of converges on this core question of modeling hidden state change over time Andrew. + + +SPEAKER_01: +yeah thanks um yeah i think so with this kind of like these temporal aspects and we're looking at generalized states and generalized noise i also wanted to comment on um this aspect of colored noise um it's it's very interesting because a lot of the things we've been looking at so far could uh in some respects be treated as like these kind of + +markovian processes of like step by step there's one update that largely just depends on what's going on currently and not on the past or history or any kind of assumed correlations between particular things and so with noise uh yeah i just wanted to comment that so far whenever we've added noise to to functions and kind of consider them to be you know probabilistic rather than purely deterministic these are all things we've already seen uh in part one of the book + +But whenever we were doing that with noise, we were assuming that the noise was itself Gaussian or like this kind of white noise, something like put in a sentence sort of like an ideally like, you know, random noise. + +There's no. + +patterns in the noise that are assumed to be there. + +And that's why we end up with these sort of like, you know, whenever we plot the linear line of how observations are generated from states, then the observation noise is just sort of scattered around the line, no particular pattern. + +But now we're moving into colored noise. + +And so what happens is that we start assuming that actually there might be some sort of temporal correlation in the noise, that there's some kind of pattern within the noise itself. + +So this kind of renders noise to be a potentially useful aspect for the model to sort of exploit for making predictions. + +As opposed to just assuming, oh, this is just something that is always going to minutely throw off probably what would otherwise be a perfect relationship between my predictions in my internal model versus what I think is going on in the world. + +So, yeah, it's all to say that it's a very interesting thing, and that's where we start seeing these sort of covariance matrices and the rest and looking at how noise relates to itself. + +And then finally, some aspects of this are not purely just, oh, let's add another thing to the model to make it better at performance. + +We do see in a lot of areas of neuroscience and other forms of time series analysis, different kinds of data, that whenever there's temporal correlation in the noise, that maybe helps us better understand what's really going on. + +um in in reality and uh it's super fascinating how the the brain itself like noise is one of the things that neuroscientists have attempted the most to try and understand over time because it's within noise that we start seeing these patterns and like uh you know the the kind of frequencies of neural activity that that are generated that correspond or correlate to different kinds of + +behavioral decision-making that an organism shows or otherwise. + +So noise itself, previously we've been treating noise basically as this thing that kind of throws our model off a little bit, but as long as we can get a good enough handle on that, then it's fine. + +but now suddenly noise is this kind of colored temporally rich and and covariating thing so just that's another aspect of this chapter is like like taking noise seriously rather than it's just some messy stuff that throws things off yeah great comment and fraser wrote i think section 6.1 6.2 6.3 are the theoretical core of the chapter i would agree six four and six five + + +SPEAKER_02: +are going to take that state space generative model that was of a single variable and put it into generalized coordinates of motion, which are the first derivative, second derivative, and higher order derivatives, leading to some generalizations about how different kinds of temporal dynamics in signal and noise can be brought into this model. + +Fraser's right though that in the first three sections it's a focus on a variable at the sort of most core filtering layer um without going into the generalized coordinates which gets sort of layered into the same method but just as a different way to encode temporal information + + +SPEAKER_05: +Yeah, just really quickly, I think the takeaways for this chapter, the core ideas are this idea of state dynamics, number one, that you can say is going to be changing. + +Number two is this idea that we're going to have a very specific kind of approximation, namely the Laplace approximation. + +I think that's number two. + +And that is used throughout this chapter. + +But I do think it needs to be stressed that that is actually quite an important + +core sort of thing that we're going to, that shows up in this chapter and will show up in other chapters as well. + +That's an assumption about the nature of the variational posterior. + +And then number three would be this idea of generalized coordinates of motion. + +I think I suspect those are the three kind of big + +ideas that you would do well to try and grok um throughout this this chapter and then there's a whole bunch of other discussion about implications of this and various you know multivariate this is univariate things but i think that those are the core takeaways to my mind at least yeah that's a great comment + + +SPEAKER_02: +State dynamics, hidden states can be changing. + +There are different approximation methods, like the Laplace approximation, which is long known. + +And then there's this different approximation approach. + +The mean field that has some different functions or some different analytical relationships or software implementations or abilities, like, to be run incrementally. + +Using a update learning. + +Function then there's the generalized coordinates of motion. + +which is one particular way to, a kind of physics informed way, like position, speed, acceleration, that sort of physics derivatives. + + +SPEAKER_05: +I was just going to say, for those interested in, there are a lot of people interested in implementation and experimentation with the code. + +Obviously, we're going to do our best to implement some things ourselves. + +But example 6.1, univariate Gaussian filtering, and then 6.2, multivariate Gaussian filtering, are extremely fundamental examples that will show up again and again and again throughout the entire book, actually. + +stuff to them and things like that. + +It's not quite that simple, but they are very canonical. + +So if you were looking for like bang for buck examples to study and even to implement, I mean, Sanjeev himself says you would do well to maybe play around with trying to implement, you know, at least 6.1, you know, the univariate case, that's a bit easier in code. + +So I think that's something that I'll endeavor to create maybe in a Google Colab or something like that. + +put it in the examples repo. + +Because you could just spend forever playing around with various parameters and things like that and learn by the bit. + +Because that's going to be the foundation of chapter 7, really, active generalized filtering. + +We're going to do this whole story again, but just by putting action into the mix. + +And then we're going to have a whole bunch of different things we can look at once we do that. + +So 6.1, 6.2, they're very, very, very fundamental. + + +SPEAKER_02: +Great comment. + +And the multivariate is in this case, like, and in many related kind of statistical cases or algorithmic cases, is like a vectorized or parallel version of the univariate case. + +And then there's sort of a simpler multivariate, which is just doing all the operations in parallel, but they don't interact with each other. + +And then there's more of like a joint multivariate where you also are looking at relationships amongst different variables. + +But whatever you do with multivariate, however much you use information across the different variables, it comes down to what is the one variable doing? + +And figure 6.4 shows basically what is happening, is there's some X star true latent variable and a belief about that latent variable, the mean centeredness of the belief about that variable. + +Then there's an observation sequence in the blue. + +and a belief about an observation sequence in blue and uh the x-axis is happening through time and the initial values of the beliefs are set to 15 and 13 which are too high on both um but they really quickly drop in to to the area and then kind of follow along with this lag it's correlated + +but clearly informational tracking. + +And, + +You could assess whether this performance is good enough or what other kind of cousin models or transformations of data or other improvements could improve the error rate, like if you were refining the model. + +But this is just showing, this is kind of in character what these algorithms do, which is from a given starting point, which is the initial value problem, it can track onto and then continue to follow + +with some secondary questions about like time scales of learning and how to bring complex causal models into picture a loss functional driven observable model output setting where initially when it was way out of the range of observations the loss function was high + +And then a gradient was followed to lead to a number that, while it doesn't really have a value like in itself, like the fact that it's 21 or 2,600 isn't really what's relevant, but you can look at this loss training trace like you would look at the loss training trace of another kind of Bayesian model or a neural network. + +and use it to model or relate to other model like absolute fit parameters or model adequacy or comparative model performance in relation to these free energy functionals to add more like context because it is a low dimensional representation of what the model is doing. + +And it tells you that even if you're in a flat, you might be still way off base or totally working in the wrong distribution family, et cetera. + +But in this pedagogical example, it drops in and then stays like doing smoothing, filtering with this lagged, smoothed belief updating process that happens time point by time point. + +And you could imagine other sort of non-base ways to do this. + +Like you could do a sliding window and then just take the average of the last end data points and then divide by the new data point. + +Or you could do a linear regression + +against the most recent data points or you could fit a classical time series model with like autocorrelation or time series and recalculate it at some interval so there's a total and you could do the laplace approximation there are many ways to approach the the setting that this is described by in the equations + +And in this chapter, the focus is on with this, I wonder if it would be clearer if this branch were more, well, here's variational free energy, but Laplace encoded energy is only on this side. + +So I'm just wondering whether, + +Here it's the Laplace approximation, but when is the mean field approximation come in? + + +SPEAKER_05: +Yeah, I guess maybe the difference is that this is emphasizing the approximations we're making about the VFE itself, as opposed to our approximate posterior, which would be your mean field approximation there. + +Yes. + +At least that's kind of how I read it, that we're really concerned about the form of the VFE here specifically. + + +SPEAKER_02: +Yeah. + +Setting up the domain generative model + +so that the VFE is trainable representing what it's supposed to mean, like putting labels on images. + +It's like, then, let's see how Laplace encoded energy is used in 6.1, and if that's the same or related to the Laplace approximation. + +Or let's look at an example of 6.1. + +Good call. + + +SPEAKER_05: +And 6.1 is expressed in terms of the language of chapter 5, where we've got prediction errors and precision weighted prediction errors. + +We could equally well have not represented it that way. + +It just happens to be + +how we've chosen to represent things here. + +I think that that emphasizes that we do have two different kinds of errors that we can, that are apparent now. + +We obviously have our observation errors and state dynamic beliefs errors as well. + +So it does kind of help highlight that we now have this additional thing. + +Andrew, I think you've got to add up. + +Yeah. + + +SPEAKER_01: +Yeah, sorry. + +I don't want to detract from what was being said, but just like, + +I mean, the big thing here is that by incorporating this notion of state transitions, or there being a flow to states, there are different ways that we can sort of model that relationship. + +So, one way that we can do that, and it's frequently done in active inference and other things leading up to active inference, including this perception only sort of example that we're looking at in Chapter 6, is that the Laplace is essentially related to taking Taylor expansions. + +And so, you know, + +With the Laplace, it's specifically up to the second derivative. + +So you can sort of think of that because the equation is not precisely the same. + +But you can think of each state having its own position as well as its own speed, if it's moving up or down. + +over time right that's a temporal like we're actually thinking of states changing as opposed to just potentially being static and then finally the second derivative which is something like an acceleration we can liken it to that and so we're actually doing this for for many things we're not just doing it for states we're also doing it for observations we're doing it for noise um so so all of these things kind of come together + +But that's the big distinction here. + +It's just that the Laplace approximation is a particular way of modeling these sort of temporal derivatives and how something relates to itself in time. + +Yeah, there are other ways you could do it. + +It wouldn't have to be Laplace. + +You could go with a single derivative or like there are many other instances and like other writings on generalized coordinates of motion where we go up to like six different temporal derivatives, but the Laplace generally revolves around using two. + +And then you could kind of potentially modify that in different areas, depending on what you're trying to do or how you think your agent should be modeled. + + +SPEAKER_02: +That's a great point. + +The Laplace approximation can be run iteratively classically because it makes certain constrained assumptions, like by only considering up to the quadratic, not exploring optimization over + +what higher orders of motion, like higher derivatives of what the Taylor series describes are parameterized, which is like the general polynomial setting. + +But the Laplace approximation can just quickly scoot around a quadratic pretty similarly to how the variational distribution can scoot around a Gaussian. + +But I believe, and let's nuance this if this is not accurate, but the Gaussian, the variational Gaussian that we move around is a true statistical distribution. + +Whereas the quadratic is a heuristic, but it pierces down, and it's not necessarily a probability distribution itself. + +It's just like a central tendency and curvature estimate, very locally. + +Whereas the variational setting is more composable, and we're talking about real statistical variables, not just a local curvature. + +Verena? + + +SPEAKER_00: +I have a question about those smooth trajectories. + +So assuming that maybe we have to deal with discontinuous events or degraded information, how is this interpreted as a prediction error or does the generative model + + +SPEAKER_02: +need something like explicit change point dynamics or something like that i think a good example that explores that is um i'll put it in the chat too this is the rx infer example of a recurrent switching linear dynamical system so um there are + +two different discreetly switching settings and then there's noisy data coming in that are um being on a first order fit by a smoothing filtering and then at the same time the the latent state estimation is like am i in state one or state two with low very low volatility and a high volatility uh regime + +And then, uh, this is the estimates through time for that latent state. + +So that's, that's kind of one way to model that, like to include what was the discontinuity of, of one type or of, of a causal type into the same sensor by, by just, um, testing, like, should there be, um, + +one, two, three latent variables. + +So it kind of goes, okay, now we're going to have a prior over how many latent states there are. + +So it's kind of the process of going from building these hierarchical models that at the end of the day, like the thermometer or the microphone or the EEG sensor, it is going to be some continuous one by one stream of discrete data points. + +And then there are like well now we can take into account sequences of that's the whole building statistical models that that explore state spaces related to that primary filter in question pressure. + + +SPEAKER_05: +Yeah, just really quickly, maybe an example from my own work. + +I've got this robotic arm that has, you know, the states are the angles of the segments. + +Right. + +So and we have a belief about all of those angles. + +But I'm approximating that belief by taking a couple of samples and then I stick those samples through my state dynamics. + +but if you if if a link wraps around 180 degrees then some of those samples can wrap around and you don't know if you're at you know where you are or 180 degrees removed from now and that would be an example of a discontinuity so at the moment i just have to be careful to not ever bend + +all the way around. + +That would be one example of a discontinuity in my state dynamics that would cause me a lot of problems. + +It just so happens that I don't actually have to worry about it in practice for whatever reason most of the time. + +But those are big problems that do come up now. + +You have to worry about, okay, well, is there a discontinuity here? + +Is there some region where something isn't defined? + + +SPEAKER_03: +So that is definitely going to be the thing that's added to our list of concerns, basically. + + +SPEAKER_01: +Yeah, I just want to maybe provide another intuitive example, even though both of these are nice. + +So imagine, you know, someone who is a catcher in baseball, or basically someone who has to catch a ball for whatever reason. + +You know, maybe they're just having fun playing catch. + +So in previous sort of settings that we've seen in the textbook prior to Chapter 6, we have models that are probabilistic Bayesian models that would be able to predict potentially where a ball is at in time. + +um but in order to actually be able to make a full prediction it would need to be able to see the entire trajectory of the ball coming towards the the catcher the person who's trying to predict where the ball is who wants to catch it they would have to like see and advance the entire trajectory and then even then + +there's nothing in the model that's saying that that trajectory is specifically a trajectory in time that there's some kind of relationship between where the ball is at in time step one versus time step four so it's just a bunch of these like kind of static observations that the the agent is getting um whenever we provide for temporality in the way that we are + +Uh, here, this is where we're able to have an agent who says, oh, well, the ball is here. + +But it was here just prior and because I'm using these temporal derivatives, I'm able to sort of approximate estimate predict, not just precisely where the ball is, but. + +what is its current speed and direction, sort of acceleration, and potentially more if you include more terms. + +And so I hope that that gives a sense of like, we're able to actually establish sort of a trajectory that allows for an expectation of where the ball will be next. + +Because if we just stay in this static world without a sense of time, + +then it would be very surprising to us to see that the ball suddenly changed position from one time step to the next. + +We have no sense of how that could happen if we don't include some sense of flow or temporal dynamics into our model. + + +SPEAKER_00: +Thank you. + +I was thinking about cochlear implant recipients. + +who have a very degraded signal input over average 22 electrodes and to actually become the full you know voice or the full information over the words takes time through learning and this is not a very linear process so that was the reason why i was asking + + +SPEAKER_02: +Yeah, just as like one place to explore, but I don't think we'll cover it, is that it's the hearing active inference setting for DeVry, which is related to the RX and for streaming real time. + +So definitely like the audio processing and hearing is related to what people are working on. + +Okay, I'm going to stop the recording there. diff --git a/docs/SCHEMA.md b/docs/SCHEMA.md index e39a1bede..822fc228b 100644 --- a/docs/SCHEMA.md +++ b/docs/SCHEMA.md @@ -98,6 +98,11 @@ corresponding display fields without losing the source text: - **Split transcript identity:** when source transcripts are split into sessions, corresponding `transcript.txt` headings and `transcript.json` `video_id` fields use the stable `_sessNN` session name rather than an empty placeholder. +- **`parts[].speakers`** (journal-owned): human speaker identifications for + WhisperX-diarized transcripts — `{"SPEAKER_NN": "Name", ...}` per video. + Record the mapping here, then run + `Journal-Utilities/scripts/apply_speaker_names.py` to replace the labels in + transcript.txt/.json (idempotent; unmapped labels stay SPEAKER_NN). - **`duplicate_of`** marks an `Other/` item whose content is a duplicate of a curated item (the two symposium full uploads). The duplicate intentionally repeats the canonical video's ID; coverage reconciliation From 9e633ca748f80bd9e177253462e4a9987d4e1d32 Mon Sep 17 00:00:00 2001 From: Holly Grimm Date: Fri, 17 Jul 2026 16:38:15 -0600 Subject: [PATCH 2/4] data: speaker names for Session_023 (44 labels, 7 speakers) Fraser Paterson (two diarization labels), Magdalena Hurtado, Marc Broberg, Andrew Pashea, Jim DeLong, Dorsa, Sun Xin; one audience questioner (SPEAKER_00) still unidentified. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_015NjFiMWkgrVy4foGF1Zbax --- .../Cohort_1/Session_023/metadata.json | 12 ++- .../Cohort_1/Session_023/transcript.json | 2 +- .../Cohort_1/Session_023/transcript.txt | 88 +++++++++---------- 3 files changed, 56 insertions(+), 46 deletions(-) diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json index 34f3954f1..1d4f7a5dd 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json @@ -15,7 +15,17 @@ "url": "https://www.youtube.com/watch?v=MjeQeWeyYhE", "title": "Fundamentals of Active Inference (Part 1 review, Session 23) June 30, 2026", "duration": null, - "upload_date": "" + "upload_date": "", + "speakers": { + "SPEAKER_05": "Fraser Paterson", + "SPEAKER_01": "Magdalena Hurtado", + "SPEAKER_02": "Marc Broberg", + "SPEAKER_09": "Andrew Pashea", + "SPEAKER_03": "Jim DeLong", + "SPEAKER_04": "Fraser Paterson", + "SPEAKER_07": "Dorsa", + "SPEAKER_08": "Sun Xin" + } } ] } diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json index bd6cacdf1..1f0b4ce89 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json @@ -1 +1 @@ -[{"video_id": "MjeQeWeyYhE", "segments": [{"start": 3.727, "end": 29.498, "text": " all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense", "speaker": "SPEAKER_05"}, {"start": 29.798, "end": 40.317, "text": " And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two.", "speaker": "SPEAKER_05"}, {"start": 40.337, "end": 43.923, "text": "We're going to do active inputs proper, which is going to be very, very exciting going forward.", "speaker": "SPEAKER_05"}, {"start": 43.943, "end": 50.575, "text": "So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts.", "speaker": "SPEAKER_05"}, {"start": 51.027, "end": 56.833, "text": " that we're going to start doing actual, quote unquote, active inference in the second part.", "speaker": "SPEAKER_05"}, {"start": 56.853, "end": 67.444, "text": "But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection.", "speaker": "SPEAKER_05"}, {"start": 68.305, "end": 72.409, "text": "We're going to sort of slow down and think about part one in general, chapters one to five.", "speaker": "SPEAKER_05"}, {"start": 73.531, "end": 80.578, "text": "This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one.", "speaker": "SPEAKER_05"}, {"start": 80.812, "end": 85.079, "text": " you know, going so that we can be adequately grounded going forward.", "speaker": "SPEAKER_05"}, {"start": 85.981, "end": 94.696, "text": "I will say we have one one question is, you know, we've done each chapter across two weeks.", "speaker": "SPEAKER_05"}, {"start": 95.658, "end": 100.947, "text": "Historically, you know, chapter one, we've had two weeks going forward.", "speaker": "SPEAKER_05"}, {"start": 101.72, "end": 128.115, "text": " depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week", "speaker": "SPEAKER_05"}, {"start": 128.095, "end": 152.685, "text": " of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe", "speaker": "SPEAKER_05"}, {"start": 152.935, "end": 159.784, "text": " It would be a good idea to reach out through the blankets email.", "speaker": "SPEAKER_05"}, {"start": 160.425, "end": 166.993, "text": "Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place.", "speaker": "SPEAKER_05"}, {"start": 168.415, "end": 169.296, "text": "That would be amazing.", "speaker": "SPEAKER_05"}, {"start": 169.416, "end": 170.057, "text": "Good idea.", "speaker": "SPEAKER_05"}, {"start": 170.698, "end": 171.198, "text": "Yes, please.", "speaker": "SPEAKER_05"}, {"start": 171.238, "end": 176.445, "text": "I see a lot of people are saying yes.", "speaker": "SPEAKER_05"}, {"start": 176.695, "end": 182.461, "text": " In this chat here, it would be very helpful if you could give your yay or nay to that as well.", "speaker": "SPEAKER_05"}, {"start": 182.581, "end": 185.384, "text": "So we can actually look at that just directly through the chat here.", "speaker": "SPEAKER_05"}, {"start": 185.464, "end": 188.848, "text": "So if you want a second week review, yes.", "speaker": "SPEAKER_05"}, {"start": 189.108, "end": 190.089, "text": "If you don't, no.", "speaker": "SPEAKER_05"}, {"start": 190.71, "end": 192.252, "text": "And then we can go forward.", "speaker": "SPEAKER_05"}, {"start": 193.553, "end": 193.893, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 195.495, "end": 196.096, "text": "Enough of that.", "speaker": "SPEAKER_05"}, {"start": 196.616, "end": 197.898, "text": "I'll start sharing my screen here.", "speaker": "SPEAKER_05"}, {"start": 199.259, "end": 200.28, "text": "Probably my entire screen.", "speaker": "SPEAKER_04"}, {"start": 204.084, "end": 204.505, "text": "Okie dokie.", "speaker": "SPEAKER_05"}, {"start": 204.525, "end": 205.446, "text": "So people should be able to see.", "speaker": "SPEAKER_05"}, {"start": 205.486, "end": 206.567, "text": "Maybe just get rid of...", "speaker": "SPEAKER_05"}, {"start": 207.357, "end": 214.009, "text": " The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise.", "speaker": "SPEAKER_05"}, {"start": 215.011, "end": 227.172, "text": "about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here.", "speaker": "SPEAKER_05"}, {"start": 227.928, "end": 229.152, "text": " So you've got your overview.", "speaker": "SPEAKER_05"}, {"start": 229.935, "end": 232.945, "text": "I've added one or two additional resources.", "speaker": "SPEAKER_05"}, {"start": 232.965, "end": 233.827, "text": "So these actually came up.", "speaker": "SPEAKER_05"}, {"start": 233.908, "end": 235.593, "text": "Andrew shared the first of these.", "speaker": "SPEAKER_05"}, {"start": 235.613, "end": 236.436, "text": "These are all videos.", "speaker": "SPEAKER_05"}, {"start": 237.239, "end": 238.242, "text": "One of them is an active inference", "speaker": "SPEAKER_05"}, {"start": 240.668, "end": 241.791, "text": " coding and active inference.", "speaker": "SPEAKER_05"}, {"start": 243.094, "end": 245.139, "text": "So that's Ryan Smith and his team.", "speaker": "SPEAKER_05"}, {"start": 245.841, "end": 248.227, "text": "They've done excellent work on predictive coding.", "speaker": "SPEAKER_05"}, {"start": 248.708, "end": 249.45, "text": "I'll play that just now.", "speaker": "SPEAKER_05"}, {"start": 249.991, "end": 258.171, "text": "This is a great stream just about predictive coding more generally and its relation to active inference and its relation to", "speaker": "SPEAKER_05"}, {"start": 258.151, "end": 260.855, "text": " you know, its status in actual brains.", "speaker": "SPEAKER_05"}, {"start": 261.616, "end": 265.781, "text": "So if you're interested in more context, this would be excellent to go through.", "speaker": "SPEAKER_05"}, {"start": 265.801, "end": 267.003, "text": "And there's two more links there as well.", "speaker": "SPEAKER_05"}, {"start": 267.043, "end": 271.99, "text": "One by Jakob Howie, he's a philosopher over in Australia in the University of Monash.", "speaker": "SPEAKER_05"}, {"start": 272.01, "end": 274.013, "text": "What is predictive processing and what is it good for?", "speaker": "SPEAKER_05"}, {"start": 274.133, "end": 274.914, "text": "Another excellent talk.", "speaker": "SPEAKER_05"}, {"start": 275.495, "end": 280.041, "text": "And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke.", "speaker": "SPEAKER_05"}, {"start": 280.325, "end": 281.426, "text": " from the University of Toronto.", "speaker": "SPEAKER_05"}, {"start": 281.446, "end": 286.772, "text": "He's actually speaking about Neoplatonism and mystical experiences, believe it or not.", "speaker": "SPEAKER_05"}, {"start": 287.513, "end": 289.135, "text": "And I know it sounds weird.", "speaker": "SPEAKER_05"}, {"start": 289.575, "end": 292.799, "text": "Predictive processing actually shows up in here in a very significant way.", "speaker": "SPEAKER_05"}, {"start": 292.819, "end": 297.084, "text": "So that might be a nice mood setter, as it were.", "speaker": "SPEAKER_05"}, {"start": 297.164, "end": 298.365, "text": "So there's some additional resources there.", "speaker": "SPEAKER_05"}, {"start": 298.686, "end": 301.128, "text": "I've begun filling in the map for chapter five.", "speaker": "SPEAKER_05"}, {"start": 301.949, "end": 307.075, "text": "So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid.", "speaker": "SPEAKER_05"}, {"start": 307.73, "end": 334.446, "text": " the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that", "speaker": "SPEAKER_05"}, {"start": 334.814, "end": 338.84, "text": " A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation.", "speaker": "SPEAKER_05"}, {"start": 339.401, "end": 353.881, "text": "What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you.", "speaker": "SPEAKER_05"}, {"start": 354.322, "end": 359.269, "text": "You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you.", "speaker": "SPEAKER_05"}, {"start": 359.249, "end": 363.355, "text": " So it's not ideal, but the content is there at least.", "speaker": "SPEAKER_05"}, {"start": 363.736, "end": 367.081, "text": "And especially for those who don't have the book, I think this is quite useful.", "speaker": "SPEAKER_05"}, {"start": 367.101, "end": 374.753, "text": "So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter.", "speaker": "SPEAKER_05"}, {"start": 375.274, "end": 382.405, "text": "I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts", "speaker": "SPEAKER_05"}, {"start": 382.975, "end": 403.431, "text": " uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there", "speaker": "SPEAKER_05"}, {"start": 404.373, "end": 404.954, "text": " Any questions?", "speaker": "SPEAKER_05"}, {"start": 404.994, "end": 416.831, "text": "Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that.", "speaker": "SPEAKER_05"}, {"start": 418.033, "end": 419.795, "text": "I'll start my share temporarily.", "speaker": "SPEAKER_05"}, {"start": 420.756, "end": 421.858, "text": "Just coming back.", "speaker": "SPEAKER_05"}, {"start": 423.863, "end": 424.404, "text": " Very good.", "speaker": "SPEAKER_05"}, {"start": 425.745, "end": 426.386, "text": "Yes, but it's limited.", "speaker": "SPEAKER_05"}, {"start": 428.008, "end": 430.991, "text": "Formatting in LaTeX is a bit strange in Coda.", "speaker": "SPEAKER_05"}, {"start": 432.032, "end": 442.664, "text": "We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda.", "speaker": "SPEAKER_05"}, {"start": 444.026, "end": 445.928, "text": "OK, so I'll share, come back.", "speaker": "SPEAKER_05"}, {"start": 447.89, "end": 452.976, "text": "Let's go all the way back, hopefully people can see my screen again, to the introduction.", "speaker": "SPEAKER_05"}, {"start": 454.441, "end": 456.885, "text": " So everyone's seen this figure here.", "speaker": "SPEAKER_05"}, {"start": 456.905, "end": 458.408, "text": "This is essentially the first figure in the book.", "speaker": "SPEAKER_05"}, {"start": 458.849, "end": 461.814, "text": "This is the breakdown of the partitions of the book, so part one, part two, part three.", "speaker": "SPEAKER_05"}, {"start": 462.396, "end": 468.046, "text": "We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding.", "speaker": "SPEAKER_05"}, {"start": 469.649, "end": 471.232, "text": "Part two is active inference core.", "speaker": "SPEAKER_05"}, {"start": 471.252, "end": 473.215, "text": "So this is the heart of the book, really.", "speaker": "SPEAKER_05"}, {"start": 473.255, "end": 476.922, "text": "This is the fundamentals of active inference per se.", "speaker": "SPEAKER_05"}, {"start": 476.902, "end": 498.64, "text": " what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end", "speaker": "SPEAKER_05"}, {"start": 499.328, "end": 504.875, "text": " But as has been the case, everything has been just perception only.", "speaker": "SPEAKER_05"}, {"start": 504.895, "end": 506.057, "text": "So we haven't actually dealt with actions.", "speaker": "SPEAKER_05"}, {"start": 506.437, "end": 508.24, "text": "And we're going to be moving into that with part two.", "speaker": "SPEAKER_05"}, {"start": 509.221, "end": 511.564, "text": "But even more fundamentally, everything has been static.", "speaker": "SPEAKER_05"}, {"start": 511.824, "end": 516.651, "text": "So all of the models we've been looking at, I don't know if I can get a nice example.", "speaker": "SPEAKER_05"}, {"start": 517.592, "end": 519.014, "text": "All of the models, so this is chapter one.", "speaker": "SPEAKER_05"}, {"start": 519.995, "end": 523.159, "text": "Let's maybe go down to some of the equations.", "speaker": "SPEAKER_05"}, {"start": 523.443, "end": 524.244, "text": " They've all been static.", "speaker": "SPEAKER_05"}, {"start": 524.264, "end": 527.387, "text": "So the environment hasn't really been changing at all.", "speaker": "SPEAKER_05"}, {"start": 527.407, "end": 530.371, "text": "And this is very simple.", "speaker": "SPEAKER_05"}, {"start": 531.372, "end": 534.235, "text": "So part one, really going to chapter two, I think.", "speaker": "SPEAKER_05"}, {"start": 536.858, "end": 538.88, "text": "So we have an environment.", "speaker": "SPEAKER_05"}, {"start": 539.501, "end": 541.503, "text": "This is one we've seen this many times now.", "speaker": "SPEAKER_05"}, {"start": 541.523, "end": 548.691, "text": "This is a representation of the generative process, the real process in and of itself out there.", "speaker": "SPEAKER_05"}, {"start": 549.262, "end": 553.407, "text": " This is fairly static, so nothing is really changing.", "speaker": "SPEAKER_05"}, {"start": 553.467, "end": 555.609, "text": "The hidden states aren't really changing from moment to moment.", "speaker": "SPEAKER_05"}, {"start": 555.629, "end": 562.397, "text": "We're just presented with a situation and we have to update our beliefs about what hidden states might be.", "speaker": "SPEAKER_05"}, {"start": 562.437, "end": 564.199, "text": "That's been constant throughout all of Chapter 1.", "speaker": "SPEAKER_05"}, {"start": 564.879, "end": 566.721, "text": "Things have gotten progressively more complicated.", "speaker": "SPEAKER_05"}, {"start": 566.741, "end": 570.906, "text": "We've looked at univariate hidden states where there's literally just one hidden state.", "speaker": "SPEAKER_05"}, {"start": 571.567, "end": 577.193, "text": "We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things.", "speaker": "SPEAKER_05"}, {"start": 577.662, "end": 579.304, "text": " But in every case, it's been static.", "speaker": "SPEAKER_05"}, {"start": 579.884, "end": 585.331, "text": "The dynamics of the hidden state have been non-existent.", "speaker": "SPEAKER_05"}, {"start": 585.351, "end": 586.992, "text": "That is going to change going into part two.", "speaker": "SPEAKER_05"}, {"start": 587.052, "end": 591.918, "text": "We're going to start looking at generative processes which change over time.", "speaker": "SPEAKER_05"}, {"start": 592.579, "end": 593.94, "text": "And they're the interesting ones.", "speaker": "SPEAKER_05"}, {"start": 593.98, "end": 596.283, "text": "They're the ones that are actually useful.", "speaker": "SPEAKER_05"}, {"start": 596.343, "end": 602.55, "text": "And they're the ones that allow us to actually begin to have a reason to act in the environment.", "speaker": "SPEAKER_05"}, {"start": 602.61, "end": 605.593, "text": "We haven't really had any reason to perform actions yet.", "speaker": "SPEAKER_05"}, {"start": 606.805, "end": 611.032, "text": " So that's a big move that's going to take place and that we're going to have to deal with.", "speaker": "SPEAKER_05"}, {"start": 612.074, "end": 613.517, "text": "So maybe just coming back.", "speaker": "SPEAKER_05"}, {"start": 613.557, "end": 618.325, "text": "So chapter one was about, what was this about?", "speaker": "SPEAKER_05"}, {"start": 618.345, "end": 619.187, "text": "It was about perception.", "speaker": "SPEAKER_05"}, {"start": 619.427, "end": 619.588, "text": "Okay.", "speaker": "SPEAKER_05"}, {"start": 619.628, "end": 624.176, "text": "So like in terms of how we're going to think about perception in active inference.", "speaker": "SPEAKER_05"}, {"start": 624.757, "end": 627.261, "text": "So in chapter one, we cast or we framed", "speaker": "SPEAKER_05"}, {"start": 630.565, "end": 654.058, "text": " of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at", "speaker": "SPEAKER_05"}, {"start": 655.118, "end": 661.187, "text": " you know, how to do like exact Bayesian inference, okay, if we were able to do everything.", "speaker": "SPEAKER_05"}, {"start": 663.11, "end": 665.213, "text": "We saw how to use Bayes' rule basically.", "speaker": "SPEAKER_05"}, {"start": 665.233, "end": 675.187, "text": "So we cemented the distinction personally before that around the difference between the generative process and the generative model.", "speaker": "SPEAKER_05"}, {"start": 675.989, "end": 679.354, "text": "So I come all the way down to the figures here.", "speaker": "SPEAKER_05"}, {"start": 679.374, "end": 681.076, "text": "I don't have a particularly good figure for that.", "speaker": "SPEAKER_05"}, {"start": 681.096, "end": 682.358, "text": "Here we go, 2.4.", "speaker": "SPEAKER_05"}, {"start": 682.793, "end": 691.766, "text": " The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world.", "speaker": "SPEAKER_05"}, {"start": 692.307, "end": 695.191, "text": "This was the crucial thing, I think, in chapter two, basically.", "speaker": "SPEAKER_05"}, {"start": 695.992, "end": 699.978, "text": "And we have a way of notating this convention.", "speaker": "SPEAKER_05"}, {"start": 699.998, "end": 708.531, "text": "We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah.", "speaker": "SPEAKER_05"}, {"start": 708.971, "end": 712.176, "text": "These belong to the generative process, the real world out there.", "speaker": "SPEAKER_05"}, {"start": 712.797, "end": 716.822, "text": " Okay, and then modeled variables are without a star.", "speaker": "SPEAKER_05"}, {"start": 717.142, "end": 719.184, "text": "So they correspond to our model.", "speaker": "SPEAKER_05"}, {"start": 720.746, "end": 721.467, "text": "And what is our model?", "speaker": "SPEAKER_05"}, {"start": 722.668, "end": 730.037, "text": "It is quite literally a joint probability distribution over hidden states, latent states, and observations.", "speaker": "SPEAKER_05"}, {"start": 730.978, "end": 733.421, "text": "We use x for latent states and y for observations.", "speaker": "SPEAKER_05"}, {"start": 734.722, "end": 742.551, "text": "But you can see, of course, the x in here is our model of the real latent state outside of here, x star.", "speaker": "SPEAKER_05"}, {"start": 743.172, "end": 744.755, "text": " And they need not agree with each other.", "speaker": "SPEAKER_05"}, {"start": 744.795, "end": 745.737, "text": "They need not be the same.", "speaker": "SPEAKER_05"}, {"start": 746.839, "end": 752.69, "text": "But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be.", "speaker": "SPEAKER_05"}, {"start": 752.731, "end": 754.795, "text": "That is p of x and y.", "speaker": "SPEAKER_05"}, {"start": 755.817, "end": 760.185, "text": "And then if we have the model evidence, the probability of observing anything,", "speaker": "SPEAKER_05"}, {"start": 760.722, "end": 769.453, "text": " or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation.", "speaker": "SPEAKER_05"}, {"start": 770.174, "end": 772.037, "text": "And that was kind of chapter two.", "speaker": "SPEAKER_05"}, {"start": 772.057, "end": 779.526, "text": "And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions.", "speaker": "SPEAKER_05"}, {"start": 779.546, "end": 780.648, "text": "And these are very ubiquitous.", "speaker": "SPEAKER_05"}, {"start": 780.668, "end": 781.369, "text": "They show up everywhere.", "speaker": "SPEAKER_05"}, {"start": 782.43, "end": 783.331, "text": "We're going to continue that.", "speaker": "SPEAKER_05"}, {"start": 783.351, "end": 790.24, "text": "In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference.", "speaker": "SPEAKER_05"}, {"start": 790.642, "end": 791.784, "text": " There are other distributions as well.", "speaker": "SPEAKER_05"}, {"start": 791.804, "end": 792.725, "text": "We haven't really looked at those yet.", "speaker": "SPEAKER_05"}, {"start": 793.967, "end": 799.554, "text": "But the real sort of, there's a problem with this, which is that, yes, we can do inference.", "speaker": "SPEAKER_05"}, {"start": 799.574, "end": 803.079, "text": "Yes, we might even be able to do exact inference for really, really simple problems.", "speaker": "SPEAKER_05"}, {"start": 804.14, "end": 812.952, "text": "But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model.", "speaker": "SPEAKER_05"}, {"start": 813.793, "end": 815.075, "text": "And then we're going to do inference.", "speaker": "SPEAKER_05"}, {"start": 815.095, "end": 820.142, "text": "And inference is like, I set the dials, and then my machine does some inference stuff, right?", "speaker": "SPEAKER_05"}, {"start": 820.983, "end": 847.716, "text": " that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation", "speaker": "SPEAKER_05"}, {"start": 848.303, "end": 850.966, "text": " And we're doing good old fashioned Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 851.346, "end": 854.99, "text": "But now we're having to do parameter learning and estimation as well.", "speaker": "SPEAKER_05"}, {"start": 856.692, "end": 867.903, "text": "And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference.", "speaker": "SPEAKER_05"}, {"start": 867.923, "end": 872.448, "text": "So we have these two processes, inference and learning parameter estimation.", "speaker": "SPEAKER_05"}, {"start": 873.109, "end": 877.213, "text": "And they're both able to be done by means of Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 878.56, "end": 879.421, "text": " That's the crucial thing.", "speaker": "SPEAKER_05"}, {"start": 879.441, "end": 889.438, "text": "So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah.", "speaker": "SPEAKER_05"}, {"start": 889.458, "end": 904.163, "text": "And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent.", "speaker": "SPEAKER_05"}, {"start": 904.868, "end": 908.953, "text": " And we saw many versions of this across the chapter and indeed in other chapters.", "speaker": "SPEAKER_05"}, {"start": 908.993, "end": 920.085, "text": "It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters.", "speaker": "SPEAKER_05"}, {"start": 920.185, "end": 926.172, "text": "So let's imagine we have our two parameters, I don't know, beta 1, beta 0.", "speaker": "SPEAKER_05"}, {"start": 926.232, "end": 934.521, "text": "And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space.", "speaker": "SPEAKER_05"}, {"start": 934.872, "end": 935.733, "text": " That's this surface.", "speaker": "SPEAKER_05"}, {"start": 936.754, "end": 944.063, "text": "And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point.", "speaker": "SPEAKER_05"}, {"start": 945.064, "end": 950.271, "text": "And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters.", "speaker": "SPEAKER_05"}, {"start": 950.291, "end": 952.654, "text": "Very general way of doing optimization.", "speaker": "SPEAKER_05"}, {"start": 954.015, "end": 964.288, "text": "However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be.", "speaker": "SPEAKER_05"}, {"start": 965.72, "end": 973.231, "text": " And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization.", "speaker": "SPEAKER_05"}, {"start": 975.574, "end": 985.308, "text": "So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now.", "speaker": "SPEAKER_05"}, {"start": 986.25, "end": 993.921, "text": "A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding.", "speaker": "SPEAKER_05"}, {"start": 994.357, "end": 1021.09, "text": " um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model", "speaker": "SPEAKER_05"}, {"start": 1021.525, "end": 1025.652, "text": " And also, how should I use those settings to do good inference?", "speaker": "SPEAKER_05"}, {"start": 1026.934, "end": 1040.558, "text": "It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters.", "speaker": "SPEAKER_05"}, {"start": 1040.578, "end": 1042.442, "text": "So imagine you've got learning up here.", "speaker": "SPEAKER_05"}, {"start": 1042.582, "end": 1043.744, "text": "What should the model parameters be?", "speaker": "SPEAKER_05"}, {"start": 1043.764, "end": 1046.048, "text": "And you can do inference on model parameters.", "speaker": "SPEAKER_05"}, {"start": 1046.45, "end": 1050.538, "text": " And then that can inform how you do inference at the low-level hidden states.", "speaker": "SPEAKER_05"}, {"start": 1050.558, "end": 1053.805, "text": "So there's kind of these two processes that are happening in two different timescales.", "speaker": "SPEAKER_05"}, {"start": 1053.825, "end": 1062.062, "text": "That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally.", "speaker": "SPEAKER_05"}, {"start": 1063.122, "end": 1064.023, "text": " So that's kind of chapter three.", "speaker": "SPEAKER_05"}, {"start": 1064.143, "end": 1065.264, "text": "I'm going to race through this.", "speaker": "SPEAKER_05"}, {"start": 1065.925, "end": 1068.828, "text": "And we'll get to chapter five and then to some actual questions from you guys.", "speaker": "SPEAKER_05"}, {"start": 1069.749, "end": 1081.062, "text": "Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference.", "speaker": "SPEAKER_05"}, {"start": 1081.082, "end": 1084.966, "text": "So chapter three, we're still doing exact inference.", "speaker": "SPEAKER_05"}, {"start": 1084.986, "end": 1090.632, "text": "But we saw that really we culminated in an algorithm", "speaker": "SPEAKER_05"}, {"start": 1091.658, "end": 1113.632, "text": " allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be", "speaker": "SPEAKER_05"}, {"start": 1113.95, "end": 1115.672, "text": " Those two problems are related to one another.", "speaker": "SPEAKER_05"}, {"start": 1115.752, "end": 1118.615, "text": "To know the hidden states, you need to know what the good parameters are.", "speaker": "SPEAKER_05"}, {"start": 1118.935, "end": 1121.378, "text": "But to know what the good parameters are, you have to know how to infer hidden states well.", "speaker": "SPEAKER_05"}, {"start": 1121.958, "end": 1131.588, "text": "So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm.", "speaker": "SPEAKER_05"}, {"start": 1131.608, "end": 1132.689, "text": "I won't go over that again here.", "speaker": "SPEAKER_05"}, {"start": 1132.73, "end": 1134.131, "text": "Maybe we will if people want me to.", "speaker": "SPEAKER_05"}, {"start": 1135.072, "end": 1139.957, "text": "But that was the solution to the problem of chapter 3, that chicken and egg problem.", "speaker": "SPEAKER_05"}, {"start": 1140.781, "end": 1151.908, "text": " The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations.", "speaker": "SPEAKER_05"}, {"start": 1151.928, "end": 1155.938, "text": "And that, unfortunately, is usually not something we can do.", "speaker": "SPEAKER_05"}, {"start": 1156.542, "end": 1161.508, "text": " Being able to find the exact posterior is usually impossible, for reasons that we saw.", "speaker": "SPEAKER_05"}, {"start": 1161.708, "end": 1162.75, "text": "We can review again if you want.", "speaker": "SPEAKER_05"}, {"start": 1163.33, "end": 1166.895, "text": "So then chapter 4 says, all right, well, we can't do exact Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 1167.536, "end": 1168.096, "text": "What are we going to do?", "speaker": "SPEAKER_05"}, {"start": 1168.116, "end": 1177.107, "text": "We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like.", "speaker": "SPEAKER_05"}, {"start": 1178.029, "end": 1180.812, "text": "So the idea with chapter 4 is, ah, OK, we're going to now", "speaker": "SPEAKER_05"}, {"start": 1181.804, "end": 1193.779, "text": " have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior.", "speaker": "SPEAKER_05"}, {"start": 1194.7, "end": 1198.304, "text": "We're going to try and find a posterior that's close enough to the true posterior.", "speaker": "SPEAKER_05"}, {"start": 1198.705, "end": 1208.537, "text": "And this then motivated the idea of variational free energy as a quantity that is a measurement of how well", "speaker": "SPEAKER_05"}, {"start": 1208.838, "end": 1213.784, "text": " an approximate posterior fits or how close it is to the true posterior.", "speaker": "SPEAKER_05"}, {"start": 1214.885, "end": 1222.794, "text": "And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal.", "speaker": "SPEAKER_05"}, {"start": 1223.055, "end": 1225.858, "text": "We saw from chapter three that surprisal is something we want to make very small.", "speaker": "SPEAKER_05"}, {"start": 1226.719, "end": 1237.191, "text": "So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal,", "speaker": "SPEAKER_05"}, {"start": 1237.778, "end": 1242.586, "text": " which is something we can't do directly, and therefore we can do approximate Bayesian inference.", "speaker": "SPEAKER_05"}, {"start": 1242.606, "end": 1248.976, "text": "So in a lot of ways, that really kind of spiritually is the end of part one, I would say.", "speaker": "SPEAKER_05"}, {"start": 1249.958, "end": 1265.102, "text": "The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE.", "speaker": "SPEAKER_05"}, {"start": 1265.842, "end": 1269.509, "text": " That's, in my mind, really kind of the end of part one.", "speaker": "SPEAKER_05"}, {"start": 1270.311, "end": 1278.908, "text": "Chapter five was a nice kind of detour into a different way of thinking about what the VFE is.", "speaker": "SPEAKER_05"}, {"start": 1280.451, "end": 1287.445, "text": "So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 1288.387, "end": 1314.799, "text": " we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical", "speaker": "SPEAKER_05"}, {"start": 1314.965, "end": 1334.813, "text": " know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was", "speaker": "SPEAKER_05"}, {"start": 1335.535, "end": 1349.242, "text": " lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning.", "speaker": "SPEAKER_05"}, {"start": 1350.083, "end": 1357.257, "text": "It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization", "speaker": "SPEAKER_05"}, {"start": 1357.574, "end": 1369.356, "text": " in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure", "speaker": "SPEAKER_05"}, {"start": 1370.163, "end": 1387.641, "text": " And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow.", "speaker": "SPEAKER_05"}, {"start": 1388.962, "end": 1391.865, "text": "And that's kind of where we left off with part one.", "speaker": "SPEAKER_05"}, {"start": 1391.885, "end": 1396.39, "text": "So that is, in effect, what we've done, where we've gone.", "speaker": "SPEAKER_05"}, {"start": 1396.943, "end": 1400.268, "text": " That's a lot of stuff to cover, but again, we haven't looked at actions yet.", "speaker": "SPEAKER_05"}, {"start": 1400.829, "end": 1407.12, "text": "We're gonna be doing that in part two, and that will bring us full circle to active inference.", "speaker": "SPEAKER_05"}, {"start": 1407.14, "end": 1420.021, "text": "And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception.", "speaker": "SPEAKER_05"}, {"start": 1420.922, "end": 1423.086, "text": "Okay, we're gonna see that these are deeply related to each other.", "speaker": "SPEAKER_05"}, {"start": 1423.623, "end": 1426.045, "text": " But that's the ground that we've covered.", "speaker": "SPEAKER_05"}, {"start": 1426.846, "end": 1429.829, "text": "I'd be interested if people have questions related to that.", "speaker": "SPEAKER_05"}, {"start": 1429.849, "end": 1431.29, "text": "I see that there's lots of questions.", "speaker": "SPEAKER_05"}, {"start": 1432.371, "end": 1432.912, "text": "Stop sharing.", "speaker": "SPEAKER_05"}, {"start": 1434.533, "end": 1437.476, "text": "And I see Mark has a question.", "speaker": "SPEAKER_05"}, {"start": 1437.496, "end": 1438.397, "text": "Please fire away, Mark.", "speaker": "SPEAKER_05"}, {"start": 1440.039, "end": 1440.519, "text": "As usual.", "speaker": "SPEAKER_02"}, {"start": 1441.16, "end": 1443.041, "text": "Thank you so much for this review.", "speaker": "SPEAKER_02"}, {"start": 1443.061, "end": 1443.602, "text": "This is great.", "speaker": "SPEAKER_02"}, {"start": 1445.724, "end": 1451.389, "text": "I was wondering if you could pull up that slide that showed the generative process and the generative model.", "speaker": "SPEAKER_02"}, {"start": 1451.429, "end": 1453.191, "text": "That might be helpful as a reference.", "speaker": "SPEAKER_02"}, {"start": 1453.964, "end": 1455.166, "text": " Oh, yeah, okay.", "speaker": "SPEAKER_05"}, {"start": 1455.186, "end": 1456.187, "text": "So let me come back here.", "speaker": "SPEAKER_05"}, {"start": 1458.43, "end": 1462.415, "text": "That's the one in chapter two, I assume you're referring to?", "speaker": "SPEAKER_05"}, {"start": 1462.435, "end": 1462.836, "text": "I think so.", "speaker": "SPEAKER_02"}, {"start": 1463.697, "end": 1464.138, "text": "This one here?", "speaker": "SPEAKER_05"}, {"start": 1464.618, "end": 1465.479, "text": "Yes, thank you.", "speaker": "SPEAKER_02"}, {"start": 1466.1, "end": 1470.126, "text": "Yeah, I really just kind of appreciate what this textbook is trying to do.", "speaker": "SPEAKER_02"}, {"start": 1470.146, "end": 1473.57, "text": "And I think it took me a while to appreciate that.", "speaker": "SPEAKER_02"}, {"start": 1473.69, "end": 1482.402, "text": "But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening.", "speaker": "SPEAKER_02"}, {"start": 1482.382, "end": 1505.689, "text": " in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise", "speaker": "SPEAKER_02"}, {"start": 1505.905, "end": 1532.082, "text": " in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of", "speaker": "SPEAKER_05"}, {"start": 1532.973, "end": 1538.437, "text": " It seems like chapter four kind of serves as a background where the story gets started for the most part.", "speaker": "SPEAKER_02"}, {"start": 1538.458, "end": 1542.375, "text": "And this is kind of more of a ground level", "speaker": "SPEAKER_02"}, {"start": 1542.457, "end": 1544.78, "text": " view, which is nice and helpful.", "speaker": "SPEAKER_02"}, {"start": 1546.603, "end": 1554.433, "text": "What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function.", "speaker": "SPEAKER_02"}, {"start": 1554.453, "end": 1559.68, "text": "Because in my mind, I'm kind of conditioned to seeing these Gaussians, right?", "speaker": "SPEAKER_02"}, {"start": 1559.7, "end": 1566.089, "text": "You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior.", "speaker": "SPEAKER_02"}, {"start": 1566.75, "end": 1571.316, "text": "However, here we have this linear function, which I guess is", "speaker": "SPEAKER_02"}, {"start": 1572.207, "end": 1575.451, "text": " then becomes part of, say, a likelihood.", "speaker": "SPEAKER_02"}, {"start": 1575.491, "end": 1588.409, "text": "And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense.", "speaker": "SPEAKER_02"}, {"start": 1588.429, "end": 1588.83, "text": "Yeah, yeah.", "speaker": "SPEAKER_05"}, {"start": 1589.651, "end": 1590.212, "text": "No, absolutely.", "speaker": "SPEAKER_05"}, {"start": 1590.252, "end": 1600.746, "text": "I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1601.131, "end": 1628.91, "text": " um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so", "speaker": "SPEAKER_05"}, {"start": 1629.464, "end": 1638.597, "text": " This is, in some sense, still our assumption about what the relationship is between observations and the hidden states.", "speaker": "SPEAKER_05"}, {"start": 1639.066, "end": 1646.86, "text": " For our purpose, we've assumed that the real world, we're sort of creating a world here.", "speaker": "SPEAKER_05"}, {"start": 1646.94, "end": 1651.348, "text": "This is the example of the food size and light intensity world.", "speaker": "SPEAKER_05"}, {"start": 1651.468, "end": 1653.572, "text": "This is the really simple world that we've seen for all of part one.", "speaker": "SPEAKER_05"}, {"start": 1654.153, "end": 1655.475, "text": "So what are the hidden states?", "speaker": "SPEAKER_05"}, {"start": 1656.076, "end": 1658.04, "text": "They are the sizes of the food.", "speaker": "SPEAKER_05"}, {"start": 1658.46, "end": 1661.666, "text": "And we can see that there's five different sizes, one to five.", "speaker": "SPEAKER_05"}, {"start": 1661.907, "end": 1662.728, "text": "That's the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1663.552, "end": 1665.214, "text": " And the observations are light intensities.", "speaker": "SPEAKER_05"}, {"start": 1665.354, "end": 1668.318, "text": "So light comes in, hits the food, and you get some sort of light intensity.", "speaker": "SPEAKER_05"}, {"start": 1668.758, "end": 1672.903, "text": "And depending on the size of the food, you get some specific light intensity.", "speaker": "SPEAKER_05"}, {"start": 1672.924, "end": 1685.619, "text": "So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size.", "speaker": "SPEAKER_05"}, {"start": 1685.639, "end": 1688.783, "text": "So we're saying that the relationship between the observation you get in", "speaker": "SPEAKER_05"}, {"start": 1689.235, "end": 1709.573, "text": " is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters", "speaker": "SPEAKER_05"}, {"start": 1710.548, "end": 1714.279, "text": " So again, this is the environment regenerative process.", "speaker": "SPEAKER_05"}, {"start": 1714.941, "end": 1721.942, "text": "The corresponding model is this one, 2.11, equation 2.11, I believe.", "speaker": "SPEAKER_05"}, {"start": 1722.513, "end": 1751.187, "text": " now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs", "speaker": "SPEAKER_05"}, {"start": 1751.555, "end": 1762.486, "text": " are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution.", "speaker": "SPEAKER_05"}, {"start": 1762.526, "end": 1780.684, "text": "So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations.", "speaker": "SPEAKER_05"}, {"start": 1781.204, "end": 1783.246, "text": " And the thing is, we're uncertain.", "speaker": "SPEAKER_05"}, {"start": 1784.268, "end": 1789.194, "text": "So the agent doesn't get to see what the real relationship is between food size and light intensity.", "speaker": "SPEAKER_05"}, {"start": 1789.334, "end": 1789.995, "text": "It has no idea.", "speaker": "SPEAKER_05"}, {"start": 1791.136, "end": 1793.679, "text": "It has a model about what it thinks the relationship is.", "speaker": "SPEAKER_05"}, {"start": 1794.24, "end": 1796.443, "text": "Now, we know that the relationship is linear.", "speaker": "SPEAKER_05"}, {"start": 1796.483, "end": 1805.994, "text": "We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity.", "speaker": "SPEAKER_05"}, {"start": 1806.014, "end": 1809.118, "text": "But in the real world, we have no idea what the relationship is.", "speaker": "SPEAKER_05"}, {"start": 1809.385, "end": 1813.791, "text": " It turns out that in this example, we've specified exactly the same relationship.", "speaker": "SPEAKER_05"}, {"start": 1814.752, "end": 1816.975, "text": "So that's quite a nice scenario to be in.", "speaker": "SPEAKER_05"}, {"start": 1817.676, "end": 1826.968, "text": "You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size.", "speaker": "SPEAKER_05"}, {"start": 1828.429, "end": 1834.397, "text": "Now that's wrong, but it's OK, maybe, for some settings.", "speaker": "SPEAKER_05"}, {"start": 1834.985, "end": 1843.699, "text": " That would be an example of a likelihood mapping that would be incorrect with respect to its generating function.", "speaker": "SPEAKER_05"}, {"start": 1844.721, "end": 1861.448, "text": "So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount.", "speaker": "SPEAKER_05"}, {"start": 1862.103, "end": 1869.693, "text": " And it's centered about the relationship we think is the case between the hidden states and the observations.", "speaker": "SPEAKER_05"}, {"start": 1870.234, "end": 1873.658, "text": "So the reason why we have this probability distribution is because we don't know.", "speaker": "SPEAKER_05"}, {"start": 1873.678, "end": 1877.924, "text": "We do not actually ever get to know what the real relationship is between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1877.944, "end": 1881.208, "text": "So does that answer your question, Mark, about why", "speaker": "SPEAKER_05"}, {"start": 1881.458, "end": 1887.967, "text": " why probability distributions, maybe why normal distributions, or did that help at all?", "speaker": "SPEAKER_05"}, {"start": 1888.007, "end": 1889.469, "text": "Yeah, that helped a lot.", "speaker": "SPEAKER_02"}, {"start": 1890.511, "end": 1890.951, "text": "Thank you.", "speaker": "SPEAKER_02"}, {"start": 1891.011, "end": 1900.164, "text": "Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right?", "speaker": "SPEAKER_02"}, {"start": 1900.464, "end": 1908.175, "text": "And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right?", "speaker": "SPEAKER_02"}, {"start": 1909.235, "end": 1934.061, "text": " uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all", "speaker": "SPEAKER_05"}, {"start": 1934.648, "end": 1938.613, "text": " We could plug, we could say, okay, the relationship is just equal to the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1938.813, "end": 1942.738, "text": "I think that the observation is exactly the hidden state, right?", "speaker": "SPEAKER_05"}, {"start": 1942.758, "end": 1943.438, "text": "That's an assumption.", "speaker": "SPEAKER_05"}, {"start": 1944.86, "end": 1952.529, "text": "Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state.", "speaker": "SPEAKER_05"}, {"start": 1953.49, "end": 1955.653, "text": "Or we can have a quadratic relationship.", "speaker": "SPEAKER_05"}, {"start": 1955.673, "end": 1957.635, "text": "We can have anything we want in there at all, actually.", "speaker": "SPEAKER_05"}, {"start": 1957.856, "end": 1963.843, "text": "And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations.", "speaker": "SPEAKER_05"}, {"start": 1964.083, "end": 1964.183, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 1966.036, "end": 1983.758, "text": " very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B,", "speaker": "SPEAKER_04"}, {"start": 1986.978, "end": 2006.903, "text": " Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from.", "speaker": "SPEAKER_05"}, {"start": 2006.923, "end": 2009.526, "text": "Michael Morehead , Other questions other questions Magdalena.", "speaker": "SPEAKER_05"}, {"start": 2011.582, "end": 2015.987, "text": " Okay, my question is this.", "speaker": "SPEAKER_01"}, {"start": 2016.608, "end": 2019.672, "text": "So could we do a thought experiment for a minute?", "speaker": "SPEAKER_01"}, {"start": 2019.712, "end": 2025.319, "text": "Because I think if I set it up as a thought experiment, I'll get the answer.", "speaker": "SPEAKER_01"}, {"start": 2026.821, "end": 2027.121, "text": "Okay.", "speaker": "SPEAKER_01"}, {"start": 2027.221, "end": 2027.942, "text": "Oh, so I'm inside.", "speaker": "SPEAKER_01"}, {"start": 2027.962, "end": 2030.966, "text": "Okay, so let's imagine the following.", "speaker": "SPEAKER_01"}, {"start": 2031.907, "end": 2040.357, "text": "In present-day societies, we have this societal", "speaker": "SPEAKER_01"}, {"start": 2040.775, "end": 2050.769, "text": " back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists.", "speaker": "SPEAKER_01"}, {"start": 2052.391, "end": 2060.623, "text": "And there's a tension in societies to update that belief by some group to science.", "speaker": "SPEAKER_01"}, {"start": 2060.89, "end": 2063.737, "text": " Exists a science so God exists.", "speaker": "SPEAKER_01"}, {"start": 2063.837, "end": 2071.114, "text": "So science tells us that God may not exist or something to that effect or quantum physics explains reality.", "speaker": "SPEAKER_01"}, {"start": 2071.214, "end": 2071.735, "text": "Not God.", "speaker": "SPEAKER_01"}, {"start": 2072.016, "end": 2072.597, "text": "Okay.", "speaker": "SPEAKER_01"}, {"start": 2072.617, "end": 2074.682, "text": "So there's a process, right?", "speaker": "SPEAKER_01"}, {"start": 2074.782, "end": 2077.228, "text": "That's in in the environment.", "speaker": "SPEAKER_01"}, {"start": 2077.208, "end": 2096.28, "text": " that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed", "speaker": "SPEAKER_01"}, {"start": 2096.733, "end": 2118.994, "text": " And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications.", "speaker": "SPEAKER_01"}, {"start": 2118.974, "end": 2148.88, "text": " for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is", "speaker": "SPEAKER_01"}, {"start": 2148.86, "end": 2163.055, "text": " Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis?", "speaker": "SPEAKER_01"}, {"start": 2165.296, "end": 2191.017, "text": " are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds", "speaker": "SPEAKER_01"}, {"start": 2191.25, "end": 2195.815, "text": " Okay, as a third level of information processing.", "speaker": "SPEAKER_01"}, {"start": 2196.296, "end": 2217.46, "text": "And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis?", "speaker": "SPEAKER_01"}, {"start": 2217.44, "end": 2219.723, "text": " Yeah, no, excellent, excellent question, Magdalena.", "speaker": "SPEAKER_05"}, {"start": 2219.964, "end": 2242.857, "text": "This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents.", "speaker": "SPEAKER_05"}, {"start": 2243.748, "end": 2250.721, "text": " Under what conditions do they interact so as to form a larger composite active inference agent?", "speaker": "SPEAKER_05"}, {"start": 2251.462, "end": 2257.513, "text": "So it's kind of a question like, how many active inference agents are there really in any given picture?", "speaker": "SPEAKER_05"}, {"start": 2260.018, "end": 2266.87, "text": "So your question, I think, insofar as I understand it, is, look, we have some notion about agency.", "speaker": "SPEAKER_05"}, {"start": 2267.423, "end": 2268.825, "text": " inactive inference.", "speaker": "SPEAKER_05"}, {"start": 2268.845, "end": 2274.433, "text": "We have some notion about something interacting with its environment, whatever the environment is, right?", "speaker": "SPEAKER_05"}, {"start": 2274.453, "end": 2278.098, "text": "So on the right hand side, we've got an agent, and on the left hand side, we have an environment.", "speaker": "SPEAKER_05"}, {"start": 2279.36, "end": 2293.699, "text": "But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent.", "speaker": "SPEAKER_05"}, {"start": 2293.84, "end": 2295.642, "text": "Okay, we haven't been able to tell that story yet.", "speaker": "SPEAKER_05"}, {"start": 2296.837, "end": 2311.04, "text": " Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers.", "speaker": "SPEAKER_05"}, {"start": 2311.12, "end": 2314.385, "text": "Let me see if I can find that figure.", "speaker": "SPEAKER_05"}, {"start": 2315.206, "end": 2317.83, "text": "Various layers within the predictive coding hierarchy.", "speaker": "SPEAKER_05"}, {"start": 2318.672, "end": 2319.433, "text": "Which one is it here?", "speaker": "SPEAKER_05"}, {"start": 2319.473, "end": 2320.074, "text": "This one?", "speaker": "SPEAKER_05"}, {"start": 2320.354, "end": 2320.735, "text": "No, this one.", "speaker": "SPEAKER_05"}, {"start": 2322.487, "end": 2329.997, "text": " You might interpret the various layers here as individual active implementations that are just doing local free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 2330.017, "end": 2333.682, "text": "So maybe this guy is an active implementation in some sense.", "speaker": "SPEAKER_05"}, {"start": 2333.702, "end": 2335.825, "text": "And then he's passing messages, blah, blah, blah.", "speaker": "SPEAKER_05"}, {"start": 2336.445, "end": 2338.989, "text": "And then the overall thing can be regarded as an active implementation.", "speaker": "SPEAKER_05"}, {"start": 2339.029, "end": 2342.393, "text": "That's the only inkling that we've seen thus far of this idea yet.", "speaker": "SPEAKER_05"}, {"start": 2343.234, "end": 2347.6, "text": "We're also not really going to see a lot of it in the rest of the book.", "speaker": "SPEAKER_05"}, {"start": 2348.052, "end": 2349.956, "text": " because it is a very open question.", "speaker": "SPEAKER_05"}, {"start": 2350.017, "end": 2351.38, "text": "It's a very difficult question.", "speaker": "SPEAKER_05"}, {"start": 2352.302, "end": 2360.36, "text": "And it gets at the heart of what agency even is at all, which is a question over and above active inference per se.", "speaker": "SPEAKER_05"}, {"start": 2360.781, "end": 2367.276, "text": "However, my personal interpretation or my personal thoughts about this issue", "speaker": "SPEAKER_05"}, {"start": 2367.678, "end": 2374.214, "text": " is that the thing that is currently missing from active inference is exactly this idea of compositionality.", "speaker": "SPEAKER_05"}, {"start": 2374.234, "end": 2375.938, "text": "So I've got agent here, agent here.", "speaker": "SPEAKER_05"}, {"start": 2376.219, "end": 2376.841, "text": "They interact.", "speaker": "SPEAKER_05"}, {"start": 2377.482, "end": 2381.873, "text": "Under what conditions is that whole thing one active inference agent?", "speaker": "SPEAKER_05"}, {"start": 2382.275, "end": 2388.688, "text": " That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research.", "speaker": "SPEAKER_05"}, {"start": 2388.709, "end": 2390.352, "text": "So I hope that answers your question in some respect.", "speaker": "SPEAKER_05"}, {"start": 2390.532, "end": 2391.494, "text": "Very, very important question.", "speaker": "SPEAKER_05"}, {"start": 2392.015, "end": 2397.828, "text": "But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding.", "speaker": "SPEAKER_05"}, {"start": 2398.81, "end": 2399.912, "text": "OK.", "speaker": "SPEAKER_01"}, {"start": 2400.23, "end": 2425.562, "text": " really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups", "speaker": "SPEAKER_01"}, {"start": 2425.542, "end": 2430.369, "text": " What I see and I'm going to use this language, I'm not a mathematician.", "speaker": "SPEAKER_01"}, {"start": 2430.429, "end": 2435.577, "text": "OK, so forgive me, but but I kind of I kind of get some things in math.", "speaker": "SPEAKER_01"}, {"start": 2435.597, "end": 2442.167, "text": "OK, so topologically generative models really work as huge attractors.", "speaker": "SPEAKER_01"}, {"start": 2442.147, "end": 2444.451, "text": " in informational systems in humans.", "speaker": "SPEAKER_01"}, {"start": 2444.912, "end": 2458.638, "text": "So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right?", "speaker": "SPEAKER_01"}, {"start": 2458.718, "end": 2462.265, "text": "So your language is a technology in human systems.", "speaker": "SPEAKER_01"}, {"start": 2462.245, "end": 2490.784, "text": " so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model", "speaker": "SPEAKER_01"}, {"start": 2491.692, "end": 2504.479, "text": " And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math.", "speaker": "SPEAKER_01"}, {"start": 2505.702, "end": 2509.55, "text": "If I just kind of sorry, is that okay?", "speaker": "SPEAKER_09"}, {"start": 2509.63, "end": 2509.951, "text": "All right.", "speaker": "SPEAKER_09"}, {"start": 2510.091, "end": 2510.632, "text": "Thanks.", "speaker": "SPEAKER_09"}, {"start": 2510.652, "end": 2512.536, "text": "Just ask someone who has a.", "speaker": "SPEAKER_09"}, {"start": 2512.516, "end": 2515.059, "text": " More immediate background in this social sciences.", "speaker": "SPEAKER_09"}, {"start": 2515.72, "end": 2516.001, "text": "Yeah.", "speaker": "SPEAKER_09"}, {"start": 2516.141, "end": 2516.822, "text": "So, so.", "speaker": "SPEAKER_09"}, {"start": 2517.323, "end": 2521.729, "text": "And I like how the question that Mark had prior to this had to do with, like.", "speaker": "SPEAKER_09"}, {"start": 2522.33, "end": 2524.332, "text": "You know, what is it to make linear assumptions?", "speaker": "SPEAKER_09"}, {"start": 2524.853, "end": 2527.777, "text": "So, in the context of the textbook, what we've seen thus far, like.", "speaker": "SPEAKER_09"}, {"start": 2528.278, "end": 2535.908, "text": "This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise.", "speaker": "SPEAKER_09"}, {"start": 2535.948, "end": 2539.774, "text": "And it's just to show how a model can be composed.", "speaker": "SPEAKER_09"}, {"start": 2540.755, "end": 2542.337, "text": "Like, how can.", "speaker": "SPEAKER_09"}, {"start": 2542.57, "end": 2551.986, "text": " In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right?", "speaker": "SPEAKER_09"}, {"start": 2552.407, "end": 2560.16, "text": "It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right?", "speaker": "SPEAKER_09"}, {"start": 2560.18, "end": 2561.422, "text": "So that's the action part.", "speaker": "SPEAKER_09"}, {"start": 2561.783, "end": 2563.245, "text": "We haven't reached that yet.", "speaker": "SPEAKER_09"}, {"start": 2563.225, "end": 2572.756, "text": " But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function?", "speaker": "SPEAKER_09"}, {"start": 2573.317, "end": 2590.597, "text": "So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right?", "speaker": "SPEAKER_09"}, {"start": 2591.198, "end": 2593.08, "text": "So that linear model we were looking at", "speaker": "SPEAKER_09"}, {"start": 2593.06, "end": 2616.688, "text": " we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people", "speaker": "SPEAKER_09"}, {"start": 2616.938, "end": 2619.842, "text": " what are the kinds of sensory information that I receive?", "speaker": "SPEAKER_09"}, {"start": 2620.522, "end": 2622.965, "text": "And what are the kinds of actions that I can do?", "speaker": "SPEAKER_09"}, {"start": 2623.005, "end": 2625.308, "text": "And what do I have beliefs about?", "speaker": "SPEAKER_09"}, {"start": 2625.408, "end": 2629.173, "text": "Or what variables do I think explains everything, right?", "speaker": "SPEAKER_09"}, {"start": 2629.774, "end": 2640.887, "text": "So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human", "speaker": "SPEAKER_09"}, {"start": 2640.867, "end": 2655.518, "text": " uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like", "speaker": "SPEAKER_09"}, {"start": 2656.021, "end": 2658.344, "text": " There's, there's a lot to track.", "speaker": "SPEAKER_09"}, {"start": 2658.364, "end": 2662.21, "text": "So there are some people who have tried to model this sort of thing.", "speaker": "SPEAKER_09"}, {"start": 2662.23, "end": 2663.572, "text": "And so I've shared this paper.", "speaker": "SPEAKER_09"}, {"start": 2664.313, "end": 2672.665, "text": "It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference.", "speaker": "SPEAKER_09"}, {"start": 2673.226, "end": 2676.19, "text": "And this kind of has to do with.", "speaker": "SPEAKER_09"}, {"start": 2676.507, "end": 2693.875, "text": " Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere.", "speaker": "SPEAKER_09"}, {"start": 2693.855, "end": 2711.429, "text": " And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing.", "speaker": "SPEAKER_09"}, {"start": 2711.489, "end": 2714.715, "text": "And that's one way of trying to look at a problem.", "speaker": "SPEAKER_09"}, {"start": 2714.695, "end": 2734.948, "text": " like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one", "speaker": "SPEAKER_09"}, {"start": 2735.333, "end": 2737.86, "text": " And a lot of agents who rally around idea two.", "speaker": "SPEAKER_09"}, {"start": 2737.9, "end": 2743.315, "text": "And as they do that, the two groups start to communicate directly with each other less and less.", "speaker": "SPEAKER_09"}, {"start": 2743.997, "end": 2749.613, "text": "The agent's actual actions that are simulated has to do with them communicating with one another and sharing their", "speaker": "SPEAKER_09"}, {"start": 2750.015, "end": 2753.178, "text": " They're kind of more verbal beliefs with one another.", "speaker": "SPEAKER_09"}, {"start": 2753.199, "end": 2764.771, "text": "And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing.", "speaker": "SPEAKER_09"}, {"start": 2764.791, "end": 2769.256, "text": "And somehow that gets all blended in with our assumptions of how science works and things like that.", "speaker": "SPEAKER_09"}, {"start": 2769.757, "end": 2771.338, "text": "It's much more complex.", "speaker": "SPEAKER_09"}, {"start": 2771.479, "end": 2779.988, "text": "And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy", "speaker": "SPEAKER_09"}, {"start": 2779.968, "end": 2783.633, "text": " But remember that free energy, based on how we've discussed it so far,", "speaker": "SPEAKER_09"}, {"start": 2783.967, "end": 2787.932, "text": " It's not some magical quantity of energy or something like that.", "speaker": "SPEAKER_09"}, {"start": 2788.914, "end": 2790.736, "text": "It's really just a proxy.", "speaker": "SPEAKER_09"}, {"start": 2790.796, "end": 2797.485, "text": "It's a word that is a proxy for something called surprisal, which is more or less uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2797.625, "end": 2799.608, "text": "So we want to minimize our uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2800.169, "end": 2813.687, "text": "So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another,", "speaker": "SPEAKER_09"}, {"start": 2813.667, "end": 2817.235, "text": " about is it idea one or is it idea two?", "speaker": "SPEAKER_09"}, {"start": 2817.716, "end": 2819.781, "text": "Is it science or is it something else?", "speaker": "SPEAKER_09"}, {"start": 2820.443, "end": 2824.993, "text": "One way to minimize uncertainty is to repeatedly tell each other", "speaker": "SPEAKER_09"}, {"start": 2825.328, "end": 2828.993, "text": " what the truth is and you all converge on what you all think the truth is.", "speaker": "SPEAKER_09"}, {"start": 2829.073, "end": 2829.293, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 2829.834, "end": 2838.545, "text": "And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much.", "speaker": "SPEAKER_09"}, {"start": 2839.085, "end": 2841.929, "text": "Because the other group is adding to your uncertainty.", "speaker": "SPEAKER_09"}, {"start": 2842.129, "end": 2849.338, "text": "Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe.", "speaker": "SPEAKER_09"}, {"start": 2849.418, "end": 2854.845, "text": "So, so you see that there, there is a kind of like logic where the free energy principle is involved.", "speaker": "SPEAKER_09"}, {"start": 2855.129, "end": 2864.026, "text": " where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise.", "speaker": "SPEAKER_09"}, {"start": 2864.467, "end": 2870.679, "text": "And that has nothing to do with the true validity of science, right?", "speaker": "SPEAKER_09"}, {"start": 2871.641, "end": 2875.348, "text": "Like I'm attempting to be a scientist explaining all this stuff,", "speaker": "SPEAKER_09"}, {"start": 2875.683, "end": 2878.748, "text": " But the people who, you see what I'm saying?", "speaker": "SPEAKER_09"}, {"start": 2878.788, "end": 2884.538, "text": "Like to take in sensory information is what we do.", "speaker": "SPEAKER_09"}, {"start": 2884.578, "end": 2888.544, "text": "It's not about like following proper logic and stuff.", "speaker": "SPEAKER_09"}, {"start": 2888.564, "end": 2892.19, "text": "It's people learn to do that, right?", "speaker": "SPEAKER_09"}, {"start": 2892.17, "end": 2894.737, "text": " Can I interject something real quick?", "speaker": "SPEAKER_01"}, {"start": 2894.757, "end": 2906.128, "text": "I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're", "speaker": "SPEAKER_01"}, {"start": 2906.581, "end": 2923.723, "text": " Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality.", "speaker": "SPEAKER_01"}, {"start": 2924.364, "end": 2931.593, "text": "So what happens in humans groups is that the active inferencing that's happening in a lot of individuals,", "speaker": "SPEAKER_01"}, {"start": 2931.573, "end": 2937.991, "text": " is to analyze and update whether or not they believe the person who's giving the message.", "speaker": "SPEAKER_01"}, {"start": 2938.552, "end": 2942.864, "text": "It's not about evaluating the validity of the premises.", "speaker": "SPEAKER_01"}, {"start": 2943.486, "end": 2947.236, "text": "It's not about the... That gets at a very...", "speaker": "SPEAKER_01"}, {"start": 2947.216, "end": 2957.913, "text": " That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer.", "speaker": "SPEAKER_05"}, {"start": 2958.734, "end": 2961.178, "text": "We might imagine an active inference agent doing inference about this.", "speaker": "SPEAKER_05"}, {"start": 2961.759, "end": 2967.688, "text": "But above that is a question, how reliable is this thing that I'm hearing at all?", "speaker": "SPEAKER_05"}, {"start": 2967.668, "end": 2992.039, "text": " and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant", "speaker": "SPEAKER_05"}, {"start": 2992.019, "end": 3018.583, "text": " uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation", "speaker": "SPEAKER_05"}, {"start": 3018.563, "end": 3022.371, "text": " for what it means for one entity to be an active inference agent.", "speaker": "SPEAKER_05"}, {"start": 3022.391, "end": 3031.651, "text": "Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents.", "speaker": "SPEAKER_05"}, {"start": 3031.671, "end": 3033.595, "text": "But that's not going to be the focus of the book.", "speaker": "SPEAKER_05"}, {"start": 3034.396, "end": 3038.445, "text": "A lot of that is at the cutting edge of active inference research now.", "speaker": "SPEAKER_05"}, {"start": 3039.961, "end": 3044.874, "text": " Thanks for keeping it closer to the textbook by bringing up precision.", "speaker": "SPEAKER_09"}, {"start": 3044.934, "end": 3046.117, "text": "I very much agree with that.", "speaker": "SPEAKER_09"}, {"start": 3046.157, "end": 3052.915, "text": "And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but.", "speaker": "SPEAKER_09"}, {"start": 3054.459, "end": 3055.422, "text": "This.", "speaker": "SPEAKER_09"}, {"start": 3055.655, "end": 3065.723, "text": " Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference.", "speaker": "SPEAKER_09"}, {"start": 3065.743, "end": 3070.095, "text": "So active inferences is generally setting up these principles.", "speaker": "SPEAKER_09"}, {"start": 3070.115, "end": 3070.997, "text": "It.", "speaker": "SPEAKER_09"}, {"start": 3070.977, "end": 3092.985, "text": " definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different", "speaker": "SPEAKER_09"}, {"start": 3092.965, "end": 3093.386, "text": " Right?", "speaker": "SPEAKER_09"}, {"start": 3093.806, "end": 3097.411, "text": "Like, our variables that we're including in our respective models are different.", "speaker": "SPEAKER_09"}, {"start": 3097.772, "end": 3099.694, "text": "The precisions can definitely differ.", "speaker": "SPEAKER_09"}, {"start": 3099.714, "end": 3106.303, "text": "If I get information from someone in my group, I might have a higher precision and trust what they say more.", "speaker": "SPEAKER_09"}, {"start": 3106.764, "end": 3114.654, "text": "I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right?", "speaker": "SPEAKER_09"}, {"start": 3115.155, "end": 3118.72, "text": "So a lot of the, you know, a lot of the", "speaker": "SPEAKER_09"}, {"start": 3119.054, "end": 3131.018, "text": " political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right?", "speaker": "SPEAKER_09"}, {"start": 3131.159, "end": 3134.245, "text": "It's like, oh, I just lump them all in with political group", "speaker": "SPEAKER_09"}, {"start": 3134.225, "end": 3156.351, "text": " be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available", "speaker": "SPEAKER_09"}, {"start": 3156.331, "end": 3172.112, "text": " then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone", "speaker": "SPEAKER_09"}, {"start": 3172.092, "end": 3191.299, "text": " ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it", "speaker": "SPEAKER_09"}, {"start": 3192.308, "end": 3195.736, "text": " Do make sure, I've been absent for a little bit.", "speaker": "SPEAKER_05"}, {"start": 3196.237, "end": 3200.687, "text": "I'm gonna be much more active on the questions side of things on the CODA.", "speaker": "SPEAKER_05"}, {"start": 3200.748, "end": 3205.418, "text": "So do make sure if you've got questions, please do put them down in the questions tab here.", "speaker": "SPEAKER_05"}, {"start": 3205.458, "end": 3209.608, "text": "I've gone through and I've already answered quite a few of them for chapter five.", "speaker": "SPEAKER_05"}, {"start": 3209.588, "end": 3224.188, "text": " uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see", "speaker": "SPEAKER_05"}, {"start": 3224.573, "end": 3226.135, "text": " Are there any more live questions?", "speaker": "SPEAKER_05"}, {"start": 3227.777, "end": 3228.718, "text": "That would be nice.", "speaker": "SPEAKER_05"}, {"start": 3228.738, "end": 3233.584, "text": "We can stay around for maybe five or so minutes after the deadline if people so choose.", "speaker": "SPEAKER_05"}, {"start": 3234.645, "end": 3236.267, "text": "But we are coming up to the top of the hour.", "speaker": "SPEAKER_05"}, {"start": 3236.287, "end": 3241.994, "text": "So yeah, all of the questions in chapter five now are answered, or I've given my attempt.", "speaker": "SPEAKER_05"}, {"start": 3242.655, "end": 3246.379, "text": "As I say, I'll go through and attempt to answer all the previous ones as well.", "speaker": "SPEAKER_05"}, {"start": 3249.937, "end": 3276.737, "text": " think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry", "speaker": "SPEAKER_05"}, {"start": 3277.595, "end": 3301.122, "text": " okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian", "speaker": "SPEAKER_08"}, {"start": 3301.102, "end": 3308.391, "text": " inference to engage in parameter learning so that it can update beliefs about the current state.", "speaker": "SPEAKER_08"}, {"start": 3309.172, "end": 3319.545, "text": "If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself.", "speaker": "SPEAKER_08"}, {"start": 3319.985, "end": 3323.97, "text": "And I think this is what we call plasticity.", "speaker": "SPEAKER_08"}, {"start": 3323.95, "end": 3331.683, "text": " The ultimate goal of all this is to minimize expected free energy.", "speaker": "SPEAKER_08"}, {"start": 3332.404, "end": 3333.085, "text": "Is that right?", "speaker": "SPEAKER_08"}, {"start": 3334.548, "end": 3335.088, "text": "Wow, okay.", "speaker": "SPEAKER_05"}, {"start": 3335.129, "end": 3338.174, "text": "I mean, basically, yeah, that's substantively correct.", "speaker": "SPEAKER_05"}, {"start": 3339.115, "end": 3342.06, "text": "I will note, however, we haven't yet talked about expected free energy.", "speaker": "SPEAKER_05"}, {"start": 3342.395, "end": 3344.237, "text": " and planning and action selection.", "speaker": "SPEAKER_05"}, {"start": 3344.257, "end": 3345.358, "text": "So that is going to come later.", "speaker": "SPEAKER_05"}, {"start": 3346.379, "end": 3346.959, "text": "But yeah, you're right.", "speaker": "SPEAKER_05"}, {"start": 3346.979, "end": 3352.825, "text": "In terms of the general flavor of things, we have beliefs about hidden states in the world.", "speaker": "SPEAKER_05"}, {"start": 3353.686, "end": 3365.076, "text": "And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization.", "speaker": "SPEAKER_05"}, {"start": 3365.997, "end": 3369.921, "text": "Or we've seen that we can also recast this in terms of predictive processing.", "speaker": "SPEAKER_05"}, {"start": 3370.66, "end": 3372.943, "text": " minimizing the precision way to prediction errors.", "speaker": "SPEAKER_05"}, {"start": 3372.963, "end": 3373.924, "text": "It's the same kind of story.", "speaker": "SPEAKER_05"}, {"start": 3374.926, "end": 3386.221, "text": "And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states.", "speaker": "SPEAKER_05"}, {"start": 3387.063, "end": 3387.283, "text": "Yes.", "speaker": "SPEAKER_05"}, {"start": 3388.725, "end": 3391.328, "text": "The other thing, you mentioned parameter learning.", "speaker": "SPEAKER_05"}, {"start": 3391.388, "end": 3392.77, "text": "So yes, exactly.", "speaker": "SPEAKER_05"}, {"start": 3392.79, "end": 3393.932, "text": "There's these two problems.", "speaker": "SPEAKER_05"}, {"start": 3394.032, "end": 3397.317, "text": "We have the problem of inferring the hidden states.", "speaker": "SPEAKER_05"}, {"start": 3397.357, "end": 3398.338, "text": "What should the hidden states be?", "speaker": "SPEAKER_05"}, {"start": 3398.977, "end": 3407.788, "text": " But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference.", "speaker": "SPEAKER_05"}, {"start": 3408.529, "end": 3415.017, "text": "So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like.", "speaker": "SPEAKER_05"}, {"start": 3415.738, "end": 3419.183, "text": "But then above that, we need to deal with the problem of what should the setting of the parameters be?", "speaker": "SPEAKER_05"}, {"start": 3420.204, "end": 3426.853, "text": "And we've only really just begun to look at this in terms of prediction errors and hierarchical models.", "speaker": "SPEAKER_05"}, {"start": 3428.034, "end": 3428.815, "text": "But we saw", "speaker": "SPEAKER_05"}, {"start": 3429.149, "end": 3432.913, "text": " how to deal with that in terms of expectation maximization and earlier chapters.", "speaker": "SPEAKER_05"}, {"start": 3433.634, "end": 3443.644, "text": "We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture.", "speaker": "SPEAKER_05"}, {"start": 3443.964, "end": 3450.251, "text": "Because we generally can't do expectation maximization for the reason that we don't know the exact posterior.", "speaker": "SPEAKER_05"}, {"start": 3450.291, "end": 3458.299, "text": "So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah.", "speaker": "SPEAKER_05"}, {"start": 3458.339, "end": 3458.599, "text": "Thank you.", "speaker": "SPEAKER_08"}, {"start": 3459.49, "end": 3461.011, "text": " No problem.", "speaker": "SPEAKER_05"}, {"start": 3461.032, "end": 3462.473, "text": "Do we have more questions?", "speaker": "SPEAKER_05"}, {"start": 3462.633, "end": 3463.254, "text": "More questions?", "speaker": "SPEAKER_05"}, {"start": 3464.655, "end": 3466.857, "text": "Maybe one more question, if there is one more.", "speaker": "SPEAKER_05"}, {"start": 3467.117, "end": 3469.52, "text": "And then I'll stop the recording.", "speaker": "SPEAKER_05"}, {"start": 3475.566, "end": 3475.866, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 3475.886, "end": 3476.887, "text": "I might stop the recording here.", "speaker": "SPEAKER_05"}, {"start": 3478.068, "end": 3478.549, "text": "Well, Mark.", "speaker": "SPEAKER_05"}, {"start": 3478.569, "end": 3479.71, "text": "Hello, everyone, and welcome.", "speaker": "SPEAKER_05"}, {"start": 3479.91, "end": 3481.732, "text": "I've got my esteemed friend.", "speaker": "SPEAKER_05"}, {"start": 3482.753, "end": 3484.975, "text": "Oh, hang on.", "speaker": "SPEAKER_05"}, {"start": 3486.423, "end": 3489.528, "text": " That was interesting audio feedback.", "speaker": "SPEAKER_02"}, {"start": 3489.548, "end": 3492.053, "text": "I thought I would jump in since nobody else asked us.", "speaker": "SPEAKER_02"}, {"start": 3493.135, "end": 3499.105, "text": "But earlier you mentioned using gradient descent to learn parameters, right?", "speaker": "SPEAKER_02"}, {"start": 3499.245, "end": 3501.289, "text": "Is that also used in perception?", "speaker": "SPEAKER_02"}, {"start": 3501.69, "end": 3503.893, "text": "Is it the same approach?", "speaker": "SPEAKER_02"}, {"start": 3503.994, "end": 3508.581, "text": "Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things.", "speaker": "SPEAKER_05"}, {"start": 3508.602, "end": 3510.625, "text": "So it's a very, very general", "speaker": "SPEAKER_05"}, {"start": 3511.06, "end": 3512.522, "text": " optimization techniques.", "speaker": "SPEAKER_05"}, {"start": 3512.602, "end": 3520.19, "text": "So yes, it certainly can be used for learning, well, for perception.", "speaker": "SPEAKER_05"}, {"start": 3520.711, "end": 3522.513, "text": "I'm trying to find a nice picture here.", "speaker": "SPEAKER_05"}, {"start": 3523.173, "end": 3531.983, "text": "Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent.", "speaker": "SPEAKER_05"}, {"start": 3532.524, "end": 3533.445, "text": "Let me just make sure of that.", "speaker": "SPEAKER_05"}, {"start": 3533.565, "end": 3535.107, "text": "Yes, in other words, is the answer.", "speaker": "SPEAKER_05"}, {"start": 3535.547, "end": 3537.049, "text": "It's used for a lot of things.", "speaker": "SPEAKER_05"}, {"start": 3538.39, "end": 3539.852, "text": "Very, very general technique.", "speaker": "SPEAKER_05"}, {"start": 3540.237, "end": 3549.91, "text": " Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six.", "speaker": "SPEAKER_05"}, {"start": 3549.93, "end": 3550.39, "text": "Absolutely.", "speaker": "SPEAKER_05"}, {"start": 3550.51, "end": 3550.891, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3550.911, "end": 3552.593, "text": "We're gonna talk about phase planes and so on.", "speaker": "SPEAKER_05"}, {"start": 3553.394, "end": 3553.635, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3553.955, "end": 3556.258, "text": "There are problems with it, but yes, we're gonna see it going forward.", "speaker": "SPEAKER_05"}, {"start": 3556.278, "end": 3559.242, "text": "It's a very, it's a very foundational technique.", "speaker": "SPEAKER_05"}, {"start": 3559.382, "end": 3562.807, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3562.827, "end": 3563.027, "text": "All right.", "speaker": "SPEAKER_05"}, {"start": 3563.047, "end": 3566.211, "text": "Any, any last minute questions for the YouTube recording before we, uh,", "speaker": "SPEAKER_05"}, {"start": 3569.093, "end": 3570.975, "text": " If not, I think I'm sorry.", "speaker": "SPEAKER_05"}, {"start": 3571.236, "end": 3572.377, "text": "Can I ask a quick question?", "speaker": "SPEAKER_07"}, {"start": 3572.437, "end": 3574.2, "text": "Yeah, sure.", "speaker": "SPEAKER_05"}, {"start": 3574.74, "end": 3575.461, "text": "I do apologize.", "speaker": "SPEAKER_07"}, {"start": 3575.702, "end": 3577.424, "text": "I have joined the group very late.", "speaker": "SPEAKER_07"}, {"start": 3577.484, "end": 3579.767, "text": "So maybe this was addressed in like the past weeks.", "speaker": "SPEAKER_07"}, {"start": 3579.827, "end": 3593.245, "text": "But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning?", "speaker": "SPEAKER_07"}, {"start": 3593.478, "end": 3619.553, "text": " well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms", "speaker": "SPEAKER_05"}, {"start": 3619.955, "end": 3629.547, "text": " If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control.", "speaker": "SPEAKER_05"}, {"start": 3630.147, "end": 3643.784, "text": "There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference.", "speaker": "SPEAKER_05"}, {"start": 3643.804, "end": 3646.988, "text": "So the answer is yes, there is a profound relationship.", "speaker": "SPEAKER_05"}, {"start": 3647.008, "end": 3649.371, "text": "We probably don't have time to get into it here.", "speaker": "SPEAKER_05"}, {"start": 3649.84, "end": 3659.489, "text": " One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain.", "speaker": "SPEAKER_05"}, {"start": 3660.41, "end": 3665.735, "text": "Because we've seen with, well, we haven't yet seen with expected free energy how that works.", "speaker": "SPEAKER_05"}, {"start": 3665.755, "end": 3672.461, "text": "But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning.", "speaker": "SPEAKER_05"}, {"start": 3673.062, "end": 3679.628, "text": "And we don't just have this scale of reward, this signal in reinforcement learning.", "speaker": "SPEAKER_05"}, {"start": 3680.114, "end": 3708.67, "text": " we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um", "speaker": "SPEAKER_05"}, {"start": 3708.97, "end": 3717.107, "text": " a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning.", "speaker": "SPEAKER_09"}, {"start": 3717.147, "end": 3727.688, "text": "Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning,", "speaker": "SPEAKER_09"}, {"start": 3727.668, "end": 3741.341, "text": " Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two.", "speaker": "SPEAKER_09"}, {"start": 3742.282, "end": 3752.331, "text": "But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven.", "speaker": "SPEAKER_09"}, {"start": 3753.192, "end": 3756.595, "text": "So there can be like a KL control agent or there can be a variety.", "speaker": "SPEAKER_09"}, {"start": 3756.575, "end": 3779.157, "text": " other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that", "speaker": "SPEAKER_09"}, {"start": 3779.137, "end": 3782.622, "text": " that things like curiosity exist.", "speaker": "SPEAKER_09"}, {"start": 3782.742, "end": 3787.429, "text": "Why does curiosity exist if we're actually always just driven towards a reward?", "speaker": "SPEAKER_09"}, {"start": 3787.509, "end": 3790.173, "text": "Once you know what the reward is, you should just go for that, right?", "speaker": "SPEAKER_09"}, {"start": 3790.213, "end": 3796.262, "text": "And you would have no reason to find some other strategy for going for it.", "speaker": "SPEAKER_09"}, {"start": 3796.463, "end": 3798.466, "text": "You would have no reason to go out of your way.", "speaker": "SPEAKER_09"}, {"start": 3798.606, "end": 3803.032, "text": "From what you already know, you just keep going for the reward as much as possible.", "speaker": "SPEAKER_09"}, {"start": 3803.393, "end": 3808.941, "text": "Perhaps you learn other things through happenstance along the way.", "speaker": "SPEAKER_09"}, {"start": 3809.258, "end": 3811.041, "text": " That's a big difference.", "speaker": "SPEAKER_05"}, {"start": 3811.722, "end": 3814.948, "text": "In reinforcement, you can just sort of start maximizing a reward signal.", "speaker": "SPEAKER_05"}, {"start": 3815.569, "end": 3822.601, "text": "But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards.", "speaker": "SPEAKER_05"}, {"start": 3822.621, "end": 3828.972, "text": "You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards.", "speaker": "SPEAKER_05"}, {"start": 3828.952, "end": 3856.731, "text": " yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite", "speaker": "SPEAKER_09"}, {"start": 3857.234, "end": 3868.402, "text": " how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning.", "speaker": "SPEAKER_09"}, {"start": 3868.743, "end": 3872.492, "text": "Meanwhile, active inference has a much more principled way that relates to this notion of", "speaker": "SPEAKER_09"}, {"start": 3872.472, "end": 3892.793, "text": " free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually", "speaker": "SPEAKER_09"}, {"start": 3893.06, "end": 3896.992, "text": " Um, when I play a sport, I throw the ball this way.", "speaker": "SPEAKER_09"}, {"start": 3897.012, "end": 3906.681, "text": "Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person.", "speaker": "SPEAKER_09"}, {"start": 3906.897, "end": 3909.963, "text": " but maybe I will kind of curve it or change it, right?", "speaker": "SPEAKER_09"}, {"start": 3910.343, "end": 3914.03, "text": "You're not randomly throwing it in the sky or in the opposite direction.", "speaker": "SPEAKER_09"}, {"start": 3914.431, "end": 3919.481, "text": "You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right?", "speaker": "SPEAKER_09"}, {"start": 3919.541, "end": 3927.235, "text": "There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better.", "speaker": "SPEAKER_09"}, {"start": 3927.552, "end": 3953.037, "text": " based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about", "speaker": "SPEAKER_09"}, {"start": 3953.523, "end": 3956.968, "text": " How do I perform the best whenever it comes to active inference?", "speaker": "SPEAKER_09"}, {"start": 3957.129, "end": 3965.422, "text": "It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them.", "speaker": "SPEAKER_09"}, {"start": 3966.223, "end": 3975.918, "text": "But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear.", "speaker": "SPEAKER_09"}, {"start": 3976.038, "end": 3977.32, "text": "Yeah.", "speaker": "SPEAKER_05"}, {"start": 3978.194, "end": 3983.299, "text": " I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys.", "speaker": "SPEAKER_05"}, {"start": 3983.319, "end": 3986.502, "text": "So Giancuomo, fire away.", "speaker": "SPEAKER_05"}, {"start": 3986.522, "end": 3991.827, "text": "Very quickly, I think it ties in with what was just spoken and talked about now.", "speaker": "SPEAKER_00"}, {"start": 3992.648, "end": 4007.982, "text": "Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is?", "speaker": "SPEAKER_00"}, {"start": 4008.265, "end": 4033.468, "text": " Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty.", "speaker": "SPEAKER_00"}, {"start": 4033.508, "end": 4035.029, "text": "They say, this is the answer.", "speaker": "SPEAKER_00"}, {"start": 4035.109, "end": 4037.091, "text": "And you're like, no, no, no, it's not.", "speaker": "SPEAKER_00"}, {"start": 4037.594, "end": 4057.919, "text": " And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware.", "speaker": "SPEAKER_00"}, {"start": 4057.959, "end": 4064.007, "text": "Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind?", "speaker": "SPEAKER_00"}, {"start": 4065.455, "end": 4093.251, "text": " well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions", "speaker": "SPEAKER_05"}, {"start": 4093.687, "end": 4094.989, "text": " This is built in from the start.", "speaker": "SPEAKER_05"}, {"start": 4095.029, "end": 4103.639, "text": "So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right?", "speaker": "SPEAKER_05"}, {"start": 4104.22, "end": 4105.982, "text": "And okay, we can't do Bayes' rule exactly.", "speaker": "SPEAKER_05"}, {"start": 4106.022, "end": 4107.023, "text": "We have to do it approximately.", "speaker": "SPEAKER_05"}, {"start": 4107.083, "end": 4108.365, "text": "We do variational inference.", "speaker": "SPEAKER_05"}, {"start": 4108.405, "end": 4110.407, "text": "We do prediction error and minimization, that kind of thing.", "speaker": "SPEAKER_05"}, {"start": 4110.828, "end": 4114.052, "text": "But we're always, always, always reasoning about probability distribution.", "speaker": "SPEAKER_05"}, {"start": 4114.072, "end": 4118.597, "text": "So yes, every single active implementation ever is always", "speaker": "SPEAKER_05"}, {"start": 4120.535, "end": 4147.757, "text": " predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet", "speaker": "SPEAKER_05"}, {"start": 4147.737, "end": 4149.84, "text": " And there's issues around that and how to do that.", "speaker": "SPEAKER_05"}, {"start": 4150.821, "end": 4151.722, "text": "But we haven't seen that just yet.", "speaker": "SPEAKER_05"}, {"start": 4151.742, "end": 4153.024, "text": "So, yes, absolutely.", "speaker": "SPEAKER_05"}, {"start": 4153.084, "end": 4155.808, "text": "It's part and parcel of what it is to be an active infatuation.", "speaker": "SPEAKER_05"}, {"start": 4155.828, "end": 4157.15, "text": "It's the reason about uncertainty now.", "speaker": "SPEAKER_05"}, {"start": 4158.411, "end": 4158.932, "text": "Okay, thank you.", "speaker": "SPEAKER_05"}, {"start": 4159.613, "end": 4159.913, "text": "No problem.", "speaker": "SPEAKER_05"}, {"start": 4159.933, "end": 4162.036, "text": "And then, Mark, and then I think we can do... Oh, hang on.", "speaker": "SPEAKER_05"}, {"start": 4162.337, "end": 4167.163, "text": "Maybe just really quickly, did you have a comment, Andrew?", "speaker": "SPEAKER_05"}, {"start": 4167.413, "end": 4171.819, "text": " I also really have to go, but yeah, just briefly, you basically answered it.", "speaker": "SPEAKER_09"}, {"start": 4171.839, "end": 4173.622, "text": "Sorry, I was looking at the chat.", "speaker": "SPEAKER_09"}, {"start": 4173.642, "end": 4177.928, "text": "It's just, yeah, it depends on how the question is being asked.", "speaker": "SPEAKER_09"}, {"start": 4177.948, "end": 4185.339, "text": "Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are?", "speaker": "SPEAKER_09"}, {"start": 4185.459, "end": 4191.348, "text": "Like that's going to take a little bit more than just like only looking at the simplified models.", "speaker": "SPEAKER_09"}, {"start": 4191.368, "end": 4193.05, "text": "We're looking at the textbook here.", "speaker": "SPEAKER_09"}, {"start": 4193.03, "end": 4216.928, "text": " um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization", "speaker": "SPEAKER_09"}, {"start": 4216.908, "end": 4230.176, "text": " of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots.", "speaker": "SPEAKER_09"}, {"start": 4230.256, "end": 4232.18, "text": "There are 2 possible things.", "speaker": "SPEAKER_09"}, {"start": 4232.22, "end": 4235.507, "text": "It could be is completely 50 50 is fully uncertain.", "speaker": "SPEAKER_09"}, {"start": 4235.487, "end": 4246.867, "text": " Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision.", "speaker": "SPEAKER_09"}, {"start": 4247.388, "end": 4247.609, "text": "Right?", "speaker": "SPEAKER_09"}, {"start": 4247.709, "end": 4255.423, "text": "Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob.", "speaker": "SPEAKER_09"}, {"start": 4255.803, "end": 4256.044, "text": "Right?", "speaker": "SPEAKER_09"}, {"start": 4256.104, "end": 4258.508, "text": "The more you turn up precision, the more you're.", "speaker": "SPEAKER_09"}, {"start": 4258.488, "end": 4263.134, "text": " you're going to take into account the prediction errors you're receiving and vice versa.", "speaker": "SPEAKER_09"}, {"start": 4263.214, "end": 4275.609, "text": "So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving.", "speaker": "SPEAKER_09"}, {"start": 4275.669, "end": 4275.89, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 4275.91, "end": 4279.294, "text": "So so uncertainty is very much like throughout.", "speaker": "SPEAKER_09"}, {"start": 4279.474, "end": 4284.18, "text": "I mean, it's a very ubiquitous term through many parts of active inference.", "speaker": "SPEAKER_09"}, {"start": 4284.2, "end": 4284.34, "text": "Yeah.", "speaker": "SPEAKER_09"}, {"start": 4284.48, "end": 4285.742, "text": "So it's essential.", "speaker": "SPEAKER_09"}, {"start": 4287.224, "end": 4288.225, "text": "So all right.", "speaker": "SPEAKER_03"}, {"start": 4288.677, "end": 4311.58, "text": " so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see", "speaker": "SPEAKER_03"}, {"start": 4312.673, "end": 4337.179, "text": " right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh", "speaker": "SPEAKER_09"}, {"start": 4337.547, "end": 4338.789, "text": " who cares about priors.", "speaker": "SPEAKER_09"}, {"start": 4338.829, "end": 4342.634, "text": "And that's what we see with maximum likelihood estimation in chapter two there.", "speaker": "SPEAKER_09"}, {"start": 4343.655, "end": 4354.31, "text": "But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something.", "speaker": "SPEAKER_09"}, {"start": 4354.35, "end": 4366.086, "text": "It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like,", "speaker": "SPEAKER_09"}, {"start": 4366.403, "end": 4392.141, "text": " thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um", "speaker": "SPEAKER_09"}, {"start": 4392.661, "end": 4394.263, "text": " you know, how those connect up with each other.", "speaker": "SPEAKER_09"}, {"start": 4394.704, "end": 4403.396, "text": "So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there.", "speaker": "SPEAKER_09"}, {"start": 4403.476, "end": 4413.409, "text": "And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe.", "speaker": "SPEAKER_09"}, {"start": 4413.55, "end": 4416.193, "text": "And so I fully trust it without question.", "speaker": "SPEAKER_09"}, {"start": 4416.654, "end": 4416.854, "text": "Right.", "speaker": "SPEAKER_09"}, {"start": 4416.874, "end": 4419.057, "text": "So I say all these things because, uh,", "speaker": "SPEAKER_09"}, {"start": 4419.037, "end": 4426.085, "text": " Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe.", "speaker": "SPEAKER_09"}, {"start": 4426.666, "end": 4442.303, "text": "And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things.", "speaker": "SPEAKER_09"}, {"start": 4442.343, "end": 4444.065, "text": "Yeah, it's very interesting to think about.", "speaker": "SPEAKER_09"}, {"start": 4445.294, "end": 4449.623, "text": " Yeah, do take a look at, I think I'm sharing, but figure 2.11.", "speaker": "SPEAKER_05"}, {"start": 4450.525, "end": 4459.965, "text": "This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief.", "speaker": "SPEAKER_05"}, {"start": 4459.985, "end": 4462.29, "text": "So if the precision is really tight around the prior,", "speaker": "SPEAKER_05"}, {"start": 4462.692, "end": 4472.808, "text": " the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model.", "speaker": "SPEAKER_05"}, {"start": 4472.828, "end": 4474.511, "text": "So yeah, do take a look at that.", "speaker": "SPEAKER_05"}, {"start": 4474.531, "end": 4482.584, "text": "That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have.", "speaker": "SPEAKER_05"}, {"start": 4482.684, "end": 4483.786, "text": "Yeah.", "speaker": "SPEAKER_03"}, {"start": 4483.806, "end": 4484.146, "text": "Excellent.", "speaker": "SPEAKER_03"}, {"start": 4484.206, "end": 4484.667, "text": "Thank you.", "speaker": "SPEAKER_03"}, {"start": 4485.372, "end": 4485.512, "text": " Cool.", "speaker": "SPEAKER_05"}, {"start": 4485.532, "end": 4487.255, "text": "All right, guys, we're going to have to call it there.", "speaker": "SPEAKER_05"}, {"start": 4487.756, "end": 4491.963, "text": "Do put questions in the chats on the page, the code page.", "speaker": "SPEAKER_05"}, {"start": 4492.003, "end": 4493.606, "text": "I'll be a bit more attentive going forward.", "speaker": "SPEAKER_05"}, {"start": 4494.267, "end": 4495.489, "text": "I look forward to next week.", "speaker": "SPEAKER_05"}, {"start": 4495.509, "end": 4499.796, "text": "I'll be doing another session of this kind on Friday for the people who usually attend that session.", "speaker": "SPEAKER_05"}, {"start": 4499.816, "end": 4500.557, "text": "So thank you very much.", "speaker": "SPEAKER_05"}, {"start": 4500.918, "end": 4502.22, "text": "Stop sharing and stop the recording.", "speaker": "SPEAKER_05"}, {"start": 4503.522, "end": 4504.704, "text": "Have we got a recording here?", "speaker": "SPEAKER_05"}, {"start": 4506.247, "end": 4507.269, "text": "Okay, stop the recording.", "speaker": "SPEAKER_05"}, {"start": 4507.749, "end": 4508.551, "text": "Goodbye, YouTube people.", "speaker": "SPEAKER_05"}, {"start": 4508.751, "end": 4509.372, "text": "Until next time.", "speaker": "SPEAKER_05"}]}] \ No newline at end of file +[{"video_id": "MjeQeWeyYhE", "segments": [{"start": 3.727, "end": 29.498, "text": " all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense", "speaker": "Fraser Paterson"}, {"start": 29.798, "end": 40.317, "text": " And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two.", "speaker": "Fraser Paterson"}, {"start": 40.337, "end": 43.923, "text": "We're going to do active inputs proper, which is going to be very, very exciting going forward.", "speaker": "Fraser Paterson"}, {"start": 43.943, "end": 50.575, "text": "So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts.", "speaker": "Fraser Paterson"}, {"start": 51.027, "end": 56.833, "text": " that we're going to start doing actual, quote unquote, active inference in the second part.", "speaker": "Fraser Paterson"}, {"start": 56.853, "end": 67.444, "text": "But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection.", "speaker": "Fraser Paterson"}, {"start": 68.305, "end": 72.409, "text": "We're going to sort of slow down and think about part one in general, chapters one to five.", "speaker": "Fraser Paterson"}, {"start": 73.531, "end": 80.578, "text": "This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one.", "speaker": "Fraser Paterson"}, {"start": 80.812, "end": 85.079, "text": " you know, going so that we can be adequately grounded going forward.", "speaker": "Fraser Paterson"}, {"start": 85.981, "end": 94.696, "text": "I will say we have one one question is, you know, we've done each chapter across two weeks.", "speaker": "Fraser Paterson"}, {"start": 95.658, "end": 100.947, "text": "Historically, you know, chapter one, we've had two weeks going forward.", "speaker": "Fraser Paterson"}, {"start": 101.72, "end": 128.115, "text": " depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week", "speaker": "Fraser Paterson"}, {"start": 128.095, "end": 152.685, "text": " of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe", "speaker": "Fraser Paterson"}, {"start": 152.935, "end": 159.784, "text": " It would be a good idea to reach out through the blankets email.", "speaker": "Fraser Paterson"}, {"start": 160.425, "end": 166.993, "text": "Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place.", "speaker": "Fraser Paterson"}, {"start": 168.415, "end": 169.296, "text": "That would be amazing.", "speaker": "Fraser Paterson"}, {"start": 169.416, "end": 170.057, "text": "Good idea.", "speaker": "Fraser Paterson"}, {"start": 170.698, "end": 171.198, "text": "Yes, please.", "speaker": "Fraser Paterson"}, {"start": 171.238, "end": 176.445, "text": "I see a lot of people are saying yes.", "speaker": "Fraser Paterson"}, {"start": 176.695, "end": 182.461, "text": " In this chat here, it would be very helpful if you could give your yay or nay to that as well.", "speaker": "Fraser Paterson"}, {"start": 182.581, "end": 185.384, "text": "So we can actually look at that just directly through the chat here.", "speaker": "Fraser Paterson"}, {"start": 185.464, "end": 188.848, "text": "So if you want a second week review, yes.", "speaker": "Fraser Paterson"}, {"start": 189.108, "end": 190.089, "text": "If you don't, no.", "speaker": "Fraser Paterson"}, {"start": 190.71, "end": 192.252, "text": "And then we can go forward.", "speaker": "Fraser Paterson"}, {"start": 193.553, "end": 193.893, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 195.495, "end": 196.096, "text": "Enough of that.", "speaker": "Fraser Paterson"}, {"start": 196.616, "end": 197.898, "text": "I'll start sharing my screen here.", "speaker": "Fraser Paterson"}, {"start": 199.259, "end": 200.28, "text": "Probably my entire screen.", "speaker": "Fraser Paterson"}, {"start": 204.084, "end": 204.505, "text": "Okie dokie.", "speaker": "Fraser Paterson"}, {"start": 204.525, "end": 205.446, "text": "So people should be able to see.", "speaker": "Fraser Paterson"}, {"start": 205.486, "end": 206.567, "text": "Maybe just get rid of...", "speaker": "Fraser Paterson"}, {"start": 207.357, "end": 214.009, "text": " The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise.", "speaker": "Fraser Paterson"}, {"start": 215.011, "end": 227.172, "text": "about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here.", "speaker": "Fraser Paterson"}, {"start": 227.928, "end": 229.152, "text": " So you've got your overview.", "speaker": "Fraser Paterson"}, {"start": 229.935, "end": 232.945, "text": "I've added one or two additional resources.", "speaker": "Fraser Paterson"}, {"start": 232.965, "end": 233.827, "text": "So these actually came up.", "speaker": "Fraser Paterson"}, {"start": 233.908, "end": 235.593, "text": "Andrew shared the first of these.", "speaker": "Fraser Paterson"}, {"start": 235.613, "end": 236.436, "text": "These are all videos.", "speaker": "Fraser Paterson"}, {"start": 237.239, "end": 238.242, "text": "One of them is an active inference", "speaker": "Fraser Paterson"}, {"start": 240.668, "end": 241.791, "text": " coding and active inference.", "speaker": "Fraser Paterson"}, {"start": 243.094, "end": 245.139, "text": "So that's Ryan Smith and his team.", "speaker": "Fraser Paterson"}, {"start": 245.841, "end": 248.227, "text": "They've done excellent work on predictive coding.", "speaker": "Fraser Paterson"}, {"start": 248.708, "end": 249.45, "text": "I'll play that just now.", "speaker": "Fraser Paterson"}, {"start": 249.991, "end": 258.171, "text": "This is a great stream just about predictive coding more generally and its relation to active inference and its relation to", "speaker": "Fraser Paterson"}, {"start": 258.151, "end": 260.855, "text": " you know, its status in actual brains.", "speaker": "Fraser Paterson"}, {"start": 261.616, "end": 265.781, "text": "So if you're interested in more context, this would be excellent to go through.", "speaker": "Fraser Paterson"}, {"start": 265.801, "end": 267.003, "text": "And there's two more links there as well.", "speaker": "Fraser Paterson"}, {"start": 267.043, "end": 271.99, "text": "One by Jakob Howie, he's a philosopher over in Australia in the University of Monash.", "speaker": "Fraser Paterson"}, {"start": 272.01, "end": 274.013, "text": "What is predictive processing and what is it good for?", "speaker": "Fraser Paterson"}, {"start": 274.133, "end": 274.914, "text": "Another excellent talk.", "speaker": "Fraser Paterson"}, {"start": 275.495, "end": 280.041, "text": "And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke.", "speaker": "Fraser Paterson"}, {"start": 280.325, "end": 281.426, "text": " from the University of Toronto.", "speaker": "Fraser Paterson"}, {"start": 281.446, "end": 286.772, "text": "He's actually speaking about Neoplatonism and mystical experiences, believe it or not.", "speaker": "Fraser Paterson"}, {"start": 287.513, "end": 289.135, "text": "And I know it sounds weird.", "speaker": "Fraser Paterson"}, {"start": 289.575, "end": 292.799, "text": "Predictive processing actually shows up in here in a very significant way.", "speaker": "Fraser Paterson"}, {"start": 292.819, "end": 297.084, "text": "So that might be a nice mood setter, as it were.", "speaker": "Fraser Paterson"}, {"start": 297.164, "end": 298.365, "text": "So there's some additional resources there.", "speaker": "Fraser Paterson"}, {"start": 298.686, "end": 301.128, "text": "I've begun filling in the map for chapter five.", "speaker": "Fraser Paterson"}, {"start": 301.949, "end": 307.075, "text": "So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid.", "speaker": "Fraser Paterson"}, {"start": 307.73, "end": 334.446, "text": " the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that", "speaker": "Fraser Paterson"}, {"start": 334.814, "end": 338.84, "text": " A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation.", "speaker": "Fraser Paterson"}, {"start": 339.401, "end": 353.881, "text": "What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you.", "speaker": "Fraser Paterson"}, {"start": 354.322, "end": 359.269, "text": "You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you.", "speaker": "Fraser Paterson"}, {"start": 359.249, "end": 363.355, "text": " So it's not ideal, but the content is there at least.", "speaker": "Fraser Paterson"}, {"start": 363.736, "end": 367.081, "text": "And especially for those who don't have the book, I think this is quite useful.", "speaker": "Fraser Paterson"}, {"start": 367.101, "end": 374.753, "text": "So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter.", "speaker": "Fraser Paterson"}, {"start": 375.274, "end": 382.405, "text": "I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts", "speaker": "Fraser Paterson"}, {"start": 382.975, "end": 403.431, "text": " uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there", "speaker": "Fraser Paterson"}, {"start": 404.373, "end": 404.954, "text": " Any questions?", "speaker": "Fraser Paterson"}, {"start": 404.994, "end": 416.831, "text": "Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that.", "speaker": "Fraser Paterson"}, {"start": 418.033, "end": 419.795, "text": "I'll start my share temporarily.", "speaker": "Fraser Paterson"}, {"start": 420.756, "end": 421.858, "text": "Just coming back.", "speaker": "Fraser Paterson"}, {"start": 423.863, "end": 424.404, "text": " Very good.", "speaker": "Fraser Paterson"}, {"start": 425.745, "end": 426.386, "text": "Yes, but it's limited.", "speaker": "Fraser Paterson"}, {"start": 428.008, "end": 430.991, "text": "Formatting in LaTeX is a bit strange in Coda.", "speaker": "Fraser Paterson"}, {"start": 432.032, "end": 442.664, "text": "We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda.", "speaker": "Fraser Paterson"}, {"start": 444.026, "end": 445.928, "text": "OK, so I'll share, come back.", "speaker": "Fraser Paterson"}, {"start": 447.89, "end": 452.976, "text": "Let's go all the way back, hopefully people can see my screen again, to the introduction.", "speaker": "Fraser Paterson"}, {"start": 454.441, "end": 456.885, "text": " So everyone's seen this figure here.", "speaker": "Fraser Paterson"}, {"start": 456.905, "end": 458.408, "text": "This is essentially the first figure in the book.", "speaker": "Fraser Paterson"}, {"start": 458.849, "end": 461.814, "text": "This is the breakdown of the partitions of the book, so part one, part two, part three.", "speaker": "Fraser Paterson"}, {"start": 462.396, "end": 468.046, "text": "We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding.", "speaker": "Fraser Paterson"}, {"start": 469.649, "end": 471.232, "text": "Part two is active inference core.", "speaker": "Fraser Paterson"}, {"start": 471.252, "end": 473.215, "text": "So this is the heart of the book, really.", "speaker": "Fraser Paterson"}, {"start": 473.255, "end": 476.922, "text": "This is the fundamentals of active inference per se.", "speaker": "Fraser Paterson"}, {"start": 476.902, "end": 498.64, "text": " what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end", "speaker": "Fraser Paterson"}, {"start": 499.328, "end": 504.875, "text": " But as has been the case, everything has been just perception only.", "speaker": "Fraser Paterson"}, {"start": 504.895, "end": 506.057, "text": "So we haven't actually dealt with actions.", "speaker": "Fraser Paterson"}, {"start": 506.437, "end": 508.24, "text": "And we're going to be moving into that with part two.", "speaker": "Fraser Paterson"}, {"start": 509.221, "end": 511.564, "text": "But even more fundamentally, everything has been static.", "speaker": "Fraser Paterson"}, {"start": 511.824, "end": 516.651, "text": "So all of the models we've been looking at, I don't know if I can get a nice example.", "speaker": "Fraser Paterson"}, {"start": 517.592, "end": 519.014, "text": "All of the models, so this is chapter one.", "speaker": "Fraser Paterson"}, {"start": 519.995, "end": 523.159, "text": "Let's maybe go down to some of the equations.", "speaker": "Fraser Paterson"}, {"start": 523.443, "end": 524.244, "text": " They've all been static.", "speaker": "Fraser Paterson"}, {"start": 524.264, "end": 527.387, "text": "So the environment hasn't really been changing at all.", "speaker": "Fraser Paterson"}, {"start": 527.407, "end": 530.371, "text": "And this is very simple.", "speaker": "Fraser Paterson"}, {"start": 531.372, "end": 534.235, "text": "So part one, really going to chapter two, I think.", "speaker": "Fraser Paterson"}, {"start": 536.858, "end": 538.88, "text": "So we have an environment.", "speaker": "Fraser Paterson"}, {"start": 539.501, "end": 541.503, "text": "This is one we've seen this many times now.", "speaker": "Fraser Paterson"}, {"start": 541.523, "end": 548.691, "text": "This is a representation of the generative process, the real process in and of itself out there.", "speaker": "Fraser Paterson"}, {"start": 549.262, "end": 553.407, "text": " This is fairly static, so nothing is really changing.", "speaker": "Fraser Paterson"}, {"start": 553.467, "end": 555.609, "text": "The hidden states aren't really changing from moment to moment.", "speaker": "Fraser Paterson"}, {"start": 555.629, "end": 562.397, "text": "We're just presented with a situation and we have to update our beliefs about what hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 562.437, "end": 564.199, "text": "That's been constant throughout all of Chapter 1.", "speaker": "Fraser Paterson"}, {"start": 564.879, "end": 566.721, "text": "Things have gotten progressively more complicated.", "speaker": "Fraser Paterson"}, {"start": 566.741, "end": 570.906, "text": "We've looked at univariate hidden states where there's literally just one hidden state.", "speaker": "Fraser Paterson"}, {"start": 571.567, "end": 577.193, "text": "We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things.", "speaker": "Fraser Paterson"}, {"start": 577.662, "end": 579.304, "text": " But in every case, it's been static.", "speaker": "Fraser Paterson"}, {"start": 579.884, "end": 585.331, "text": "The dynamics of the hidden state have been non-existent.", "speaker": "Fraser Paterson"}, {"start": 585.351, "end": 586.992, "text": "That is going to change going into part two.", "speaker": "Fraser Paterson"}, {"start": 587.052, "end": 591.918, "text": "We're going to start looking at generative processes which change over time.", "speaker": "Fraser Paterson"}, {"start": 592.579, "end": 593.94, "text": "And they're the interesting ones.", "speaker": "Fraser Paterson"}, {"start": 593.98, "end": 596.283, "text": "They're the ones that are actually useful.", "speaker": "Fraser Paterson"}, {"start": 596.343, "end": 602.55, "text": "And they're the ones that allow us to actually begin to have a reason to act in the environment.", "speaker": "Fraser Paterson"}, {"start": 602.61, "end": 605.593, "text": "We haven't really had any reason to perform actions yet.", "speaker": "Fraser Paterson"}, {"start": 606.805, "end": 611.032, "text": " So that's a big move that's going to take place and that we're going to have to deal with.", "speaker": "Fraser Paterson"}, {"start": 612.074, "end": 613.517, "text": "So maybe just coming back.", "speaker": "Fraser Paterson"}, {"start": 613.557, "end": 618.325, "text": "So chapter one was about, what was this about?", "speaker": "Fraser Paterson"}, {"start": 618.345, "end": 619.187, "text": "It was about perception.", "speaker": "Fraser Paterson"}, {"start": 619.427, "end": 619.588, "text": "Okay.", "speaker": "Fraser Paterson"}, {"start": 619.628, "end": 624.176, "text": "So like in terms of how we're going to think about perception in active inference.", "speaker": "Fraser Paterson"}, {"start": 624.757, "end": 627.261, "text": "So in chapter one, we cast or we framed", "speaker": "Fraser Paterson"}, {"start": 630.565, "end": 654.058, "text": " of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at", "speaker": "Fraser Paterson"}, {"start": 655.118, "end": 661.187, "text": " you know, how to do like exact Bayesian inference, okay, if we were able to do everything.", "speaker": "Fraser Paterson"}, {"start": 663.11, "end": 665.213, "text": "We saw how to use Bayes' rule basically.", "speaker": "Fraser Paterson"}, {"start": 665.233, "end": 675.187, "text": "So we cemented the distinction personally before that around the difference between the generative process and the generative model.", "speaker": "Fraser Paterson"}, {"start": 675.989, "end": 679.354, "text": "So I come all the way down to the figures here.", "speaker": "Fraser Paterson"}, {"start": 679.374, "end": 681.076, "text": "I don't have a particularly good figure for that.", "speaker": "Fraser Paterson"}, {"start": 681.096, "end": 682.358, "text": "Here we go, 2.4.", "speaker": "Fraser Paterson"}, {"start": 682.793, "end": 691.766, "text": " The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world.", "speaker": "Fraser Paterson"}, {"start": 692.307, "end": 695.191, "text": "This was the crucial thing, I think, in chapter two, basically.", "speaker": "Fraser Paterson"}, {"start": 695.992, "end": 699.978, "text": "And we have a way of notating this convention.", "speaker": "Fraser Paterson"}, {"start": 699.998, "end": 708.531, "text": "We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 708.971, "end": 712.176, "text": "These belong to the generative process, the real world out there.", "speaker": "Fraser Paterson"}, {"start": 712.797, "end": 716.822, "text": " Okay, and then modeled variables are without a star.", "speaker": "Fraser Paterson"}, {"start": 717.142, "end": 719.184, "text": "So they correspond to our model.", "speaker": "Fraser Paterson"}, {"start": 720.746, "end": 721.467, "text": "And what is our model?", "speaker": "Fraser Paterson"}, {"start": 722.668, "end": 730.037, "text": "It is quite literally a joint probability distribution over hidden states, latent states, and observations.", "speaker": "Fraser Paterson"}, {"start": 730.978, "end": 733.421, "text": "We use x for latent states and y for observations.", "speaker": "Fraser Paterson"}, {"start": 734.722, "end": 742.551, "text": "But you can see, of course, the x in here is our model of the real latent state outside of here, x star.", "speaker": "Fraser Paterson"}, {"start": 743.172, "end": 744.755, "text": " And they need not agree with each other.", "speaker": "Fraser Paterson"}, {"start": 744.795, "end": 745.737, "text": "They need not be the same.", "speaker": "Fraser Paterson"}, {"start": 746.839, "end": 752.69, "text": "But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 752.731, "end": 754.795, "text": "That is p of x and y.", "speaker": "Fraser Paterson"}, {"start": 755.817, "end": 760.185, "text": "And then if we have the model evidence, the probability of observing anything,", "speaker": "Fraser Paterson"}, {"start": 760.722, "end": 769.453, "text": " or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation.", "speaker": "Fraser Paterson"}, {"start": 770.174, "end": 772.037, "text": "And that was kind of chapter two.", "speaker": "Fraser Paterson"}, {"start": 772.057, "end": 779.526, "text": "And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions.", "speaker": "Fraser Paterson"}, {"start": 779.546, "end": 780.648, "text": "And these are very ubiquitous.", "speaker": "Fraser Paterson"}, {"start": 780.668, "end": 781.369, "text": "They show up everywhere.", "speaker": "Fraser Paterson"}, {"start": 782.43, "end": 783.331, "text": "We're going to continue that.", "speaker": "Fraser Paterson"}, {"start": 783.351, "end": 790.24, "text": "In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference.", "speaker": "Fraser Paterson"}, {"start": 790.642, "end": 791.784, "text": " There are other distributions as well.", "speaker": "Fraser Paterson"}, {"start": 791.804, "end": 792.725, "text": "We haven't really looked at those yet.", "speaker": "Fraser Paterson"}, {"start": 793.967, "end": 799.554, "text": "But the real sort of, there's a problem with this, which is that, yes, we can do inference.", "speaker": "Fraser Paterson"}, {"start": 799.574, "end": 803.079, "text": "Yes, we might even be able to do exact inference for really, really simple problems.", "speaker": "Fraser Paterson"}, {"start": 804.14, "end": 812.952, "text": "But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model.", "speaker": "Fraser Paterson"}, {"start": 813.793, "end": 815.075, "text": "And then we're going to do inference.", "speaker": "Fraser Paterson"}, {"start": 815.095, "end": 820.142, "text": "And inference is like, I set the dials, and then my machine does some inference stuff, right?", "speaker": "Fraser Paterson"}, {"start": 820.983, "end": 847.716, "text": " that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation", "speaker": "Fraser Paterson"}, {"start": 848.303, "end": 850.966, "text": " And we're doing good old fashioned Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 851.346, "end": 854.99, "text": "But now we're having to do parameter learning and estimation as well.", "speaker": "Fraser Paterson"}, {"start": 856.692, "end": 867.903, "text": "And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference.", "speaker": "Fraser Paterson"}, {"start": 867.923, "end": 872.448, "text": "So we have these two processes, inference and learning parameter estimation.", "speaker": "Fraser Paterson"}, {"start": 873.109, "end": 877.213, "text": "And they're both able to be done by means of Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 878.56, "end": 879.421, "text": " That's the crucial thing.", "speaker": "Fraser Paterson"}, {"start": 879.441, "end": 889.438, "text": "So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah.", "speaker": "Fraser Paterson"}, {"start": 889.458, "end": 904.163, "text": "And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent.", "speaker": "Fraser Paterson"}, {"start": 904.868, "end": 908.953, "text": " And we saw many versions of this across the chapter and indeed in other chapters.", "speaker": "Fraser Paterson"}, {"start": 908.993, "end": 920.085, "text": "It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters.", "speaker": "Fraser Paterson"}, {"start": 920.185, "end": 926.172, "text": "So let's imagine we have our two parameters, I don't know, beta 1, beta 0.", "speaker": "Fraser Paterson"}, {"start": 926.232, "end": 934.521, "text": "And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space.", "speaker": "Fraser Paterson"}, {"start": 934.872, "end": 935.733, "text": " That's this surface.", "speaker": "Fraser Paterson"}, {"start": 936.754, "end": 944.063, "text": "And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point.", "speaker": "Fraser Paterson"}, {"start": 945.064, "end": 950.271, "text": "And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters.", "speaker": "Fraser Paterson"}, {"start": 950.291, "end": 952.654, "text": "Very general way of doing optimization.", "speaker": "Fraser Paterson"}, {"start": 954.015, "end": 964.288, "text": "However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be.", "speaker": "Fraser Paterson"}, {"start": 965.72, "end": 973.231, "text": " And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization.", "speaker": "Fraser Paterson"}, {"start": 975.574, "end": 985.308, "text": "So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now.", "speaker": "Fraser Paterson"}, {"start": 986.25, "end": 993.921, "text": "A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 994.357, "end": 1021.09, "text": " um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model", "speaker": "Fraser Paterson"}, {"start": 1021.525, "end": 1025.652, "text": " And also, how should I use those settings to do good inference?", "speaker": "Fraser Paterson"}, {"start": 1026.934, "end": 1040.558, "text": "It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters.", "speaker": "Fraser Paterson"}, {"start": 1040.578, "end": 1042.442, "text": "So imagine you've got learning up here.", "speaker": "Fraser Paterson"}, {"start": 1042.582, "end": 1043.744, "text": "What should the model parameters be?", "speaker": "Fraser Paterson"}, {"start": 1043.764, "end": 1046.048, "text": "And you can do inference on model parameters.", "speaker": "Fraser Paterson"}, {"start": 1046.45, "end": 1050.538, "text": " And then that can inform how you do inference at the low-level hidden states.", "speaker": "Fraser Paterson"}, {"start": 1050.558, "end": 1053.805, "text": "So there's kind of these two processes that are happening in two different timescales.", "speaker": "Fraser Paterson"}, {"start": 1053.825, "end": 1062.062, "text": "That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally.", "speaker": "Fraser Paterson"}, {"start": 1063.122, "end": 1064.023, "text": " So that's kind of chapter three.", "speaker": "Fraser Paterson"}, {"start": 1064.143, "end": 1065.264, "text": "I'm going to race through this.", "speaker": "Fraser Paterson"}, {"start": 1065.925, "end": 1068.828, "text": "And we'll get to chapter five and then to some actual questions from you guys.", "speaker": "Fraser Paterson"}, {"start": 1069.749, "end": 1081.062, "text": "Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference.", "speaker": "Fraser Paterson"}, {"start": 1081.082, "end": 1084.966, "text": "So chapter three, we're still doing exact inference.", "speaker": "Fraser Paterson"}, {"start": 1084.986, "end": 1090.632, "text": "But we saw that really we culminated in an algorithm", "speaker": "Fraser Paterson"}, {"start": 1091.658, "end": 1113.632, "text": " allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be", "speaker": "Fraser Paterson"}, {"start": 1113.95, "end": 1115.672, "text": " Those two problems are related to one another.", "speaker": "Fraser Paterson"}, {"start": 1115.752, "end": 1118.615, "text": "To know the hidden states, you need to know what the good parameters are.", "speaker": "Fraser Paterson"}, {"start": 1118.935, "end": 1121.378, "text": "But to know what the good parameters are, you have to know how to infer hidden states well.", "speaker": "Fraser Paterson"}, {"start": 1121.958, "end": 1131.588, "text": "So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm.", "speaker": "Fraser Paterson"}, {"start": 1131.608, "end": 1132.689, "text": "I won't go over that again here.", "speaker": "Fraser Paterson"}, {"start": 1132.73, "end": 1134.131, "text": "Maybe we will if people want me to.", "speaker": "Fraser Paterson"}, {"start": 1135.072, "end": 1139.957, "text": "But that was the solution to the problem of chapter 3, that chicken and egg problem.", "speaker": "Fraser Paterson"}, {"start": 1140.781, "end": 1151.908, "text": " The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations.", "speaker": "Fraser Paterson"}, {"start": 1151.928, "end": 1155.938, "text": "And that, unfortunately, is usually not something we can do.", "speaker": "Fraser Paterson"}, {"start": 1156.542, "end": 1161.508, "text": " Being able to find the exact posterior is usually impossible, for reasons that we saw.", "speaker": "Fraser Paterson"}, {"start": 1161.708, "end": 1162.75, "text": "We can review again if you want.", "speaker": "Fraser Paterson"}, {"start": 1163.33, "end": 1166.895, "text": "So then chapter 4 says, all right, well, we can't do exact Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1167.536, "end": 1168.096, "text": "What are we going to do?", "speaker": "Fraser Paterson"}, {"start": 1168.116, "end": 1177.107, "text": "We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like.", "speaker": "Fraser Paterson"}, {"start": 1178.029, "end": 1180.812, "text": "So the idea with chapter 4 is, ah, OK, we're going to now", "speaker": "Fraser Paterson"}, {"start": 1181.804, "end": 1193.779, "text": " have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 1194.7, "end": 1198.304, "text": "We're going to try and find a posterior that's close enough to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1198.705, "end": 1208.537, "text": "And this then motivated the idea of variational free energy as a quantity that is a measurement of how well", "speaker": "Fraser Paterson"}, {"start": 1208.838, "end": 1213.784, "text": " an approximate posterior fits or how close it is to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1214.885, "end": 1222.794, "text": "And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal.", "speaker": "Fraser Paterson"}, {"start": 1223.055, "end": 1225.858, "text": "We saw from chapter three that surprisal is something we want to make very small.", "speaker": "Fraser Paterson"}, {"start": 1226.719, "end": 1237.191, "text": "So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal,", "speaker": "Fraser Paterson"}, {"start": 1237.778, "end": 1242.586, "text": " which is something we can't do directly, and therefore we can do approximate Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1242.606, "end": 1248.976, "text": "So in a lot of ways, that really kind of spiritually is the end of part one, I would say.", "speaker": "Fraser Paterson"}, {"start": 1249.958, "end": 1265.102, "text": "The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE.", "speaker": "Fraser Paterson"}, {"start": 1265.842, "end": 1269.509, "text": " That's, in my mind, really kind of the end of part one.", "speaker": "Fraser Paterson"}, {"start": 1270.311, "end": 1278.908, "text": "Chapter five was a nice kind of detour into a different way of thinking about what the VFE is.", "speaker": "Fraser Paterson"}, {"start": 1280.451, "end": 1287.445, "text": "So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 1288.387, "end": 1314.799, "text": " we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical", "speaker": "Fraser Paterson"}, {"start": 1314.965, "end": 1334.813, "text": " know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was", "speaker": "Fraser Paterson"}, {"start": 1335.535, "end": 1349.242, "text": " lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning.", "speaker": "Fraser Paterson"}, {"start": 1350.083, "end": 1357.257, "text": "It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization", "speaker": "Fraser Paterson"}, {"start": 1357.574, "end": 1369.356, "text": " in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure", "speaker": "Fraser Paterson"}, {"start": 1370.163, "end": 1387.641, "text": " And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow.", "speaker": "Fraser Paterson"}, {"start": 1388.962, "end": 1391.865, "text": "And that's kind of where we left off with part one.", "speaker": "Fraser Paterson"}, {"start": 1391.885, "end": 1396.39, "text": "So that is, in effect, what we've done, where we've gone.", "speaker": "Fraser Paterson"}, {"start": 1396.943, "end": 1400.268, "text": " That's a lot of stuff to cover, but again, we haven't looked at actions yet.", "speaker": "Fraser Paterson"}, {"start": 1400.829, "end": 1407.12, "text": "We're gonna be doing that in part two, and that will bring us full circle to active inference.", "speaker": "Fraser Paterson"}, {"start": 1407.14, "end": 1420.021, "text": "And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception.", "speaker": "Fraser Paterson"}, {"start": 1420.922, "end": 1423.086, "text": "Okay, we're gonna see that these are deeply related to each other.", "speaker": "Fraser Paterson"}, {"start": 1423.623, "end": 1426.045, "text": " But that's the ground that we've covered.", "speaker": "Fraser Paterson"}, {"start": 1426.846, "end": 1429.829, "text": "I'd be interested if people have questions related to that.", "speaker": "Fraser Paterson"}, {"start": 1429.849, "end": 1431.29, "text": "I see that there's lots of questions.", "speaker": "Fraser Paterson"}, {"start": 1432.371, "end": 1432.912, "text": "Stop sharing.", "speaker": "Fraser Paterson"}, {"start": 1434.533, "end": 1437.476, "text": "And I see Mark has a question.", "speaker": "Fraser Paterson"}, {"start": 1437.496, "end": 1438.397, "text": "Please fire away, Mark.", "speaker": "Fraser Paterson"}, {"start": 1440.039, "end": 1440.519, "text": "As usual.", "speaker": "Marc Broberg"}, {"start": 1441.16, "end": 1443.041, "text": "Thank you so much for this review.", "speaker": "Marc Broberg"}, {"start": 1443.061, "end": 1443.602, "text": "This is great.", "speaker": "Marc Broberg"}, {"start": 1445.724, "end": 1451.389, "text": "I was wondering if you could pull up that slide that showed the generative process and the generative model.", "speaker": "Marc Broberg"}, {"start": 1451.429, "end": 1453.191, "text": "That might be helpful as a reference.", "speaker": "Marc Broberg"}, {"start": 1453.964, "end": 1455.166, "text": " Oh, yeah, okay.", "speaker": "Fraser Paterson"}, {"start": 1455.186, "end": 1456.187, "text": "So let me come back here.", "speaker": "Fraser Paterson"}, {"start": 1458.43, "end": 1462.415, "text": "That's the one in chapter two, I assume you're referring to?", "speaker": "Fraser Paterson"}, {"start": 1462.435, "end": 1462.836, "text": "I think so.", "speaker": "Marc Broberg"}, {"start": 1463.697, "end": 1464.138, "text": "This one here?", "speaker": "Fraser Paterson"}, {"start": 1464.618, "end": 1465.479, "text": "Yes, thank you.", "speaker": "Marc Broberg"}, {"start": 1466.1, "end": 1470.126, "text": "Yeah, I really just kind of appreciate what this textbook is trying to do.", "speaker": "Marc Broberg"}, {"start": 1470.146, "end": 1473.57, "text": "And I think it took me a while to appreciate that.", "speaker": "Marc Broberg"}, {"start": 1473.69, "end": 1482.402, "text": "But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening.", "speaker": "Marc Broberg"}, {"start": 1482.382, "end": 1505.689, "text": " in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise", "speaker": "Marc Broberg"}, {"start": 1505.905, "end": 1532.082, "text": " in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of", "speaker": "Fraser Paterson"}, {"start": 1532.973, "end": 1538.437, "text": " It seems like chapter four kind of serves as a background where the story gets started for the most part.", "speaker": "Marc Broberg"}, {"start": 1538.458, "end": 1542.375, "text": "And this is kind of more of a ground level", "speaker": "Marc Broberg"}, {"start": 1542.457, "end": 1544.78, "text": " view, which is nice and helpful.", "speaker": "Marc Broberg"}, {"start": 1546.603, "end": 1554.433, "text": "What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function.", "speaker": "Marc Broberg"}, {"start": 1554.453, "end": 1559.68, "text": "Because in my mind, I'm kind of conditioned to seeing these Gaussians, right?", "speaker": "Marc Broberg"}, {"start": 1559.7, "end": 1566.089, "text": "You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior.", "speaker": "Marc Broberg"}, {"start": 1566.75, "end": 1571.316, "text": "However, here we have this linear function, which I guess is", "speaker": "Marc Broberg"}, {"start": 1572.207, "end": 1575.451, "text": " then becomes part of, say, a likelihood.", "speaker": "Marc Broberg"}, {"start": 1575.491, "end": 1588.409, "text": "And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense.", "speaker": "Marc Broberg"}, {"start": 1588.429, "end": 1588.83, "text": "Yeah, yeah.", "speaker": "Fraser Paterson"}, {"start": 1589.651, "end": 1590.212, "text": "No, absolutely.", "speaker": "Fraser Paterson"}, {"start": 1590.252, "end": 1600.746, "text": "I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1601.131, "end": 1628.91, "text": " um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so", "speaker": "Fraser Paterson"}, {"start": 1629.464, "end": 1638.597, "text": " This is, in some sense, still our assumption about what the relationship is between observations and the hidden states.", "speaker": "Fraser Paterson"}, {"start": 1639.066, "end": 1646.86, "text": " For our purpose, we've assumed that the real world, we're sort of creating a world here.", "speaker": "Fraser Paterson"}, {"start": 1646.94, "end": 1651.348, "text": "This is the example of the food size and light intensity world.", "speaker": "Fraser Paterson"}, {"start": 1651.468, "end": 1653.572, "text": "This is the really simple world that we've seen for all of part one.", "speaker": "Fraser Paterson"}, {"start": 1654.153, "end": 1655.475, "text": "So what are the hidden states?", "speaker": "Fraser Paterson"}, {"start": 1656.076, "end": 1658.04, "text": "They are the sizes of the food.", "speaker": "Fraser Paterson"}, {"start": 1658.46, "end": 1661.666, "text": "And we can see that there's five different sizes, one to five.", "speaker": "Fraser Paterson"}, {"start": 1661.907, "end": 1662.728, "text": "That's the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1663.552, "end": 1665.214, "text": " And the observations are light intensities.", "speaker": "Fraser Paterson"}, {"start": 1665.354, "end": 1668.318, "text": "So light comes in, hits the food, and you get some sort of light intensity.", "speaker": "Fraser Paterson"}, {"start": 1668.758, "end": 1672.903, "text": "And depending on the size of the food, you get some specific light intensity.", "speaker": "Fraser Paterson"}, {"start": 1672.924, "end": 1685.619, "text": "So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size.", "speaker": "Fraser Paterson"}, {"start": 1685.639, "end": 1688.783, "text": "So we're saying that the relationship between the observation you get in", "speaker": "Fraser Paterson"}, {"start": 1689.235, "end": 1709.573, "text": " is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters", "speaker": "Fraser Paterson"}, {"start": 1710.548, "end": 1714.279, "text": " So again, this is the environment regenerative process.", "speaker": "Fraser Paterson"}, {"start": 1714.941, "end": 1721.942, "text": "The corresponding model is this one, 2.11, equation 2.11, I believe.", "speaker": "Fraser Paterson"}, {"start": 1722.513, "end": 1751.187, "text": " now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs", "speaker": "Fraser Paterson"}, {"start": 1751.555, "end": 1762.486, "text": " are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution.", "speaker": "Fraser Paterson"}, {"start": 1762.526, "end": 1780.684, "text": "So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations.", "speaker": "Fraser Paterson"}, {"start": 1781.204, "end": 1783.246, "text": " And the thing is, we're uncertain.", "speaker": "Fraser Paterson"}, {"start": 1784.268, "end": 1789.194, "text": "So the agent doesn't get to see what the real relationship is between food size and light intensity.", "speaker": "Fraser Paterson"}, {"start": 1789.334, "end": 1789.995, "text": "It has no idea.", "speaker": "Fraser Paterson"}, {"start": 1791.136, "end": 1793.679, "text": "It has a model about what it thinks the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1794.24, "end": 1796.443, "text": "Now, we know that the relationship is linear.", "speaker": "Fraser Paterson"}, {"start": 1796.483, "end": 1805.994, "text": "We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity.", "speaker": "Fraser Paterson"}, {"start": 1806.014, "end": 1809.118, "text": "But in the real world, we have no idea what the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1809.385, "end": 1813.791, "text": " It turns out that in this example, we've specified exactly the same relationship.", "speaker": "Fraser Paterson"}, {"start": 1814.752, "end": 1816.975, "text": "So that's quite a nice scenario to be in.", "speaker": "Fraser Paterson"}, {"start": 1817.676, "end": 1826.968, "text": "You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size.", "speaker": "Fraser Paterson"}, {"start": 1828.429, "end": 1834.397, "text": "Now that's wrong, but it's OK, maybe, for some settings.", "speaker": "Fraser Paterson"}, {"start": 1834.985, "end": 1843.699, "text": " That would be an example of a likelihood mapping that would be incorrect with respect to its generating function.", "speaker": "Fraser Paterson"}, {"start": 1844.721, "end": 1861.448, "text": "So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount.", "speaker": "Fraser Paterson"}, {"start": 1862.103, "end": 1869.693, "text": " And it's centered about the relationship we think is the case between the hidden states and the observations.", "speaker": "Fraser Paterson"}, {"start": 1870.234, "end": 1873.658, "text": "So the reason why we have this probability distribution is because we don't know.", "speaker": "Fraser Paterson"}, {"start": 1873.678, "end": 1877.924, "text": "We do not actually ever get to know what the real relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1877.944, "end": 1881.208, "text": "So does that answer your question, Mark, about why", "speaker": "Fraser Paterson"}, {"start": 1881.458, "end": 1887.967, "text": " why probability distributions, maybe why normal distributions, or did that help at all?", "speaker": "Fraser Paterson"}, {"start": 1888.007, "end": 1889.469, "text": "Yeah, that helped a lot.", "speaker": "Marc Broberg"}, {"start": 1890.511, "end": 1890.951, "text": "Thank you.", "speaker": "Marc Broberg"}, {"start": 1891.011, "end": 1900.164, "text": "Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right?", "speaker": "Marc Broberg"}, {"start": 1900.464, "end": 1908.175, "text": "And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right?", "speaker": "Marc Broberg"}, {"start": 1909.235, "end": 1934.061, "text": " uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all", "speaker": "Fraser Paterson"}, {"start": 1934.648, "end": 1938.613, "text": " We could plug, we could say, okay, the relationship is just equal to the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1938.813, "end": 1942.738, "text": "I think that the observation is exactly the hidden state, right?", "speaker": "Fraser Paterson"}, {"start": 1942.758, "end": 1943.438, "text": "That's an assumption.", "speaker": "Fraser Paterson"}, {"start": 1944.86, "end": 1952.529, "text": "Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1953.49, "end": 1955.653, "text": "Or we can have a quadratic relationship.", "speaker": "Fraser Paterson"}, {"start": 1955.673, "end": 1957.635, "text": "We can have anything we want in there at all, actually.", "speaker": "Fraser Paterson"}, {"start": 1957.856, "end": 1963.843, "text": "And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1964.083, "end": 1964.183, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 1966.036, "end": 1983.758, "text": " very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B,", "speaker": "Fraser Paterson"}, {"start": 1986.978, "end": 2006.903, "text": " Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from.", "speaker": "Fraser Paterson"}, {"start": 2006.923, "end": 2009.526, "text": "Michael Morehead , Other questions other questions Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2011.582, "end": 2015.987, "text": " Okay, my question is this.", "speaker": "Magdalena Hurtado"}, {"start": 2016.608, "end": 2019.672, "text": "So could we do a thought experiment for a minute?", "speaker": "Magdalena Hurtado"}, {"start": 2019.712, "end": 2025.319, "text": "Because I think if I set it up as a thought experiment, I'll get the answer.", "speaker": "Magdalena Hurtado"}, {"start": 2026.821, "end": 2027.121, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2027.221, "end": 2027.942, "text": "Oh, so I'm inside.", "speaker": "Magdalena Hurtado"}, {"start": 2027.962, "end": 2030.966, "text": "Okay, so let's imagine the following.", "speaker": "Magdalena Hurtado"}, {"start": 2031.907, "end": 2040.357, "text": "In present-day societies, we have this societal", "speaker": "Magdalena Hurtado"}, {"start": 2040.775, "end": 2050.769, "text": " back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2052.391, "end": 2060.623, "text": "And there's a tension in societies to update that belief by some group to science.", "speaker": "Magdalena Hurtado"}, {"start": 2060.89, "end": 2063.737, "text": " Exists a science so God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2063.837, "end": 2071.114, "text": "So science tells us that God may not exist or something to that effect or quantum physics explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2071.214, "end": 2071.735, "text": "Not God.", "speaker": "Magdalena Hurtado"}, {"start": 2072.016, "end": 2072.597, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2072.617, "end": 2074.682, "text": "So there's a process, right?", "speaker": "Magdalena Hurtado"}, {"start": 2074.782, "end": 2077.228, "text": "That's in in the environment.", "speaker": "Magdalena Hurtado"}, {"start": 2077.208, "end": 2096.28, "text": " that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed", "speaker": "Magdalena Hurtado"}, {"start": 2096.733, "end": 2118.994, "text": " And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications.", "speaker": "Magdalena Hurtado"}, {"start": 2118.974, "end": 2148.88, "text": " for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is", "speaker": "Magdalena Hurtado"}, {"start": 2148.86, "end": 2163.055, "text": " Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2165.296, "end": 2191.017, "text": " are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds", "speaker": "Magdalena Hurtado"}, {"start": 2191.25, "end": 2195.815, "text": " Okay, as a third level of information processing.", "speaker": "Magdalena Hurtado"}, {"start": 2196.296, "end": 2217.46, "text": "And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2217.44, "end": 2219.723, "text": " Yeah, no, excellent, excellent question, Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2219.964, "end": 2242.857, "text": "This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents.", "speaker": "Fraser Paterson"}, {"start": 2243.748, "end": 2250.721, "text": " Under what conditions do they interact so as to form a larger composite active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2251.462, "end": 2257.513, "text": "So it's kind of a question like, how many active inference agents are there really in any given picture?", "speaker": "Fraser Paterson"}, {"start": 2260.018, "end": 2266.87, "text": "So your question, I think, insofar as I understand it, is, look, we have some notion about agency.", "speaker": "Fraser Paterson"}, {"start": 2267.423, "end": 2268.825, "text": " inactive inference.", "speaker": "Fraser Paterson"}, {"start": 2268.845, "end": 2274.433, "text": "We have some notion about something interacting with its environment, whatever the environment is, right?", "speaker": "Fraser Paterson"}, {"start": 2274.453, "end": 2278.098, "text": "So on the right hand side, we've got an agent, and on the left hand side, we have an environment.", "speaker": "Fraser Paterson"}, {"start": 2279.36, "end": 2293.699, "text": "But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent.", "speaker": "Fraser Paterson"}, {"start": 2293.84, "end": 2295.642, "text": "Okay, we haven't been able to tell that story yet.", "speaker": "Fraser Paterson"}, {"start": 2296.837, "end": 2311.04, "text": " Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers.", "speaker": "Fraser Paterson"}, {"start": 2311.12, "end": 2314.385, "text": "Let me see if I can find that figure.", "speaker": "Fraser Paterson"}, {"start": 2315.206, "end": 2317.83, "text": "Various layers within the predictive coding hierarchy.", "speaker": "Fraser Paterson"}, {"start": 2318.672, "end": 2319.433, "text": "Which one is it here?", "speaker": "Fraser Paterson"}, {"start": 2319.473, "end": 2320.074, "text": "This one?", "speaker": "Fraser Paterson"}, {"start": 2320.354, "end": 2320.735, "text": "No, this one.", "speaker": "Fraser Paterson"}, {"start": 2322.487, "end": 2329.997, "text": " You might interpret the various layers here as individual active implementations that are just doing local free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 2330.017, "end": 2333.682, "text": "So maybe this guy is an active implementation in some sense.", "speaker": "Fraser Paterson"}, {"start": 2333.702, "end": 2335.825, "text": "And then he's passing messages, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 2336.445, "end": 2338.989, "text": "And then the overall thing can be regarded as an active implementation.", "speaker": "Fraser Paterson"}, {"start": 2339.029, "end": 2342.393, "text": "That's the only inkling that we've seen thus far of this idea yet.", "speaker": "Fraser Paterson"}, {"start": 2343.234, "end": 2347.6, "text": "We're also not really going to see a lot of it in the rest of the book.", "speaker": "Fraser Paterson"}, {"start": 2348.052, "end": 2349.956, "text": " because it is a very open question.", "speaker": "Fraser Paterson"}, {"start": 2350.017, "end": 2351.38, "text": "It's a very difficult question.", "speaker": "Fraser Paterson"}, {"start": 2352.302, "end": 2360.36, "text": "And it gets at the heart of what agency even is at all, which is a question over and above active inference per se.", "speaker": "Fraser Paterson"}, {"start": 2360.781, "end": 2367.276, "text": "However, my personal interpretation or my personal thoughts about this issue", "speaker": "Fraser Paterson"}, {"start": 2367.678, "end": 2374.214, "text": " is that the thing that is currently missing from active inference is exactly this idea of compositionality.", "speaker": "Fraser Paterson"}, {"start": 2374.234, "end": 2375.938, "text": "So I've got agent here, agent here.", "speaker": "Fraser Paterson"}, {"start": 2376.219, "end": 2376.841, "text": "They interact.", "speaker": "Fraser Paterson"}, {"start": 2377.482, "end": 2381.873, "text": "Under what conditions is that whole thing one active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2382.275, "end": 2388.688, "text": " That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research.", "speaker": "Fraser Paterson"}, {"start": 2388.709, "end": 2390.352, "text": "So I hope that answers your question in some respect.", "speaker": "Fraser Paterson"}, {"start": 2390.532, "end": 2391.494, "text": "Very, very important question.", "speaker": "Fraser Paterson"}, {"start": 2392.015, "end": 2397.828, "text": "But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 2398.81, "end": 2399.912, "text": "OK.", "speaker": "Magdalena Hurtado"}, {"start": 2400.23, "end": 2425.562, "text": " really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups", "speaker": "Magdalena Hurtado"}, {"start": 2425.542, "end": 2430.369, "text": " What I see and I'm going to use this language, I'm not a mathematician.", "speaker": "Magdalena Hurtado"}, {"start": 2430.429, "end": 2435.577, "text": "OK, so forgive me, but but I kind of I kind of get some things in math.", "speaker": "Magdalena Hurtado"}, {"start": 2435.597, "end": 2442.167, "text": "OK, so topologically generative models really work as huge attractors.", "speaker": "Magdalena Hurtado"}, {"start": 2442.147, "end": 2444.451, "text": " in informational systems in humans.", "speaker": "Magdalena Hurtado"}, {"start": 2444.912, "end": 2458.638, "text": "So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right?", "speaker": "Magdalena Hurtado"}, {"start": 2458.718, "end": 2462.265, "text": "So your language is a technology in human systems.", "speaker": "Magdalena Hurtado"}, {"start": 2462.245, "end": 2490.784, "text": " so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model", "speaker": "Magdalena Hurtado"}, {"start": 2491.692, "end": 2504.479, "text": " And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math.", "speaker": "Magdalena Hurtado"}, {"start": 2505.702, "end": 2509.55, "text": "If I just kind of sorry, is that okay?", "speaker": "Andrew Pashea"}, {"start": 2509.63, "end": 2509.951, "text": "All right.", "speaker": "Andrew Pashea"}, {"start": 2510.091, "end": 2510.632, "text": "Thanks.", "speaker": "Andrew Pashea"}, {"start": 2510.652, "end": 2512.536, "text": "Just ask someone who has a.", "speaker": "Andrew Pashea"}, {"start": 2512.516, "end": 2515.059, "text": " More immediate background in this social sciences.", "speaker": "Andrew Pashea"}, {"start": 2515.72, "end": 2516.001, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 2516.141, "end": 2516.822, "text": "So, so.", "speaker": "Andrew Pashea"}, {"start": 2517.323, "end": 2521.729, "text": "And I like how the question that Mark had prior to this had to do with, like.", "speaker": "Andrew Pashea"}, {"start": 2522.33, "end": 2524.332, "text": "You know, what is it to make linear assumptions?", "speaker": "Andrew Pashea"}, {"start": 2524.853, "end": 2527.777, "text": "So, in the context of the textbook, what we've seen thus far, like.", "speaker": "Andrew Pashea"}, {"start": 2528.278, "end": 2535.908, "text": "This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2535.948, "end": 2539.774, "text": "And it's just to show how a model can be composed.", "speaker": "Andrew Pashea"}, {"start": 2540.755, "end": 2542.337, "text": "Like, how can.", "speaker": "Andrew Pashea"}, {"start": 2542.57, "end": 2551.986, "text": " In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right?", "speaker": "Andrew Pashea"}, {"start": 2552.407, "end": 2560.16, "text": "It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right?", "speaker": "Andrew Pashea"}, {"start": 2560.18, "end": 2561.422, "text": "So that's the action part.", "speaker": "Andrew Pashea"}, {"start": 2561.783, "end": 2563.245, "text": "We haven't reached that yet.", "speaker": "Andrew Pashea"}, {"start": 2563.225, "end": 2572.756, "text": " But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function?", "speaker": "Andrew Pashea"}, {"start": 2573.317, "end": 2590.597, "text": "So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right?", "speaker": "Andrew Pashea"}, {"start": 2591.198, "end": 2593.08, "text": "So that linear model we were looking at", "speaker": "Andrew Pashea"}, {"start": 2593.06, "end": 2616.688, "text": " we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people", "speaker": "Andrew Pashea"}, {"start": 2616.938, "end": 2619.842, "text": " what are the kinds of sensory information that I receive?", "speaker": "Andrew Pashea"}, {"start": 2620.522, "end": 2622.965, "text": "And what are the kinds of actions that I can do?", "speaker": "Andrew Pashea"}, {"start": 2623.005, "end": 2625.308, "text": "And what do I have beliefs about?", "speaker": "Andrew Pashea"}, {"start": 2625.408, "end": 2629.173, "text": "Or what variables do I think explains everything, right?", "speaker": "Andrew Pashea"}, {"start": 2629.774, "end": 2640.887, "text": "So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human", "speaker": "Andrew Pashea"}, {"start": 2640.867, "end": 2655.518, "text": " uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like", "speaker": "Andrew Pashea"}, {"start": 2656.021, "end": 2658.344, "text": " There's, there's a lot to track.", "speaker": "Andrew Pashea"}, {"start": 2658.364, "end": 2662.21, "text": "So there are some people who have tried to model this sort of thing.", "speaker": "Andrew Pashea"}, {"start": 2662.23, "end": 2663.572, "text": "And so I've shared this paper.", "speaker": "Andrew Pashea"}, {"start": 2664.313, "end": 2672.665, "text": "It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference.", "speaker": "Andrew Pashea"}, {"start": 2673.226, "end": 2676.19, "text": "And this kind of has to do with.", "speaker": "Andrew Pashea"}, {"start": 2676.507, "end": 2693.875, "text": " Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere.", "speaker": "Andrew Pashea"}, {"start": 2693.855, "end": 2711.429, "text": " And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing.", "speaker": "Andrew Pashea"}, {"start": 2711.489, "end": 2714.715, "text": "And that's one way of trying to look at a problem.", "speaker": "Andrew Pashea"}, {"start": 2714.695, "end": 2734.948, "text": " like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one", "speaker": "Andrew Pashea"}, {"start": 2735.333, "end": 2737.86, "text": " And a lot of agents who rally around idea two.", "speaker": "Andrew Pashea"}, {"start": 2737.9, "end": 2743.315, "text": "And as they do that, the two groups start to communicate directly with each other less and less.", "speaker": "Andrew Pashea"}, {"start": 2743.997, "end": 2749.613, "text": "The agent's actual actions that are simulated has to do with them communicating with one another and sharing their", "speaker": "Andrew Pashea"}, {"start": 2750.015, "end": 2753.178, "text": " They're kind of more verbal beliefs with one another.", "speaker": "Andrew Pashea"}, {"start": 2753.199, "end": 2764.771, "text": "And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing.", "speaker": "Andrew Pashea"}, {"start": 2764.791, "end": 2769.256, "text": "And somehow that gets all blended in with our assumptions of how science works and things like that.", "speaker": "Andrew Pashea"}, {"start": 2769.757, "end": 2771.338, "text": "It's much more complex.", "speaker": "Andrew Pashea"}, {"start": 2771.479, "end": 2779.988, "text": "And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy", "speaker": "Andrew Pashea"}, {"start": 2779.968, "end": 2783.633, "text": " But remember that free energy, based on how we've discussed it so far,", "speaker": "Andrew Pashea"}, {"start": 2783.967, "end": 2787.932, "text": " It's not some magical quantity of energy or something like that.", "speaker": "Andrew Pashea"}, {"start": 2788.914, "end": 2790.736, "text": "It's really just a proxy.", "speaker": "Andrew Pashea"}, {"start": 2790.796, "end": 2797.485, "text": "It's a word that is a proxy for something called surprisal, which is more or less uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2797.625, "end": 2799.608, "text": "So we want to minimize our uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2800.169, "end": 2813.687, "text": "So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another,", "speaker": "Andrew Pashea"}, {"start": 2813.667, "end": 2817.235, "text": " about is it idea one or is it idea two?", "speaker": "Andrew Pashea"}, {"start": 2817.716, "end": 2819.781, "text": "Is it science or is it something else?", "speaker": "Andrew Pashea"}, {"start": 2820.443, "end": 2824.993, "text": "One way to minimize uncertainty is to repeatedly tell each other", "speaker": "Andrew Pashea"}, {"start": 2825.328, "end": 2828.993, "text": " what the truth is and you all converge on what you all think the truth is.", "speaker": "Andrew Pashea"}, {"start": 2829.073, "end": 2829.293, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 2829.834, "end": 2838.545, "text": "And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much.", "speaker": "Andrew Pashea"}, {"start": 2839.085, "end": 2841.929, "text": "Because the other group is adding to your uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2842.129, "end": 2849.338, "text": "Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe.", "speaker": "Andrew Pashea"}, {"start": 2849.418, "end": 2854.845, "text": "So, so you see that there, there is a kind of like logic where the free energy principle is involved.", "speaker": "Andrew Pashea"}, {"start": 2855.129, "end": 2864.026, "text": " where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2864.467, "end": 2870.679, "text": "And that has nothing to do with the true validity of science, right?", "speaker": "Andrew Pashea"}, {"start": 2871.641, "end": 2875.348, "text": "Like I'm attempting to be a scientist explaining all this stuff,", "speaker": "Andrew Pashea"}, {"start": 2875.683, "end": 2878.748, "text": " But the people who, you see what I'm saying?", "speaker": "Andrew Pashea"}, {"start": 2878.788, "end": 2884.538, "text": "Like to take in sensory information is what we do.", "speaker": "Andrew Pashea"}, {"start": 2884.578, "end": 2888.544, "text": "It's not about like following proper logic and stuff.", "speaker": "Andrew Pashea"}, {"start": 2888.564, "end": 2892.19, "text": "It's people learn to do that, right?", "speaker": "Andrew Pashea"}, {"start": 2892.17, "end": 2894.737, "text": " Can I interject something real quick?", "speaker": "Magdalena Hurtado"}, {"start": 2894.757, "end": 2906.128, "text": "I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're", "speaker": "Magdalena Hurtado"}, {"start": 2906.581, "end": 2923.723, "text": " Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2924.364, "end": 2931.593, "text": "So what happens in humans groups is that the active inferencing that's happening in a lot of individuals,", "speaker": "Magdalena Hurtado"}, {"start": 2931.573, "end": 2937.991, "text": " is to analyze and update whether or not they believe the person who's giving the message.", "speaker": "Magdalena Hurtado"}, {"start": 2938.552, "end": 2942.864, "text": "It's not about evaluating the validity of the premises.", "speaker": "Magdalena Hurtado"}, {"start": 2943.486, "end": 2947.236, "text": "It's not about the... That gets at a very...", "speaker": "Magdalena Hurtado"}, {"start": 2947.216, "end": 2957.913, "text": " That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer.", "speaker": "Fraser Paterson"}, {"start": 2958.734, "end": 2961.178, "text": "We might imagine an active inference agent doing inference about this.", "speaker": "Fraser Paterson"}, {"start": 2961.759, "end": 2967.688, "text": "But above that is a question, how reliable is this thing that I'm hearing at all?", "speaker": "Fraser Paterson"}, {"start": 2967.668, "end": 2992.039, "text": " and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant", "speaker": "Fraser Paterson"}, {"start": 2992.019, "end": 3018.583, "text": " uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation", "speaker": "Fraser Paterson"}, {"start": 3018.563, "end": 3022.371, "text": " for what it means for one entity to be an active inference agent.", "speaker": "Fraser Paterson"}, {"start": 3022.391, "end": 3031.651, "text": "Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents.", "speaker": "Fraser Paterson"}, {"start": 3031.671, "end": 3033.595, "text": "But that's not going to be the focus of the book.", "speaker": "Fraser Paterson"}, {"start": 3034.396, "end": 3038.445, "text": "A lot of that is at the cutting edge of active inference research now.", "speaker": "Fraser Paterson"}, {"start": 3039.961, "end": 3044.874, "text": " Thanks for keeping it closer to the textbook by bringing up precision.", "speaker": "Andrew Pashea"}, {"start": 3044.934, "end": 3046.117, "text": "I very much agree with that.", "speaker": "Andrew Pashea"}, {"start": 3046.157, "end": 3052.915, "text": "And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but.", "speaker": "Andrew Pashea"}, {"start": 3054.459, "end": 3055.422, "text": "This.", "speaker": "Andrew Pashea"}, {"start": 3055.655, "end": 3065.723, "text": " Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference.", "speaker": "Andrew Pashea"}, {"start": 3065.743, "end": 3070.095, "text": "So active inferences is generally setting up these principles.", "speaker": "Andrew Pashea"}, {"start": 3070.115, "end": 3070.997, "text": "It.", "speaker": "Andrew Pashea"}, {"start": 3070.977, "end": 3092.985, "text": " definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different", "speaker": "Andrew Pashea"}, {"start": 3092.965, "end": 3093.386, "text": " Right?", "speaker": "Andrew Pashea"}, {"start": 3093.806, "end": 3097.411, "text": "Like, our variables that we're including in our respective models are different.", "speaker": "Andrew Pashea"}, {"start": 3097.772, "end": 3099.694, "text": "The precisions can definitely differ.", "speaker": "Andrew Pashea"}, {"start": 3099.714, "end": 3106.303, "text": "If I get information from someone in my group, I might have a higher precision and trust what they say more.", "speaker": "Andrew Pashea"}, {"start": 3106.764, "end": 3114.654, "text": "I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right?", "speaker": "Andrew Pashea"}, {"start": 3115.155, "end": 3118.72, "text": "So a lot of the, you know, a lot of the", "speaker": "Andrew Pashea"}, {"start": 3119.054, "end": 3131.018, "text": " political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right?", "speaker": "Andrew Pashea"}, {"start": 3131.159, "end": 3134.245, "text": "It's like, oh, I just lump them all in with political group", "speaker": "Andrew Pashea"}, {"start": 3134.225, "end": 3156.351, "text": " be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available", "speaker": "Andrew Pashea"}, {"start": 3156.331, "end": 3172.112, "text": " then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone", "speaker": "Andrew Pashea"}, {"start": 3172.092, "end": 3191.299, "text": " ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it", "speaker": "Andrew Pashea"}, {"start": 3192.308, "end": 3195.736, "text": " Do make sure, I've been absent for a little bit.", "speaker": "Fraser Paterson"}, {"start": 3196.237, "end": 3200.687, "text": "I'm gonna be much more active on the questions side of things on the CODA.", "speaker": "Fraser Paterson"}, {"start": 3200.748, "end": 3205.418, "text": "So do make sure if you've got questions, please do put them down in the questions tab here.", "speaker": "Fraser Paterson"}, {"start": 3205.458, "end": 3209.608, "text": "I've gone through and I've already answered quite a few of them for chapter five.", "speaker": "Fraser Paterson"}, {"start": 3209.588, "end": 3224.188, "text": " uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see", "speaker": "Fraser Paterson"}, {"start": 3224.573, "end": 3226.135, "text": " Are there any more live questions?", "speaker": "Fraser Paterson"}, {"start": 3227.777, "end": 3228.718, "text": "That would be nice.", "speaker": "Fraser Paterson"}, {"start": 3228.738, "end": 3233.584, "text": "We can stay around for maybe five or so minutes after the deadline if people so choose.", "speaker": "Fraser Paterson"}, {"start": 3234.645, "end": 3236.267, "text": "But we are coming up to the top of the hour.", "speaker": "Fraser Paterson"}, {"start": 3236.287, "end": 3241.994, "text": "So yeah, all of the questions in chapter five now are answered, or I've given my attempt.", "speaker": "Fraser Paterson"}, {"start": 3242.655, "end": 3246.379, "text": "As I say, I'll go through and attempt to answer all the previous ones as well.", "speaker": "Fraser Paterson"}, {"start": 3249.937, "end": 3276.737, "text": " think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry", "speaker": "Fraser Paterson"}, {"start": 3277.595, "end": 3301.122, "text": " okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian", "speaker": "Sun Xin"}, {"start": 3301.102, "end": 3308.391, "text": " inference to engage in parameter learning so that it can update beliefs about the current state.", "speaker": "Sun Xin"}, {"start": 3309.172, "end": 3319.545, "text": "If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself.", "speaker": "Sun Xin"}, {"start": 3319.985, "end": 3323.97, "text": "And I think this is what we call plasticity.", "speaker": "Sun Xin"}, {"start": 3323.95, "end": 3331.683, "text": " The ultimate goal of all this is to minimize expected free energy.", "speaker": "Sun Xin"}, {"start": 3332.404, "end": 3333.085, "text": "Is that right?", "speaker": "Sun Xin"}, {"start": 3334.548, "end": 3335.088, "text": "Wow, okay.", "speaker": "Fraser Paterson"}, {"start": 3335.129, "end": 3338.174, "text": "I mean, basically, yeah, that's substantively correct.", "speaker": "Fraser Paterson"}, {"start": 3339.115, "end": 3342.06, "text": "I will note, however, we haven't yet talked about expected free energy.", "speaker": "Fraser Paterson"}, {"start": 3342.395, "end": 3344.237, "text": " and planning and action selection.", "speaker": "Fraser Paterson"}, {"start": 3344.257, "end": 3345.358, "text": "So that is going to come later.", "speaker": "Fraser Paterson"}, {"start": 3346.379, "end": 3346.959, "text": "But yeah, you're right.", "speaker": "Fraser Paterson"}, {"start": 3346.979, "end": 3352.825, "text": "In terms of the general flavor of things, we have beliefs about hidden states in the world.", "speaker": "Fraser Paterson"}, {"start": 3353.686, "end": 3365.076, "text": "And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 3365.997, "end": 3369.921, "text": "Or we've seen that we can also recast this in terms of predictive processing.", "speaker": "Fraser Paterson"}, {"start": 3370.66, "end": 3372.943, "text": " minimizing the precision way to prediction errors.", "speaker": "Fraser Paterson"}, {"start": 3372.963, "end": 3373.924, "text": "It's the same kind of story.", "speaker": "Fraser Paterson"}, {"start": 3374.926, "end": 3386.221, "text": "And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states.", "speaker": "Fraser Paterson"}, {"start": 3387.063, "end": 3387.283, "text": "Yes.", "speaker": "Fraser Paterson"}, {"start": 3388.725, "end": 3391.328, "text": "The other thing, you mentioned parameter learning.", "speaker": "Fraser Paterson"}, {"start": 3391.388, "end": 3392.77, "text": "So yes, exactly.", "speaker": "Fraser Paterson"}, {"start": 3392.79, "end": 3393.932, "text": "There's these two problems.", "speaker": "Fraser Paterson"}, {"start": 3394.032, "end": 3397.317, "text": "We have the problem of inferring the hidden states.", "speaker": "Fraser Paterson"}, {"start": 3397.357, "end": 3398.338, "text": "What should the hidden states be?", "speaker": "Fraser Paterson"}, {"start": 3398.977, "end": 3407.788, "text": " But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference.", "speaker": "Fraser Paterson"}, {"start": 3408.529, "end": 3415.017, "text": "So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like.", "speaker": "Fraser Paterson"}, {"start": 3415.738, "end": 3419.183, "text": "But then above that, we need to deal with the problem of what should the setting of the parameters be?", "speaker": "Fraser Paterson"}, {"start": 3420.204, "end": 3426.853, "text": "And we've only really just begun to look at this in terms of prediction errors and hierarchical models.", "speaker": "Fraser Paterson"}, {"start": 3428.034, "end": 3428.815, "text": "But we saw", "speaker": "Fraser Paterson"}, {"start": 3429.149, "end": 3432.913, "text": " how to deal with that in terms of expectation maximization and earlier chapters.", "speaker": "Fraser Paterson"}, {"start": 3433.634, "end": 3443.644, "text": "We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture.", "speaker": "Fraser Paterson"}, {"start": 3443.964, "end": 3450.251, "text": "Because we generally can't do expectation maximization for the reason that we don't know the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 3450.291, "end": 3458.299, "text": "So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah.", "speaker": "Fraser Paterson"}, {"start": 3458.339, "end": 3458.599, "text": "Thank you.", "speaker": "Sun Xin"}, {"start": 3459.49, "end": 3461.011, "text": " No problem.", "speaker": "Fraser Paterson"}, {"start": 3461.032, "end": 3462.473, "text": "Do we have more questions?", "speaker": "Fraser Paterson"}, {"start": 3462.633, "end": 3463.254, "text": "More questions?", "speaker": "Fraser Paterson"}, {"start": 3464.655, "end": 3466.857, "text": "Maybe one more question, if there is one more.", "speaker": "Fraser Paterson"}, {"start": 3467.117, "end": 3469.52, "text": "And then I'll stop the recording.", "speaker": "Fraser Paterson"}, {"start": 3475.566, "end": 3475.866, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3475.886, "end": 3476.887, "text": "I might stop the recording here.", "speaker": "Fraser Paterson"}, {"start": 3478.068, "end": 3478.549, "text": "Well, Mark.", "speaker": "Fraser Paterson"}, {"start": 3478.569, "end": 3479.71, "text": "Hello, everyone, and welcome.", "speaker": "Fraser Paterson"}, {"start": 3479.91, "end": 3481.732, "text": "I've got my esteemed friend.", "speaker": "Fraser Paterson"}, {"start": 3482.753, "end": 3484.975, "text": "Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 3486.423, "end": 3489.528, "text": " That was interesting audio feedback.", "speaker": "Marc Broberg"}, {"start": 3489.548, "end": 3492.053, "text": "I thought I would jump in since nobody else asked us.", "speaker": "Marc Broberg"}, {"start": 3493.135, "end": 3499.105, "text": "But earlier you mentioned using gradient descent to learn parameters, right?", "speaker": "Marc Broberg"}, {"start": 3499.245, "end": 3501.289, "text": "Is that also used in perception?", "speaker": "Marc Broberg"}, {"start": 3501.69, "end": 3503.893, "text": "Is it the same approach?", "speaker": "Marc Broberg"}, {"start": 3503.994, "end": 3508.581, "text": "Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3508.602, "end": 3510.625, "text": "So it's a very, very general", "speaker": "Fraser Paterson"}, {"start": 3511.06, "end": 3512.522, "text": " optimization techniques.", "speaker": "Fraser Paterson"}, {"start": 3512.602, "end": 3520.19, "text": "So yes, it certainly can be used for learning, well, for perception.", "speaker": "Fraser Paterson"}, {"start": 3520.711, "end": 3522.513, "text": "I'm trying to find a nice picture here.", "speaker": "Fraser Paterson"}, {"start": 3523.173, "end": 3531.983, "text": "Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent.", "speaker": "Fraser Paterson"}, {"start": 3532.524, "end": 3533.445, "text": "Let me just make sure of that.", "speaker": "Fraser Paterson"}, {"start": 3533.565, "end": 3535.107, "text": "Yes, in other words, is the answer.", "speaker": "Fraser Paterson"}, {"start": 3535.547, "end": 3537.049, "text": "It's used for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3538.39, "end": 3539.852, "text": "Very, very general technique.", "speaker": "Fraser Paterson"}, {"start": 3540.237, "end": 3549.91, "text": " Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six.", "speaker": "Fraser Paterson"}, {"start": 3549.93, "end": 3550.39, "text": "Absolutely.", "speaker": "Fraser Paterson"}, {"start": 3550.51, "end": 3550.891, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3550.911, "end": 3552.593, "text": "We're gonna talk about phase planes and so on.", "speaker": "Fraser Paterson"}, {"start": 3553.394, "end": 3553.635, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3553.955, "end": 3556.258, "text": "There are problems with it, but yes, we're gonna see it going forward.", "speaker": "Fraser Paterson"}, {"start": 3556.278, "end": 3559.242, "text": "It's a very, it's a very foundational technique.", "speaker": "Fraser Paterson"}, {"start": 3559.382, "end": 3562.807, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3562.827, "end": 3563.027, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3563.047, "end": 3566.211, "text": "Any, any last minute questions for the YouTube recording before we, uh,", "speaker": "Fraser Paterson"}, {"start": 3569.093, "end": 3570.975, "text": " If not, I think I'm sorry.", "speaker": "Fraser Paterson"}, {"start": 3571.236, "end": 3572.377, "text": "Can I ask a quick question?", "speaker": "Dorsa"}, {"start": 3572.437, "end": 3574.2, "text": "Yeah, sure.", "speaker": "Fraser Paterson"}, {"start": 3574.74, "end": 3575.461, "text": "I do apologize.", "speaker": "Dorsa"}, {"start": 3575.702, "end": 3577.424, "text": "I have joined the group very late.", "speaker": "Dorsa"}, {"start": 3577.484, "end": 3579.767, "text": "So maybe this was addressed in like the past weeks.", "speaker": "Dorsa"}, {"start": 3579.827, "end": 3593.245, "text": "But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning?", "speaker": "Dorsa"}, {"start": 3593.478, "end": 3619.553, "text": " well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms", "speaker": "Fraser Paterson"}, {"start": 3619.955, "end": 3629.547, "text": " If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control.", "speaker": "Fraser Paterson"}, {"start": 3630.147, "end": 3643.784, "text": "There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference.", "speaker": "Fraser Paterson"}, {"start": 3643.804, "end": 3646.988, "text": "So the answer is yes, there is a profound relationship.", "speaker": "Fraser Paterson"}, {"start": 3647.008, "end": 3649.371, "text": "We probably don't have time to get into it here.", "speaker": "Fraser Paterson"}, {"start": 3649.84, "end": 3659.489, "text": " One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain.", "speaker": "Fraser Paterson"}, {"start": 3660.41, "end": 3665.735, "text": "Because we've seen with, well, we haven't yet seen with expected free energy how that works.", "speaker": "Fraser Paterson"}, {"start": 3665.755, "end": 3672.461, "text": "But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning.", "speaker": "Fraser Paterson"}, {"start": 3673.062, "end": 3679.628, "text": "And we don't just have this scale of reward, this signal in reinforcement learning.", "speaker": "Fraser Paterson"}, {"start": 3680.114, "end": 3708.67, "text": " we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um", "speaker": "Fraser Paterson"}, {"start": 3708.97, "end": 3717.107, "text": " a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3717.147, "end": 3727.688, "text": "Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning,", "speaker": "Andrew Pashea"}, {"start": 3727.668, "end": 3741.341, "text": " Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two.", "speaker": "Andrew Pashea"}, {"start": 3742.282, "end": 3752.331, "text": "But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven.", "speaker": "Andrew Pashea"}, {"start": 3753.192, "end": 3756.595, "text": "So there can be like a KL control agent or there can be a variety.", "speaker": "Andrew Pashea"}, {"start": 3756.575, "end": 3779.157, "text": " other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that", "speaker": "Andrew Pashea"}, {"start": 3779.137, "end": 3782.622, "text": " that things like curiosity exist.", "speaker": "Andrew Pashea"}, {"start": 3782.742, "end": 3787.429, "text": "Why does curiosity exist if we're actually always just driven towards a reward?", "speaker": "Andrew Pashea"}, {"start": 3787.509, "end": 3790.173, "text": "Once you know what the reward is, you should just go for that, right?", "speaker": "Andrew Pashea"}, {"start": 3790.213, "end": 3796.262, "text": "And you would have no reason to find some other strategy for going for it.", "speaker": "Andrew Pashea"}, {"start": 3796.463, "end": 3798.466, "text": "You would have no reason to go out of your way.", "speaker": "Andrew Pashea"}, {"start": 3798.606, "end": 3803.032, "text": "From what you already know, you just keep going for the reward as much as possible.", "speaker": "Andrew Pashea"}, {"start": 3803.393, "end": 3808.941, "text": "Perhaps you learn other things through happenstance along the way.", "speaker": "Andrew Pashea"}, {"start": 3809.258, "end": 3811.041, "text": " That's a big difference.", "speaker": "Fraser Paterson"}, {"start": 3811.722, "end": 3814.948, "text": "In reinforcement, you can just sort of start maximizing a reward signal.", "speaker": "Fraser Paterson"}, {"start": 3815.569, "end": 3822.601, "text": "But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards.", "speaker": "Fraser Paterson"}, {"start": 3822.621, "end": 3828.972, "text": "You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards.", "speaker": "Fraser Paterson"}, {"start": 3828.952, "end": 3856.731, "text": " yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite", "speaker": "Andrew Pashea"}, {"start": 3857.234, "end": 3868.402, "text": " how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3868.743, "end": 3872.492, "text": "Meanwhile, active inference has a much more principled way that relates to this notion of", "speaker": "Andrew Pashea"}, {"start": 3872.472, "end": 3892.793, "text": " free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually", "speaker": "Andrew Pashea"}, {"start": 3893.06, "end": 3896.992, "text": " Um, when I play a sport, I throw the ball this way.", "speaker": "Andrew Pashea"}, {"start": 3897.012, "end": 3906.681, "text": "Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person.", "speaker": "Andrew Pashea"}, {"start": 3906.897, "end": 3909.963, "text": " but maybe I will kind of curve it or change it, right?", "speaker": "Andrew Pashea"}, {"start": 3910.343, "end": 3914.03, "text": "You're not randomly throwing it in the sky or in the opposite direction.", "speaker": "Andrew Pashea"}, {"start": 3914.431, "end": 3919.481, "text": "You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right?", "speaker": "Andrew Pashea"}, {"start": 3919.541, "end": 3927.235, "text": "There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better.", "speaker": "Andrew Pashea"}, {"start": 3927.552, "end": 3953.037, "text": " based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about", "speaker": "Andrew Pashea"}, {"start": 3953.523, "end": 3956.968, "text": " How do I perform the best whenever it comes to active inference?", "speaker": "Andrew Pashea"}, {"start": 3957.129, "end": 3965.422, "text": "It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them.", "speaker": "Andrew Pashea"}, {"start": 3966.223, "end": 3975.918, "text": "But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear.", "speaker": "Andrew Pashea"}, {"start": 3976.038, "end": 3977.32, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3978.194, "end": 3983.299, "text": " I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys.", "speaker": "Fraser Paterson"}, {"start": 3983.319, "end": 3986.502, "text": "So Giancuomo, fire away.", "speaker": "Fraser Paterson"}, {"start": 3986.522, "end": 3991.827, "text": "Very quickly, I think it ties in with what was just spoken and talked about now.", "speaker": "SPEAKER_00"}, {"start": 3992.648, "end": 4007.982, "text": "Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is?", "speaker": "SPEAKER_00"}, {"start": 4008.265, "end": 4033.468, "text": " Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty.", "speaker": "SPEAKER_00"}, {"start": 4033.508, "end": 4035.029, "text": "They say, this is the answer.", "speaker": "SPEAKER_00"}, {"start": 4035.109, "end": 4037.091, "text": "And you're like, no, no, no, it's not.", "speaker": "SPEAKER_00"}, {"start": 4037.594, "end": 4057.919, "text": " And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware.", "speaker": "SPEAKER_00"}, {"start": 4057.959, "end": 4064.007, "text": "Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind?", "speaker": "SPEAKER_00"}, {"start": 4065.455, "end": 4093.251, "text": " well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions", "speaker": "Fraser Paterson"}, {"start": 4093.687, "end": 4094.989, "text": " This is built in from the start.", "speaker": "Fraser Paterson"}, {"start": 4095.029, "end": 4103.639, "text": "So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right?", "speaker": "Fraser Paterson"}, {"start": 4104.22, "end": 4105.982, "text": "And okay, we can't do Bayes' rule exactly.", "speaker": "Fraser Paterson"}, {"start": 4106.022, "end": 4107.023, "text": "We have to do it approximately.", "speaker": "Fraser Paterson"}, {"start": 4107.083, "end": 4108.365, "text": "We do variational inference.", "speaker": "Fraser Paterson"}, {"start": 4108.405, "end": 4110.407, "text": "We do prediction error and minimization, that kind of thing.", "speaker": "Fraser Paterson"}, {"start": 4110.828, "end": 4114.052, "text": "But we're always, always, always reasoning about probability distribution.", "speaker": "Fraser Paterson"}, {"start": 4114.072, "end": 4118.597, "text": "So yes, every single active implementation ever is always", "speaker": "Fraser Paterson"}, {"start": 4120.535, "end": 4147.757, "text": " predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet", "speaker": "Fraser Paterson"}, {"start": 4147.737, "end": 4149.84, "text": " And there's issues around that and how to do that.", "speaker": "Fraser Paterson"}, {"start": 4150.821, "end": 4151.722, "text": "But we haven't seen that just yet.", "speaker": "Fraser Paterson"}, {"start": 4151.742, "end": 4153.024, "text": "So, yes, absolutely.", "speaker": "Fraser Paterson"}, {"start": 4153.084, "end": 4155.808, "text": "It's part and parcel of what it is to be an active infatuation.", "speaker": "Fraser Paterson"}, {"start": 4155.828, "end": 4157.15, "text": "It's the reason about uncertainty now.", "speaker": "Fraser Paterson"}, {"start": 4158.411, "end": 4158.932, "text": "Okay, thank you.", "speaker": "Fraser Paterson"}, {"start": 4159.613, "end": 4159.913, "text": "No problem.", "speaker": "Fraser Paterson"}, {"start": 4159.933, "end": 4162.036, "text": "And then, Mark, and then I think we can do... Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 4162.337, "end": 4167.163, "text": "Maybe just really quickly, did you have a comment, Andrew?", "speaker": "Fraser Paterson"}, {"start": 4167.413, "end": 4171.819, "text": " I also really have to go, but yeah, just briefly, you basically answered it.", "speaker": "Andrew Pashea"}, {"start": 4171.839, "end": 4173.622, "text": "Sorry, I was looking at the chat.", "speaker": "Andrew Pashea"}, {"start": 4173.642, "end": 4177.928, "text": "It's just, yeah, it depends on how the question is being asked.", "speaker": "Andrew Pashea"}, {"start": 4177.948, "end": 4185.339, "text": "Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are?", "speaker": "Andrew Pashea"}, {"start": 4185.459, "end": 4191.348, "text": "Like that's going to take a little bit more than just like only looking at the simplified models.", "speaker": "Andrew Pashea"}, {"start": 4191.368, "end": 4193.05, "text": "We're looking at the textbook here.", "speaker": "Andrew Pashea"}, {"start": 4193.03, "end": 4216.928, "text": " um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization", "speaker": "Andrew Pashea"}, {"start": 4216.908, "end": 4230.176, "text": " of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots.", "speaker": "Andrew Pashea"}, {"start": 4230.256, "end": 4232.18, "text": "There are 2 possible things.", "speaker": "Andrew Pashea"}, {"start": 4232.22, "end": 4235.507, "text": "It could be is completely 50 50 is fully uncertain.", "speaker": "Andrew Pashea"}, {"start": 4235.487, "end": 4246.867, "text": " Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision.", "speaker": "Andrew Pashea"}, {"start": 4247.388, "end": 4247.609, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4247.709, "end": 4255.423, "text": "Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob.", "speaker": "Andrew Pashea"}, {"start": 4255.803, "end": 4256.044, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4256.104, "end": 4258.508, "text": "The more you turn up precision, the more you're.", "speaker": "Andrew Pashea"}, {"start": 4258.488, "end": 4263.134, "text": " you're going to take into account the prediction errors you're receiving and vice versa.", "speaker": "Andrew Pashea"}, {"start": 4263.214, "end": 4275.609, "text": "So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving.", "speaker": "Andrew Pashea"}, {"start": 4275.669, "end": 4275.89, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4275.91, "end": 4279.294, "text": "So so uncertainty is very much like throughout.", "speaker": "Andrew Pashea"}, {"start": 4279.474, "end": 4284.18, "text": "I mean, it's a very ubiquitous term through many parts of active inference.", "speaker": "Andrew Pashea"}, {"start": 4284.2, "end": 4284.34, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 4284.48, "end": 4285.742, "text": "So it's essential.", "speaker": "Andrew Pashea"}, {"start": 4287.224, "end": 4288.225, "text": "So all right.", "speaker": "Jim DeLong"}, {"start": 4288.677, "end": 4311.58, "text": " so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see", "speaker": "Jim DeLong"}, {"start": 4312.673, "end": 4337.179, "text": " right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh", "speaker": "Andrew Pashea"}, {"start": 4337.547, "end": 4338.789, "text": " who cares about priors.", "speaker": "Andrew Pashea"}, {"start": 4338.829, "end": 4342.634, "text": "And that's what we see with maximum likelihood estimation in chapter two there.", "speaker": "Andrew Pashea"}, {"start": 4343.655, "end": 4354.31, "text": "But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something.", "speaker": "Andrew Pashea"}, {"start": 4354.35, "end": 4366.086, "text": "It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like,", "speaker": "Andrew Pashea"}, {"start": 4366.403, "end": 4392.141, "text": " thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um", "speaker": "Andrew Pashea"}, {"start": 4392.661, "end": 4394.263, "text": " you know, how those connect up with each other.", "speaker": "Andrew Pashea"}, {"start": 4394.704, "end": 4403.396, "text": "So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there.", "speaker": "Andrew Pashea"}, {"start": 4403.476, "end": 4413.409, "text": "And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe.", "speaker": "Andrew Pashea"}, {"start": 4413.55, "end": 4416.193, "text": "And so I fully trust it without question.", "speaker": "Andrew Pashea"}, {"start": 4416.654, "end": 4416.854, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4416.874, "end": 4419.057, "text": "So I say all these things because, uh,", "speaker": "Andrew Pashea"}, {"start": 4419.037, "end": 4426.085, "text": " Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe.", "speaker": "Andrew Pashea"}, {"start": 4426.666, "end": 4442.303, "text": "And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things.", "speaker": "Andrew Pashea"}, {"start": 4442.343, "end": 4444.065, "text": "Yeah, it's very interesting to think about.", "speaker": "Andrew Pashea"}, {"start": 4445.294, "end": 4449.623, "text": " Yeah, do take a look at, I think I'm sharing, but figure 2.11.", "speaker": "Fraser Paterson"}, {"start": 4450.525, "end": 4459.965, "text": "This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief.", "speaker": "Fraser Paterson"}, {"start": 4459.985, "end": 4462.29, "text": "So if the precision is really tight around the prior,", "speaker": "Fraser Paterson"}, {"start": 4462.692, "end": 4472.808, "text": " the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model.", "speaker": "Fraser Paterson"}, {"start": 4472.828, "end": 4474.511, "text": "So yeah, do take a look at that.", "speaker": "Fraser Paterson"}, {"start": 4474.531, "end": 4482.584, "text": "That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have.", "speaker": "Fraser Paterson"}, {"start": 4482.684, "end": 4483.786, "text": "Yeah.", "speaker": "Jim DeLong"}, {"start": 4483.806, "end": 4484.146, "text": "Excellent.", "speaker": "Jim DeLong"}, {"start": 4484.206, "end": 4484.667, "text": "Thank you.", "speaker": "Jim DeLong"}, {"start": 4485.372, "end": 4485.512, "text": " Cool.", "speaker": "Fraser Paterson"}, {"start": 4485.532, "end": 4487.255, "text": "All right, guys, we're going to have to call it there.", "speaker": "Fraser Paterson"}, {"start": 4487.756, "end": 4491.963, "text": "Do put questions in the chats on the page, the code page.", "speaker": "Fraser Paterson"}, {"start": 4492.003, "end": 4493.606, "text": "I'll be a bit more attentive going forward.", "speaker": "Fraser Paterson"}, {"start": 4494.267, "end": 4495.489, "text": "I look forward to next week.", "speaker": "Fraser Paterson"}, {"start": 4495.509, "end": 4499.796, "text": "I'll be doing another session of this kind on Friday for the people who usually attend that session.", "speaker": "Fraser Paterson"}, {"start": 4499.816, "end": 4500.557, "text": "So thank you very much.", "speaker": "Fraser Paterson"}, {"start": 4500.918, "end": 4502.22, "text": "Stop sharing and stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4503.522, "end": 4504.704, "text": "Have we got a recording here?", "speaker": "Fraser Paterson"}, {"start": 4506.247, "end": 4507.269, "text": "Okay, stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4507.749, "end": 4508.551, "text": "Goodbye, YouTube people.", "speaker": "Fraser Paterson"}, {"start": 4508.751, "end": 4509.372, "text": "Until next time.", "speaker": "Fraser Paterson"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt index e44125349..9d0693e97 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt @@ -1,4 +1,4 @@ -SPEAKER_05: +Fraser Paterson: all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two. @@ -54,11 +54,11 @@ Enough of that. I'll start sharing my screen here. -SPEAKER_04: +Fraser Paterson: Probably my entire screen. -SPEAKER_05: +Fraser Paterson: Okie dokie. So people should be able to see. @@ -488,7 +488,7 @@ And I see Mark has a question. Please fire away, Mark. -SPEAKER_02: +Marc Broberg: As usual. Thank you so much for this review. @@ -500,7 +500,7 @@ I was wondering if you could pull up that slide that showed the generative proce That might be helpful as a reference. -SPEAKER_05: +Fraser Paterson: Oh, yeah, okay. So let me come back here. @@ -508,15 +508,15 @@ So let me come back here. That's the one in chapter two, I assume you're referring to? -SPEAKER_02: +Marc Broberg: I think so. -SPEAKER_05: +Fraser Paterson: This one here? -SPEAKER_02: +Marc Broberg: Yes, thank you. Yeah, I really just kind of appreciate what this textbook is trying to do. @@ -528,11 +528,11 @@ But nonetheless, so we've got this model of the we've got this in the generative in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise -SPEAKER_05: +Fraser Paterson: in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of -SPEAKER_02: +Marc Broberg: It seems like chapter four kind of serves as a background where the story gets started for the most part. And this is kind of more of a ground level @@ -552,7 +552,7 @@ then becomes part of, say, a likelihood. And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense. -SPEAKER_05: +Fraser Paterson: Yeah, yeah. No, absolutely. @@ -636,7 +636,7 @@ So does that answer your question, Mark, about why why probability distributions, maybe why normal distributions, or did that help at all? -SPEAKER_02: +Marc Broberg: Yeah, that helped a lot. Thank you. @@ -646,7 +646,7 @@ Yeah, it's just recognizing, okay, there's a linear process, and then we have be And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right? -SPEAKER_05: +Fraser Paterson: uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all We could plug, we could say, okay, the relationship is just equal to the hidden state. @@ -666,17 +666,17 @@ And it's up to us as modelers to plug in what we think the relationship is betwe Yeah. -SPEAKER_04: +Fraser Paterson: very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B, -SPEAKER_05: +Fraser Paterson: Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from. Michael Morehead , Other questions other questions Magdalena. -SPEAKER_01: +Magdalena Hurtado: Okay, my question is this. So could we do a thought experiment for a minute? @@ -722,7 +722,7 @@ Okay, as a third level of information processing. And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis? -SPEAKER_05: +Fraser Paterson: Yeah, no, excellent, excellent question, Magdalena. This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents. @@ -792,7 +792,7 @@ Very, very important question. But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding. -SPEAKER_01: +Magdalena Hurtado: OK. really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups @@ -814,7 +814,7 @@ so take a generative model you present it to to in in some context social contex And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math. -SPEAKER_09: +Andrew Pashea: If I just kind of sorry, is that okay? All right. @@ -948,7 +948,7 @@ It's not about like following proper logic and stuff. It's people learn to do that, right? -SPEAKER_01: +Magdalena Hurtado: Can I interject something real quick? I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're @@ -964,7 +964,7 @@ It's not about evaluating the validity of the premises. It's not about the... That gets at a very... -SPEAKER_05: +Fraser Paterson: That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer. We might imagine an active inference agent doing inference about this. @@ -984,7 +984,7 @@ But that's not going to be the focus of the book. A lot of that is at the cutting edge of active inference research now. -SPEAKER_09: +Andrew Pashea: Thanks for keeping it closer to the textbook by bringing up precision. I very much agree with that. @@ -1024,7 +1024,7 @@ then cool but i just i'm just trying to make the point if you want to talk about ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it -SPEAKER_05: +Fraser Paterson: Do make sure, I've been absent for a little bit. I'm gonna be much more active on the questions side of things on the CODA. @@ -1050,7 +1050,7 @@ As I say, I'll go through and attempt to answer all the previous ones as well. think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry -SPEAKER_08: +Sun Xin: okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian inference to engage in parameter learning so that it can update beliefs about the current state. @@ -1064,7 +1064,7 @@ The ultimate goal of all this is to minimize expected free energy. Is that right? -SPEAKER_05: +Fraser Paterson: Wow, okay. I mean, basically, yeah, that's substantively correct. @@ -1120,11 +1120,11 @@ Because we generally can't do expectation maximization for the reason that we do So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah. -SPEAKER_08: +Sun Xin: Thank you. -SPEAKER_05: +Fraser Paterson: No problem. Do we have more questions? @@ -1148,7 +1148,7 @@ I've got my esteemed friend. Oh, hang on. -SPEAKER_02: +Marc Broberg: That was interesting audio feedback. I thought I would jump in since nobody else asked us. @@ -1160,7 +1160,7 @@ Is that also used in perception? Is it the same approach? -SPEAKER_05: +Fraser Paterson: Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things. So it's a very, very general @@ -1204,15 +1204,15 @@ Any, any last minute questions for the YouTube recording before we, uh, If not, I think I'm sorry. -SPEAKER_07: +Dorsa: Can I ask a quick question? -SPEAKER_05: +Fraser Paterson: Yeah, sure. -SPEAKER_07: +Dorsa: I do apologize. I have joined the group very late. @@ -1222,7 +1222,7 @@ So maybe this was addressed in like the past weeks. But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning? -SPEAKER_05: +Fraser Paterson: well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control. @@ -1244,7 +1244,7 @@ And we don't just have this scale of reward, this signal in reinforcement learni we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um -SPEAKER_09: +Andrew Pashea: a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning. Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning, @@ -1272,7 +1272,7 @@ From what you already know, you just keep going for the reward as much as possib Perhaps you learn other things through happenstance along the way. -SPEAKER_05: +Fraser Paterson: That's a big difference. In reinforcement, you can just sort of start maximizing a reward signal. @@ -1282,7 +1282,7 @@ But in active inference and in the, dare I say, in the real world, oftentimes yo You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards. -SPEAKER_09: +Andrew Pashea: yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning. @@ -1312,7 +1312,7 @@ It's not just about like, otherwise we could just make another deep neural netwo But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear. -SPEAKER_05: +Fraser Paterson: Yeah. I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys. @@ -1336,7 +1336,7 @@ And is there a sense in which it is a bit more nuanced, that it takes care that Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind? -SPEAKER_05: +Fraser Paterson: well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions This is built in from the start. @@ -1376,7 +1376,7 @@ And then, Mark, and then I think we can do... Oh, hang on. Maybe just really quickly, did you have a comment, Andrew? -SPEAKER_09: +Andrew Pashea: I also really have to go, but yeah, just briefly, you basically answered it. Sorry, I was looking at the chat. @@ -1422,13 +1422,13 @@ Yeah. So it's essential. -SPEAKER_03: +Jim DeLong: So all right. so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see -SPEAKER_09: +Andrew Pashea: right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh who cares about priors. @@ -1460,7 +1460,7 @@ And then furthermore, how does this actually show in like psychiatry and psychol Yeah, it's very interesting to think about. -SPEAKER_05: +Fraser Paterson: Yeah, do take a look at, I think I'm sharing, but figure 2.11. This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief. @@ -1474,7 +1474,7 @@ So yeah, do take a look at that. That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have. -SPEAKER_03: +Jim DeLong: Yeah. Excellent. @@ -1482,7 +1482,7 @@ Excellent. Thank you. -SPEAKER_05: +Fraser Paterson: Cool. All right, guys, we're going to have to call it there. From 01611b6b07c60d5aec4fcc3e9dc3ba981ff35900 Mon Sep 17 00:00:00 2001 From: Holly Grimm Date: Fri, 17 Jul 2026 16:42:39 -0600 Subject: [PATCH 3/4] =?UTF-8?q?data:=20Session=5F023=20SPEAKER=5F00=20=3D?= =?UTF-8?q?=20Giacomo=20Bruzzo=20=E2=80=94=20all=20speakers=20identified?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_015NjFiMWkgrVy4foGF1Zbax --- .../Namjoshi2026/Cohort_1/Session_023/metadata.json | 3 ++- .../Namjoshi2026/Cohort_1/Session_023/transcript.json | 2 +- .../Namjoshi2026/Cohort_1/Session_023/transcript.txt | 2 +- 3 files changed, 4 insertions(+), 3 deletions(-) diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json index 1d4f7a5dd..29e232520 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/metadata.json @@ -24,7 +24,8 @@ "SPEAKER_03": "Jim DeLong", "SPEAKER_04": "Fraser Paterson", "SPEAKER_07": "Dorsa", - "SPEAKER_08": "Sun Xin" + "SPEAKER_08": "Sun Xin", + "SPEAKER_00": "Giacomo Bruzzo" } } ] diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json index 1f0b4ce89..54ce9f0f8 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.json @@ -1 +1 @@ -[{"video_id": "MjeQeWeyYhE", "segments": [{"start": 3.727, "end": 29.498, "text": " all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense", "speaker": "Fraser Paterson"}, {"start": 29.798, "end": 40.317, "text": " And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two.", "speaker": "Fraser Paterson"}, {"start": 40.337, "end": 43.923, "text": "We're going to do active inputs proper, which is going to be very, very exciting going forward.", "speaker": "Fraser Paterson"}, {"start": 43.943, "end": 50.575, "text": "So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts.", "speaker": "Fraser Paterson"}, {"start": 51.027, "end": 56.833, "text": " that we're going to start doing actual, quote unquote, active inference in the second part.", "speaker": "Fraser Paterson"}, {"start": 56.853, "end": 67.444, "text": "But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection.", "speaker": "Fraser Paterson"}, {"start": 68.305, "end": 72.409, "text": "We're going to sort of slow down and think about part one in general, chapters one to five.", "speaker": "Fraser Paterson"}, {"start": 73.531, "end": 80.578, "text": "This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one.", "speaker": "Fraser Paterson"}, {"start": 80.812, "end": 85.079, "text": " you know, going so that we can be adequately grounded going forward.", "speaker": "Fraser Paterson"}, {"start": 85.981, "end": 94.696, "text": "I will say we have one one question is, you know, we've done each chapter across two weeks.", "speaker": "Fraser Paterson"}, {"start": 95.658, "end": 100.947, "text": "Historically, you know, chapter one, we've had two weeks going forward.", "speaker": "Fraser Paterson"}, {"start": 101.72, "end": 128.115, "text": " depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week", "speaker": "Fraser Paterson"}, {"start": 128.095, "end": 152.685, "text": " of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe", "speaker": "Fraser Paterson"}, {"start": 152.935, "end": 159.784, "text": " It would be a good idea to reach out through the blankets email.", "speaker": "Fraser Paterson"}, {"start": 160.425, "end": 166.993, "text": "Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place.", "speaker": "Fraser Paterson"}, {"start": 168.415, "end": 169.296, "text": "That would be amazing.", "speaker": "Fraser Paterson"}, {"start": 169.416, "end": 170.057, "text": "Good idea.", "speaker": "Fraser Paterson"}, {"start": 170.698, "end": 171.198, "text": "Yes, please.", "speaker": "Fraser Paterson"}, {"start": 171.238, "end": 176.445, "text": "I see a lot of people are saying yes.", "speaker": "Fraser Paterson"}, {"start": 176.695, "end": 182.461, "text": " In this chat here, it would be very helpful if you could give your yay or nay to that as well.", "speaker": "Fraser Paterson"}, {"start": 182.581, "end": 185.384, "text": "So we can actually look at that just directly through the chat here.", "speaker": "Fraser Paterson"}, {"start": 185.464, "end": 188.848, "text": "So if you want a second week review, yes.", "speaker": "Fraser Paterson"}, {"start": 189.108, "end": 190.089, "text": "If you don't, no.", "speaker": "Fraser Paterson"}, {"start": 190.71, "end": 192.252, "text": "And then we can go forward.", "speaker": "Fraser Paterson"}, {"start": 193.553, "end": 193.893, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 195.495, "end": 196.096, "text": "Enough of that.", "speaker": "Fraser Paterson"}, {"start": 196.616, "end": 197.898, "text": "I'll start sharing my screen here.", "speaker": "Fraser Paterson"}, {"start": 199.259, "end": 200.28, "text": "Probably my entire screen.", "speaker": "Fraser Paterson"}, {"start": 204.084, "end": 204.505, "text": "Okie dokie.", "speaker": "Fraser Paterson"}, {"start": 204.525, "end": 205.446, "text": "So people should be able to see.", "speaker": "Fraser Paterson"}, {"start": 205.486, "end": 206.567, "text": "Maybe just get rid of...", "speaker": "Fraser Paterson"}, {"start": 207.357, "end": 214.009, "text": " The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise.", "speaker": "Fraser Paterson"}, {"start": 215.011, "end": 227.172, "text": "about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here.", "speaker": "Fraser Paterson"}, {"start": 227.928, "end": 229.152, "text": " So you've got your overview.", "speaker": "Fraser Paterson"}, {"start": 229.935, "end": 232.945, "text": "I've added one or two additional resources.", "speaker": "Fraser Paterson"}, {"start": 232.965, "end": 233.827, "text": "So these actually came up.", "speaker": "Fraser Paterson"}, {"start": 233.908, "end": 235.593, "text": "Andrew shared the first of these.", "speaker": "Fraser Paterson"}, {"start": 235.613, "end": 236.436, "text": "These are all videos.", "speaker": "Fraser Paterson"}, {"start": 237.239, "end": 238.242, "text": "One of them is an active inference", "speaker": "Fraser Paterson"}, {"start": 240.668, "end": 241.791, "text": " coding and active inference.", "speaker": "Fraser Paterson"}, {"start": 243.094, "end": 245.139, "text": "So that's Ryan Smith and his team.", "speaker": "Fraser Paterson"}, {"start": 245.841, "end": 248.227, "text": "They've done excellent work on predictive coding.", "speaker": "Fraser Paterson"}, {"start": 248.708, "end": 249.45, "text": "I'll play that just now.", "speaker": "Fraser Paterson"}, {"start": 249.991, "end": 258.171, "text": "This is a great stream just about predictive coding more generally and its relation to active inference and its relation to", "speaker": "Fraser Paterson"}, {"start": 258.151, "end": 260.855, "text": " you know, its status in actual brains.", "speaker": "Fraser Paterson"}, {"start": 261.616, "end": 265.781, "text": "So if you're interested in more context, this would be excellent to go through.", "speaker": "Fraser Paterson"}, {"start": 265.801, "end": 267.003, "text": "And there's two more links there as well.", "speaker": "Fraser Paterson"}, {"start": 267.043, "end": 271.99, "text": "One by Jakob Howie, he's a philosopher over in Australia in the University of Monash.", "speaker": "Fraser Paterson"}, {"start": 272.01, "end": 274.013, "text": "What is predictive processing and what is it good for?", "speaker": "Fraser Paterson"}, {"start": 274.133, "end": 274.914, "text": "Another excellent talk.", "speaker": "Fraser Paterson"}, {"start": 275.495, "end": 280.041, "text": "And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke.", "speaker": "Fraser Paterson"}, {"start": 280.325, "end": 281.426, "text": " from the University of Toronto.", "speaker": "Fraser Paterson"}, {"start": 281.446, "end": 286.772, "text": "He's actually speaking about Neoplatonism and mystical experiences, believe it or not.", "speaker": "Fraser Paterson"}, {"start": 287.513, "end": 289.135, "text": "And I know it sounds weird.", "speaker": "Fraser Paterson"}, {"start": 289.575, "end": 292.799, "text": "Predictive processing actually shows up in here in a very significant way.", "speaker": "Fraser Paterson"}, {"start": 292.819, "end": 297.084, "text": "So that might be a nice mood setter, as it were.", "speaker": "Fraser Paterson"}, {"start": 297.164, "end": 298.365, "text": "So there's some additional resources there.", "speaker": "Fraser Paterson"}, {"start": 298.686, "end": 301.128, "text": "I've begun filling in the map for chapter five.", "speaker": "Fraser Paterson"}, {"start": 301.949, "end": 307.075, "text": "So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid.", "speaker": "Fraser Paterson"}, {"start": 307.73, "end": 334.446, "text": " the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that", "speaker": "Fraser Paterson"}, {"start": 334.814, "end": 338.84, "text": " A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation.", "speaker": "Fraser Paterson"}, {"start": 339.401, "end": 353.881, "text": "What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you.", "speaker": "Fraser Paterson"}, {"start": 354.322, "end": 359.269, "text": "You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you.", "speaker": "Fraser Paterson"}, {"start": 359.249, "end": 363.355, "text": " So it's not ideal, but the content is there at least.", "speaker": "Fraser Paterson"}, {"start": 363.736, "end": 367.081, "text": "And especially for those who don't have the book, I think this is quite useful.", "speaker": "Fraser Paterson"}, {"start": 367.101, "end": 374.753, "text": "So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter.", "speaker": "Fraser Paterson"}, {"start": 375.274, "end": 382.405, "text": "I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts", "speaker": "Fraser Paterson"}, {"start": 382.975, "end": 403.431, "text": " uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there", "speaker": "Fraser Paterson"}, {"start": 404.373, "end": 404.954, "text": " Any questions?", "speaker": "Fraser Paterson"}, {"start": 404.994, "end": 416.831, "text": "Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that.", "speaker": "Fraser Paterson"}, {"start": 418.033, "end": 419.795, "text": "I'll start my share temporarily.", "speaker": "Fraser Paterson"}, {"start": 420.756, "end": 421.858, "text": "Just coming back.", "speaker": "Fraser Paterson"}, {"start": 423.863, "end": 424.404, "text": " Very good.", "speaker": "Fraser Paterson"}, {"start": 425.745, "end": 426.386, "text": "Yes, but it's limited.", "speaker": "Fraser Paterson"}, {"start": 428.008, "end": 430.991, "text": "Formatting in LaTeX is a bit strange in Coda.", "speaker": "Fraser Paterson"}, {"start": 432.032, "end": 442.664, "text": "We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda.", "speaker": "Fraser Paterson"}, {"start": 444.026, "end": 445.928, "text": "OK, so I'll share, come back.", "speaker": "Fraser Paterson"}, {"start": 447.89, "end": 452.976, "text": "Let's go all the way back, hopefully people can see my screen again, to the introduction.", "speaker": "Fraser Paterson"}, {"start": 454.441, "end": 456.885, "text": " So everyone's seen this figure here.", "speaker": "Fraser Paterson"}, {"start": 456.905, "end": 458.408, "text": "This is essentially the first figure in the book.", "speaker": "Fraser Paterson"}, {"start": 458.849, "end": 461.814, "text": "This is the breakdown of the partitions of the book, so part one, part two, part three.", "speaker": "Fraser Paterson"}, {"start": 462.396, "end": 468.046, "text": "We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding.", "speaker": "Fraser Paterson"}, {"start": 469.649, "end": 471.232, "text": "Part two is active inference core.", "speaker": "Fraser Paterson"}, {"start": 471.252, "end": 473.215, "text": "So this is the heart of the book, really.", "speaker": "Fraser Paterson"}, {"start": 473.255, "end": 476.922, "text": "This is the fundamentals of active inference per se.", "speaker": "Fraser Paterson"}, {"start": 476.902, "end": 498.64, "text": " what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end", "speaker": "Fraser Paterson"}, {"start": 499.328, "end": 504.875, "text": " But as has been the case, everything has been just perception only.", "speaker": "Fraser Paterson"}, {"start": 504.895, "end": 506.057, "text": "So we haven't actually dealt with actions.", "speaker": "Fraser Paterson"}, {"start": 506.437, "end": 508.24, "text": "And we're going to be moving into that with part two.", "speaker": "Fraser Paterson"}, {"start": 509.221, "end": 511.564, "text": "But even more fundamentally, everything has been static.", "speaker": "Fraser Paterson"}, {"start": 511.824, "end": 516.651, "text": "So all of the models we've been looking at, I don't know if I can get a nice example.", "speaker": "Fraser Paterson"}, {"start": 517.592, "end": 519.014, "text": "All of the models, so this is chapter one.", "speaker": "Fraser Paterson"}, {"start": 519.995, "end": 523.159, "text": "Let's maybe go down to some of the equations.", "speaker": "Fraser Paterson"}, {"start": 523.443, "end": 524.244, "text": " They've all been static.", "speaker": "Fraser Paterson"}, {"start": 524.264, "end": 527.387, "text": "So the environment hasn't really been changing at all.", "speaker": "Fraser Paterson"}, {"start": 527.407, "end": 530.371, "text": "And this is very simple.", "speaker": "Fraser Paterson"}, {"start": 531.372, "end": 534.235, "text": "So part one, really going to chapter two, I think.", "speaker": "Fraser Paterson"}, {"start": 536.858, "end": 538.88, "text": "So we have an environment.", "speaker": "Fraser Paterson"}, {"start": 539.501, "end": 541.503, "text": "This is one we've seen this many times now.", "speaker": "Fraser Paterson"}, {"start": 541.523, "end": 548.691, "text": "This is a representation of the generative process, the real process in and of itself out there.", "speaker": "Fraser Paterson"}, {"start": 549.262, "end": 553.407, "text": " This is fairly static, so nothing is really changing.", "speaker": "Fraser Paterson"}, {"start": 553.467, "end": 555.609, "text": "The hidden states aren't really changing from moment to moment.", "speaker": "Fraser Paterson"}, {"start": 555.629, "end": 562.397, "text": "We're just presented with a situation and we have to update our beliefs about what hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 562.437, "end": 564.199, "text": "That's been constant throughout all of Chapter 1.", "speaker": "Fraser Paterson"}, {"start": 564.879, "end": 566.721, "text": "Things have gotten progressively more complicated.", "speaker": "Fraser Paterson"}, {"start": 566.741, "end": 570.906, "text": "We've looked at univariate hidden states where there's literally just one hidden state.", "speaker": "Fraser Paterson"}, {"start": 571.567, "end": 577.193, "text": "We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things.", "speaker": "Fraser Paterson"}, {"start": 577.662, "end": 579.304, "text": " But in every case, it's been static.", "speaker": "Fraser Paterson"}, {"start": 579.884, "end": 585.331, "text": "The dynamics of the hidden state have been non-existent.", "speaker": "Fraser Paterson"}, {"start": 585.351, "end": 586.992, "text": "That is going to change going into part two.", "speaker": "Fraser Paterson"}, {"start": 587.052, "end": 591.918, "text": "We're going to start looking at generative processes which change over time.", "speaker": "Fraser Paterson"}, {"start": 592.579, "end": 593.94, "text": "And they're the interesting ones.", "speaker": "Fraser Paterson"}, {"start": 593.98, "end": 596.283, "text": "They're the ones that are actually useful.", "speaker": "Fraser Paterson"}, {"start": 596.343, "end": 602.55, "text": "And they're the ones that allow us to actually begin to have a reason to act in the environment.", "speaker": "Fraser Paterson"}, {"start": 602.61, "end": 605.593, "text": "We haven't really had any reason to perform actions yet.", "speaker": "Fraser Paterson"}, {"start": 606.805, "end": 611.032, "text": " So that's a big move that's going to take place and that we're going to have to deal with.", "speaker": "Fraser Paterson"}, {"start": 612.074, "end": 613.517, "text": "So maybe just coming back.", "speaker": "Fraser Paterson"}, {"start": 613.557, "end": 618.325, "text": "So chapter one was about, what was this about?", "speaker": "Fraser Paterson"}, {"start": 618.345, "end": 619.187, "text": "It was about perception.", "speaker": "Fraser Paterson"}, {"start": 619.427, "end": 619.588, "text": "Okay.", "speaker": "Fraser Paterson"}, {"start": 619.628, "end": 624.176, "text": "So like in terms of how we're going to think about perception in active inference.", "speaker": "Fraser Paterson"}, {"start": 624.757, "end": 627.261, "text": "So in chapter one, we cast or we framed", "speaker": "Fraser Paterson"}, {"start": 630.565, "end": 654.058, "text": " of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at", "speaker": "Fraser Paterson"}, {"start": 655.118, "end": 661.187, "text": " you know, how to do like exact Bayesian inference, okay, if we were able to do everything.", "speaker": "Fraser Paterson"}, {"start": 663.11, "end": 665.213, "text": "We saw how to use Bayes' rule basically.", "speaker": "Fraser Paterson"}, {"start": 665.233, "end": 675.187, "text": "So we cemented the distinction personally before that around the difference between the generative process and the generative model.", "speaker": "Fraser Paterson"}, {"start": 675.989, "end": 679.354, "text": "So I come all the way down to the figures here.", "speaker": "Fraser Paterson"}, {"start": 679.374, "end": 681.076, "text": "I don't have a particularly good figure for that.", "speaker": "Fraser Paterson"}, {"start": 681.096, "end": 682.358, "text": "Here we go, 2.4.", "speaker": "Fraser Paterson"}, {"start": 682.793, "end": 691.766, "text": " The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world.", "speaker": "Fraser Paterson"}, {"start": 692.307, "end": 695.191, "text": "This was the crucial thing, I think, in chapter two, basically.", "speaker": "Fraser Paterson"}, {"start": 695.992, "end": 699.978, "text": "And we have a way of notating this convention.", "speaker": "Fraser Paterson"}, {"start": 699.998, "end": 708.531, "text": "We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 708.971, "end": 712.176, "text": "These belong to the generative process, the real world out there.", "speaker": "Fraser Paterson"}, {"start": 712.797, "end": 716.822, "text": " Okay, and then modeled variables are without a star.", "speaker": "Fraser Paterson"}, {"start": 717.142, "end": 719.184, "text": "So they correspond to our model.", "speaker": "Fraser Paterson"}, {"start": 720.746, "end": 721.467, "text": "And what is our model?", "speaker": "Fraser Paterson"}, {"start": 722.668, "end": 730.037, "text": "It is quite literally a joint probability distribution over hidden states, latent states, and observations.", "speaker": "Fraser Paterson"}, {"start": 730.978, "end": 733.421, "text": "We use x for latent states and y for observations.", "speaker": "Fraser Paterson"}, {"start": 734.722, "end": 742.551, "text": "But you can see, of course, the x in here is our model of the real latent state outside of here, x star.", "speaker": "Fraser Paterson"}, {"start": 743.172, "end": 744.755, "text": " And they need not agree with each other.", "speaker": "Fraser Paterson"}, {"start": 744.795, "end": 745.737, "text": "They need not be the same.", "speaker": "Fraser Paterson"}, {"start": 746.839, "end": 752.69, "text": "But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 752.731, "end": 754.795, "text": "That is p of x and y.", "speaker": "Fraser Paterson"}, {"start": 755.817, "end": 760.185, "text": "And then if we have the model evidence, the probability of observing anything,", "speaker": "Fraser Paterson"}, {"start": 760.722, "end": 769.453, "text": " or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation.", "speaker": "Fraser Paterson"}, {"start": 770.174, "end": 772.037, "text": "And that was kind of chapter two.", "speaker": "Fraser Paterson"}, {"start": 772.057, "end": 779.526, "text": "And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions.", "speaker": "Fraser Paterson"}, {"start": 779.546, "end": 780.648, "text": "And these are very ubiquitous.", "speaker": "Fraser Paterson"}, {"start": 780.668, "end": 781.369, "text": "They show up everywhere.", "speaker": "Fraser Paterson"}, {"start": 782.43, "end": 783.331, "text": "We're going to continue that.", "speaker": "Fraser Paterson"}, {"start": 783.351, "end": 790.24, "text": "In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference.", "speaker": "Fraser Paterson"}, {"start": 790.642, "end": 791.784, "text": " There are other distributions as well.", "speaker": "Fraser Paterson"}, {"start": 791.804, "end": 792.725, "text": "We haven't really looked at those yet.", "speaker": "Fraser Paterson"}, {"start": 793.967, "end": 799.554, "text": "But the real sort of, there's a problem with this, which is that, yes, we can do inference.", "speaker": "Fraser Paterson"}, {"start": 799.574, "end": 803.079, "text": "Yes, we might even be able to do exact inference for really, really simple problems.", "speaker": "Fraser Paterson"}, {"start": 804.14, "end": 812.952, "text": "But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model.", "speaker": "Fraser Paterson"}, {"start": 813.793, "end": 815.075, "text": "And then we're going to do inference.", "speaker": "Fraser Paterson"}, {"start": 815.095, "end": 820.142, "text": "And inference is like, I set the dials, and then my machine does some inference stuff, right?", "speaker": "Fraser Paterson"}, {"start": 820.983, "end": 847.716, "text": " that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation", "speaker": "Fraser Paterson"}, {"start": 848.303, "end": 850.966, "text": " And we're doing good old fashioned Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 851.346, "end": 854.99, "text": "But now we're having to do parameter learning and estimation as well.", "speaker": "Fraser Paterson"}, {"start": 856.692, "end": 867.903, "text": "And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference.", "speaker": "Fraser Paterson"}, {"start": 867.923, "end": 872.448, "text": "So we have these two processes, inference and learning parameter estimation.", "speaker": "Fraser Paterson"}, {"start": 873.109, "end": 877.213, "text": "And they're both able to be done by means of Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 878.56, "end": 879.421, "text": " That's the crucial thing.", "speaker": "Fraser Paterson"}, {"start": 879.441, "end": 889.438, "text": "So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah.", "speaker": "Fraser Paterson"}, {"start": 889.458, "end": 904.163, "text": "And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent.", "speaker": "Fraser Paterson"}, {"start": 904.868, "end": 908.953, "text": " And we saw many versions of this across the chapter and indeed in other chapters.", "speaker": "Fraser Paterson"}, {"start": 908.993, "end": 920.085, "text": "It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters.", "speaker": "Fraser Paterson"}, {"start": 920.185, "end": 926.172, "text": "So let's imagine we have our two parameters, I don't know, beta 1, beta 0.", "speaker": "Fraser Paterson"}, {"start": 926.232, "end": 934.521, "text": "And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space.", "speaker": "Fraser Paterson"}, {"start": 934.872, "end": 935.733, "text": " That's this surface.", "speaker": "Fraser Paterson"}, {"start": 936.754, "end": 944.063, "text": "And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point.", "speaker": "Fraser Paterson"}, {"start": 945.064, "end": 950.271, "text": "And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters.", "speaker": "Fraser Paterson"}, {"start": 950.291, "end": 952.654, "text": "Very general way of doing optimization.", "speaker": "Fraser Paterson"}, {"start": 954.015, "end": 964.288, "text": "However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be.", "speaker": "Fraser Paterson"}, {"start": 965.72, "end": 973.231, "text": " And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization.", "speaker": "Fraser Paterson"}, {"start": 975.574, "end": 985.308, "text": "So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now.", "speaker": "Fraser Paterson"}, {"start": 986.25, "end": 993.921, "text": "A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 994.357, "end": 1021.09, "text": " um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model", "speaker": "Fraser Paterson"}, {"start": 1021.525, "end": 1025.652, "text": " And also, how should I use those settings to do good inference?", "speaker": "Fraser Paterson"}, {"start": 1026.934, "end": 1040.558, "text": "It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters.", "speaker": "Fraser Paterson"}, {"start": 1040.578, "end": 1042.442, "text": "So imagine you've got learning up here.", "speaker": "Fraser Paterson"}, {"start": 1042.582, "end": 1043.744, "text": "What should the model parameters be?", "speaker": "Fraser Paterson"}, {"start": 1043.764, "end": 1046.048, "text": "And you can do inference on model parameters.", "speaker": "Fraser Paterson"}, {"start": 1046.45, "end": 1050.538, "text": " And then that can inform how you do inference at the low-level hidden states.", "speaker": "Fraser Paterson"}, {"start": 1050.558, "end": 1053.805, "text": "So there's kind of these two processes that are happening in two different timescales.", "speaker": "Fraser Paterson"}, {"start": 1053.825, "end": 1062.062, "text": "That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally.", "speaker": "Fraser Paterson"}, {"start": 1063.122, "end": 1064.023, "text": " So that's kind of chapter three.", "speaker": "Fraser Paterson"}, {"start": 1064.143, "end": 1065.264, "text": "I'm going to race through this.", "speaker": "Fraser Paterson"}, {"start": 1065.925, "end": 1068.828, "text": "And we'll get to chapter five and then to some actual questions from you guys.", "speaker": "Fraser Paterson"}, {"start": 1069.749, "end": 1081.062, "text": "Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference.", "speaker": "Fraser Paterson"}, {"start": 1081.082, "end": 1084.966, "text": "So chapter three, we're still doing exact inference.", "speaker": "Fraser Paterson"}, {"start": 1084.986, "end": 1090.632, "text": "But we saw that really we culminated in an algorithm", "speaker": "Fraser Paterson"}, {"start": 1091.658, "end": 1113.632, "text": " allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be", "speaker": "Fraser Paterson"}, {"start": 1113.95, "end": 1115.672, "text": " Those two problems are related to one another.", "speaker": "Fraser Paterson"}, {"start": 1115.752, "end": 1118.615, "text": "To know the hidden states, you need to know what the good parameters are.", "speaker": "Fraser Paterson"}, {"start": 1118.935, "end": 1121.378, "text": "But to know what the good parameters are, you have to know how to infer hidden states well.", "speaker": "Fraser Paterson"}, {"start": 1121.958, "end": 1131.588, "text": "So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm.", "speaker": "Fraser Paterson"}, {"start": 1131.608, "end": 1132.689, "text": "I won't go over that again here.", "speaker": "Fraser Paterson"}, {"start": 1132.73, "end": 1134.131, "text": "Maybe we will if people want me to.", "speaker": "Fraser Paterson"}, {"start": 1135.072, "end": 1139.957, "text": "But that was the solution to the problem of chapter 3, that chicken and egg problem.", "speaker": "Fraser Paterson"}, {"start": 1140.781, "end": 1151.908, "text": " The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations.", "speaker": "Fraser Paterson"}, {"start": 1151.928, "end": 1155.938, "text": "And that, unfortunately, is usually not something we can do.", "speaker": "Fraser Paterson"}, {"start": 1156.542, "end": 1161.508, "text": " Being able to find the exact posterior is usually impossible, for reasons that we saw.", "speaker": "Fraser Paterson"}, {"start": 1161.708, "end": 1162.75, "text": "We can review again if you want.", "speaker": "Fraser Paterson"}, {"start": 1163.33, "end": 1166.895, "text": "So then chapter 4 says, all right, well, we can't do exact Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1167.536, "end": 1168.096, "text": "What are we going to do?", "speaker": "Fraser Paterson"}, {"start": 1168.116, "end": 1177.107, "text": "We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like.", "speaker": "Fraser Paterson"}, {"start": 1178.029, "end": 1180.812, "text": "So the idea with chapter 4 is, ah, OK, we're going to now", "speaker": "Fraser Paterson"}, {"start": 1181.804, "end": 1193.779, "text": " have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 1194.7, "end": 1198.304, "text": "We're going to try and find a posterior that's close enough to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1198.705, "end": 1208.537, "text": "And this then motivated the idea of variational free energy as a quantity that is a measurement of how well", "speaker": "Fraser Paterson"}, {"start": 1208.838, "end": 1213.784, "text": " an approximate posterior fits or how close it is to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1214.885, "end": 1222.794, "text": "And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal.", "speaker": "Fraser Paterson"}, {"start": 1223.055, "end": 1225.858, "text": "We saw from chapter three that surprisal is something we want to make very small.", "speaker": "Fraser Paterson"}, {"start": 1226.719, "end": 1237.191, "text": "So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal,", "speaker": "Fraser Paterson"}, {"start": 1237.778, "end": 1242.586, "text": " which is something we can't do directly, and therefore we can do approximate Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1242.606, "end": 1248.976, "text": "So in a lot of ways, that really kind of spiritually is the end of part one, I would say.", "speaker": "Fraser Paterson"}, {"start": 1249.958, "end": 1265.102, "text": "The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE.", "speaker": "Fraser Paterson"}, {"start": 1265.842, "end": 1269.509, "text": " That's, in my mind, really kind of the end of part one.", "speaker": "Fraser Paterson"}, {"start": 1270.311, "end": 1278.908, "text": "Chapter five was a nice kind of detour into a different way of thinking about what the VFE is.", "speaker": "Fraser Paterson"}, {"start": 1280.451, "end": 1287.445, "text": "So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 1288.387, "end": 1314.799, "text": " we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical", "speaker": "Fraser Paterson"}, {"start": 1314.965, "end": 1334.813, "text": " know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was", "speaker": "Fraser Paterson"}, {"start": 1335.535, "end": 1349.242, "text": " lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning.", "speaker": "Fraser Paterson"}, {"start": 1350.083, "end": 1357.257, "text": "It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization", "speaker": "Fraser Paterson"}, {"start": 1357.574, "end": 1369.356, "text": " in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure", "speaker": "Fraser Paterson"}, {"start": 1370.163, "end": 1387.641, "text": " And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow.", "speaker": "Fraser Paterson"}, {"start": 1388.962, "end": 1391.865, "text": "And that's kind of where we left off with part one.", "speaker": "Fraser Paterson"}, {"start": 1391.885, "end": 1396.39, "text": "So that is, in effect, what we've done, where we've gone.", "speaker": "Fraser Paterson"}, {"start": 1396.943, "end": 1400.268, "text": " That's a lot of stuff to cover, but again, we haven't looked at actions yet.", "speaker": "Fraser Paterson"}, {"start": 1400.829, "end": 1407.12, "text": "We're gonna be doing that in part two, and that will bring us full circle to active inference.", "speaker": "Fraser Paterson"}, {"start": 1407.14, "end": 1420.021, "text": "And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception.", "speaker": "Fraser Paterson"}, {"start": 1420.922, "end": 1423.086, "text": "Okay, we're gonna see that these are deeply related to each other.", "speaker": "Fraser Paterson"}, {"start": 1423.623, "end": 1426.045, "text": " But that's the ground that we've covered.", "speaker": "Fraser Paterson"}, {"start": 1426.846, "end": 1429.829, "text": "I'd be interested if people have questions related to that.", "speaker": "Fraser Paterson"}, {"start": 1429.849, "end": 1431.29, "text": "I see that there's lots of questions.", "speaker": "Fraser Paterson"}, {"start": 1432.371, "end": 1432.912, "text": "Stop sharing.", "speaker": "Fraser Paterson"}, {"start": 1434.533, "end": 1437.476, "text": "And I see Mark has a question.", "speaker": "Fraser Paterson"}, {"start": 1437.496, "end": 1438.397, "text": "Please fire away, Mark.", "speaker": "Fraser Paterson"}, {"start": 1440.039, "end": 1440.519, "text": "As usual.", "speaker": "Marc Broberg"}, {"start": 1441.16, "end": 1443.041, "text": "Thank you so much for this review.", "speaker": "Marc Broberg"}, {"start": 1443.061, "end": 1443.602, "text": "This is great.", "speaker": "Marc Broberg"}, {"start": 1445.724, "end": 1451.389, "text": "I was wondering if you could pull up that slide that showed the generative process and the generative model.", "speaker": "Marc Broberg"}, {"start": 1451.429, "end": 1453.191, "text": "That might be helpful as a reference.", "speaker": "Marc Broberg"}, {"start": 1453.964, "end": 1455.166, "text": " Oh, yeah, okay.", "speaker": "Fraser Paterson"}, {"start": 1455.186, "end": 1456.187, "text": "So let me come back here.", "speaker": "Fraser Paterson"}, {"start": 1458.43, "end": 1462.415, "text": "That's the one in chapter two, I assume you're referring to?", "speaker": "Fraser Paterson"}, {"start": 1462.435, "end": 1462.836, "text": "I think so.", "speaker": "Marc Broberg"}, {"start": 1463.697, "end": 1464.138, "text": "This one here?", "speaker": "Fraser Paterson"}, {"start": 1464.618, "end": 1465.479, "text": "Yes, thank you.", "speaker": "Marc Broberg"}, {"start": 1466.1, "end": 1470.126, "text": "Yeah, I really just kind of appreciate what this textbook is trying to do.", "speaker": "Marc Broberg"}, {"start": 1470.146, "end": 1473.57, "text": "And I think it took me a while to appreciate that.", "speaker": "Marc Broberg"}, {"start": 1473.69, "end": 1482.402, "text": "But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening.", "speaker": "Marc Broberg"}, {"start": 1482.382, "end": 1505.689, "text": " in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise", "speaker": "Marc Broberg"}, {"start": 1505.905, "end": 1532.082, "text": " in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of", "speaker": "Fraser Paterson"}, {"start": 1532.973, "end": 1538.437, "text": " It seems like chapter four kind of serves as a background where the story gets started for the most part.", "speaker": "Marc Broberg"}, {"start": 1538.458, "end": 1542.375, "text": "And this is kind of more of a ground level", "speaker": "Marc Broberg"}, {"start": 1542.457, "end": 1544.78, "text": " view, which is nice and helpful.", "speaker": "Marc Broberg"}, {"start": 1546.603, "end": 1554.433, "text": "What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function.", "speaker": "Marc Broberg"}, {"start": 1554.453, "end": 1559.68, "text": "Because in my mind, I'm kind of conditioned to seeing these Gaussians, right?", "speaker": "Marc Broberg"}, {"start": 1559.7, "end": 1566.089, "text": "You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior.", "speaker": "Marc Broberg"}, {"start": 1566.75, "end": 1571.316, "text": "However, here we have this linear function, which I guess is", "speaker": "Marc Broberg"}, {"start": 1572.207, "end": 1575.451, "text": " then becomes part of, say, a likelihood.", "speaker": "Marc Broberg"}, {"start": 1575.491, "end": 1588.409, "text": "And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense.", "speaker": "Marc Broberg"}, {"start": 1588.429, "end": 1588.83, "text": "Yeah, yeah.", "speaker": "Fraser Paterson"}, {"start": 1589.651, "end": 1590.212, "text": "No, absolutely.", "speaker": "Fraser Paterson"}, {"start": 1590.252, "end": 1600.746, "text": "I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1601.131, "end": 1628.91, "text": " um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so", "speaker": "Fraser Paterson"}, {"start": 1629.464, "end": 1638.597, "text": " This is, in some sense, still our assumption about what the relationship is between observations and the hidden states.", "speaker": "Fraser Paterson"}, {"start": 1639.066, "end": 1646.86, "text": " For our purpose, we've assumed that the real world, we're sort of creating a world here.", "speaker": "Fraser Paterson"}, {"start": 1646.94, "end": 1651.348, "text": "This is the example of the food size and light intensity world.", "speaker": "Fraser Paterson"}, {"start": 1651.468, "end": 1653.572, "text": "This is the really simple world that we've seen for all of part one.", "speaker": "Fraser Paterson"}, {"start": 1654.153, "end": 1655.475, "text": "So what are the hidden states?", "speaker": "Fraser Paterson"}, {"start": 1656.076, "end": 1658.04, "text": "They are the sizes of the food.", "speaker": "Fraser Paterson"}, {"start": 1658.46, "end": 1661.666, "text": "And we can see that there's five different sizes, one to five.", "speaker": "Fraser Paterson"}, {"start": 1661.907, "end": 1662.728, "text": "That's the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1663.552, "end": 1665.214, "text": " And the observations are light intensities.", "speaker": "Fraser Paterson"}, {"start": 1665.354, "end": 1668.318, "text": "So light comes in, hits the food, and you get some sort of light intensity.", "speaker": "Fraser Paterson"}, {"start": 1668.758, "end": 1672.903, "text": "And depending on the size of the food, you get some specific light intensity.", "speaker": "Fraser Paterson"}, {"start": 1672.924, "end": 1685.619, "text": "So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size.", "speaker": "Fraser Paterson"}, {"start": 1685.639, "end": 1688.783, "text": "So we're saying that the relationship between the observation you get in", "speaker": "Fraser Paterson"}, {"start": 1689.235, "end": 1709.573, "text": " is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters", "speaker": "Fraser Paterson"}, {"start": 1710.548, "end": 1714.279, "text": " So again, this is the environment regenerative process.", "speaker": "Fraser Paterson"}, {"start": 1714.941, "end": 1721.942, "text": "The corresponding model is this one, 2.11, equation 2.11, I believe.", "speaker": "Fraser Paterson"}, {"start": 1722.513, "end": 1751.187, "text": " now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs", "speaker": "Fraser Paterson"}, {"start": 1751.555, "end": 1762.486, "text": " are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution.", "speaker": "Fraser Paterson"}, {"start": 1762.526, "end": 1780.684, "text": "So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations.", "speaker": "Fraser Paterson"}, {"start": 1781.204, "end": 1783.246, "text": " And the thing is, we're uncertain.", "speaker": "Fraser Paterson"}, {"start": 1784.268, "end": 1789.194, "text": "So the agent doesn't get to see what the real relationship is between food size and light intensity.", "speaker": "Fraser Paterson"}, {"start": 1789.334, "end": 1789.995, "text": "It has no idea.", "speaker": "Fraser Paterson"}, {"start": 1791.136, "end": 1793.679, "text": "It has a model about what it thinks the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1794.24, "end": 1796.443, "text": "Now, we know that the relationship is linear.", "speaker": "Fraser Paterson"}, {"start": 1796.483, "end": 1805.994, "text": "We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity.", "speaker": "Fraser Paterson"}, {"start": 1806.014, "end": 1809.118, "text": "But in the real world, we have no idea what the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1809.385, "end": 1813.791, "text": " It turns out that in this example, we've specified exactly the same relationship.", "speaker": "Fraser Paterson"}, {"start": 1814.752, "end": 1816.975, "text": "So that's quite a nice scenario to be in.", "speaker": "Fraser Paterson"}, {"start": 1817.676, "end": 1826.968, "text": "You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size.", "speaker": "Fraser Paterson"}, {"start": 1828.429, "end": 1834.397, "text": "Now that's wrong, but it's OK, maybe, for some settings.", "speaker": "Fraser Paterson"}, {"start": 1834.985, "end": 1843.699, "text": " That would be an example of a likelihood mapping that would be incorrect with respect to its generating function.", "speaker": "Fraser Paterson"}, {"start": 1844.721, "end": 1861.448, "text": "So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount.", "speaker": "Fraser Paterson"}, {"start": 1862.103, "end": 1869.693, "text": " And it's centered about the relationship we think is the case between the hidden states and the observations.", "speaker": "Fraser Paterson"}, {"start": 1870.234, "end": 1873.658, "text": "So the reason why we have this probability distribution is because we don't know.", "speaker": "Fraser Paterson"}, {"start": 1873.678, "end": 1877.924, "text": "We do not actually ever get to know what the real relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1877.944, "end": 1881.208, "text": "So does that answer your question, Mark, about why", "speaker": "Fraser Paterson"}, {"start": 1881.458, "end": 1887.967, "text": " why probability distributions, maybe why normal distributions, or did that help at all?", "speaker": "Fraser Paterson"}, {"start": 1888.007, "end": 1889.469, "text": "Yeah, that helped a lot.", "speaker": "Marc Broberg"}, {"start": 1890.511, "end": 1890.951, "text": "Thank you.", "speaker": "Marc Broberg"}, {"start": 1891.011, "end": 1900.164, "text": "Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right?", "speaker": "Marc Broberg"}, {"start": 1900.464, "end": 1908.175, "text": "And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right?", "speaker": "Marc Broberg"}, {"start": 1909.235, "end": 1934.061, "text": " uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all", "speaker": "Fraser Paterson"}, {"start": 1934.648, "end": 1938.613, "text": " We could plug, we could say, okay, the relationship is just equal to the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1938.813, "end": 1942.738, "text": "I think that the observation is exactly the hidden state, right?", "speaker": "Fraser Paterson"}, {"start": 1942.758, "end": 1943.438, "text": "That's an assumption.", "speaker": "Fraser Paterson"}, {"start": 1944.86, "end": 1952.529, "text": "Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1953.49, "end": 1955.653, "text": "Or we can have a quadratic relationship.", "speaker": "Fraser Paterson"}, {"start": 1955.673, "end": 1957.635, "text": "We can have anything we want in there at all, actually.", "speaker": "Fraser Paterson"}, {"start": 1957.856, "end": 1963.843, "text": "And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1964.083, "end": 1964.183, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 1966.036, "end": 1983.758, "text": " very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B,", "speaker": "Fraser Paterson"}, {"start": 1986.978, "end": 2006.903, "text": " Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from.", "speaker": "Fraser Paterson"}, {"start": 2006.923, "end": 2009.526, "text": "Michael Morehead , Other questions other questions Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2011.582, "end": 2015.987, "text": " Okay, my question is this.", "speaker": "Magdalena Hurtado"}, {"start": 2016.608, "end": 2019.672, "text": "So could we do a thought experiment for a minute?", "speaker": "Magdalena Hurtado"}, {"start": 2019.712, "end": 2025.319, "text": "Because I think if I set it up as a thought experiment, I'll get the answer.", "speaker": "Magdalena Hurtado"}, {"start": 2026.821, "end": 2027.121, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2027.221, "end": 2027.942, "text": "Oh, so I'm inside.", "speaker": "Magdalena Hurtado"}, {"start": 2027.962, "end": 2030.966, "text": "Okay, so let's imagine the following.", "speaker": "Magdalena Hurtado"}, {"start": 2031.907, "end": 2040.357, "text": "In present-day societies, we have this societal", "speaker": "Magdalena Hurtado"}, {"start": 2040.775, "end": 2050.769, "text": " back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2052.391, "end": 2060.623, "text": "And there's a tension in societies to update that belief by some group to science.", "speaker": "Magdalena Hurtado"}, {"start": 2060.89, "end": 2063.737, "text": " Exists a science so God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2063.837, "end": 2071.114, "text": "So science tells us that God may not exist or something to that effect or quantum physics explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2071.214, "end": 2071.735, "text": "Not God.", "speaker": "Magdalena Hurtado"}, {"start": 2072.016, "end": 2072.597, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2072.617, "end": 2074.682, "text": "So there's a process, right?", "speaker": "Magdalena Hurtado"}, {"start": 2074.782, "end": 2077.228, "text": "That's in in the environment.", "speaker": "Magdalena Hurtado"}, {"start": 2077.208, "end": 2096.28, "text": " that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed", "speaker": "Magdalena Hurtado"}, {"start": 2096.733, "end": 2118.994, "text": " And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications.", "speaker": "Magdalena Hurtado"}, {"start": 2118.974, "end": 2148.88, "text": " for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is", "speaker": "Magdalena Hurtado"}, {"start": 2148.86, "end": 2163.055, "text": " Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2165.296, "end": 2191.017, "text": " are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds", "speaker": "Magdalena Hurtado"}, {"start": 2191.25, "end": 2195.815, "text": " Okay, as a third level of information processing.", "speaker": "Magdalena Hurtado"}, {"start": 2196.296, "end": 2217.46, "text": "And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2217.44, "end": 2219.723, "text": " Yeah, no, excellent, excellent question, Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2219.964, "end": 2242.857, "text": "This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents.", "speaker": "Fraser Paterson"}, {"start": 2243.748, "end": 2250.721, "text": " Under what conditions do they interact so as to form a larger composite active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2251.462, "end": 2257.513, "text": "So it's kind of a question like, how many active inference agents are there really in any given picture?", "speaker": "Fraser Paterson"}, {"start": 2260.018, "end": 2266.87, "text": "So your question, I think, insofar as I understand it, is, look, we have some notion about agency.", "speaker": "Fraser Paterson"}, {"start": 2267.423, "end": 2268.825, "text": " inactive inference.", "speaker": "Fraser Paterson"}, {"start": 2268.845, "end": 2274.433, "text": "We have some notion about something interacting with its environment, whatever the environment is, right?", "speaker": "Fraser Paterson"}, {"start": 2274.453, "end": 2278.098, "text": "So on the right hand side, we've got an agent, and on the left hand side, we have an environment.", "speaker": "Fraser Paterson"}, {"start": 2279.36, "end": 2293.699, "text": "But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent.", "speaker": "Fraser Paterson"}, {"start": 2293.84, "end": 2295.642, "text": "Okay, we haven't been able to tell that story yet.", "speaker": "Fraser Paterson"}, {"start": 2296.837, "end": 2311.04, "text": " Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers.", "speaker": "Fraser Paterson"}, {"start": 2311.12, "end": 2314.385, "text": "Let me see if I can find that figure.", "speaker": "Fraser Paterson"}, {"start": 2315.206, "end": 2317.83, "text": "Various layers within the predictive coding hierarchy.", "speaker": "Fraser Paterson"}, {"start": 2318.672, "end": 2319.433, "text": "Which one is it here?", "speaker": "Fraser Paterson"}, {"start": 2319.473, "end": 2320.074, "text": "This one?", "speaker": "Fraser Paterson"}, {"start": 2320.354, "end": 2320.735, "text": "No, this one.", "speaker": "Fraser Paterson"}, {"start": 2322.487, "end": 2329.997, "text": " You might interpret the various layers here as individual active implementations that are just doing local free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 2330.017, "end": 2333.682, "text": "So maybe this guy is an active implementation in some sense.", "speaker": "Fraser Paterson"}, {"start": 2333.702, "end": 2335.825, "text": "And then he's passing messages, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 2336.445, "end": 2338.989, "text": "And then the overall thing can be regarded as an active implementation.", "speaker": "Fraser Paterson"}, {"start": 2339.029, "end": 2342.393, "text": "That's the only inkling that we've seen thus far of this idea yet.", "speaker": "Fraser Paterson"}, {"start": 2343.234, "end": 2347.6, "text": "We're also not really going to see a lot of it in the rest of the book.", "speaker": "Fraser Paterson"}, {"start": 2348.052, "end": 2349.956, "text": " because it is a very open question.", "speaker": "Fraser Paterson"}, {"start": 2350.017, "end": 2351.38, "text": "It's a very difficult question.", "speaker": "Fraser Paterson"}, {"start": 2352.302, "end": 2360.36, "text": "And it gets at the heart of what agency even is at all, which is a question over and above active inference per se.", "speaker": "Fraser Paterson"}, {"start": 2360.781, "end": 2367.276, "text": "However, my personal interpretation or my personal thoughts about this issue", "speaker": "Fraser Paterson"}, {"start": 2367.678, "end": 2374.214, "text": " is that the thing that is currently missing from active inference is exactly this idea of compositionality.", "speaker": "Fraser Paterson"}, {"start": 2374.234, "end": 2375.938, "text": "So I've got agent here, agent here.", "speaker": "Fraser Paterson"}, {"start": 2376.219, "end": 2376.841, "text": "They interact.", "speaker": "Fraser Paterson"}, {"start": 2377.482, "end": 2381.873, "text": "Under what conditions is that whole thing one active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2382.275, "end": 2388.688, "text": " That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research.", "speaker": "Fraser Paterson"}, {"start": 2388.709, "end": 2390.352, "text": "So I hope that answers your question in some respect.", "speaker": "Fraser Paterson"}, {"start": 2390.532, "end": 2391.494, "text": "Very, very important question.", "speaker": "Fraser Paterson"}, {"start": 2392.015, "end": 2397.828, "text": "But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 2398.81, "end": 2399.912, "text": "OK.", "speaker": "Magdalena Hurtado"}, {"start": 2400.23, "end": 2425.562, "text": " really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups", "speaker": "Magdalena Hurtado"}, {"start": 2425.542, "end": 2430.369, "text": " What I see and I'm going to use this language, I'm not a mathematician.", "speaker": "Magdalena Hurtado"}, {"start": 2430.429, "end": 2435.577, "text": "OK, so forgive me, but but I kind of I kind of get some things in math.", "speaker": "Magdalena Hurtado"}, {"start": 2435.597, "end": 2442.167, "text": "OK, so topologically generative models really work as huge attractors.", "speaker": "Magdalena Hurtado"}, {"start": 2442.147, "end": 2444.451, "text": " in informational systems in humans.", "speaker": "Magdalena Hurtado"}, {"start": 2444.912, "end": 2458.638, "text": "So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right?", "speaker": "Magdalena Hurtado"}, {"start": 2458.718, "end": 2462.265, "text": "So your language is a technology in human systems.", "speaker": "Magdalena Hurtado"}, {"start": 2462.245, "end": 2490.784, "text": " so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model", "speaker": "Magdalena Hurtado"}, {"start": 2491.692, "end": 2504.479, "text": " And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math.", "speaker": "Magdalena Hurtado"}, {"start": 2505.702, "end": 2509.55, "text": "If I just kind of sorry, is that okay?", "speaker": "Andrew Pashea"}, {"start": 2509.63, "end": 2509.951, "text": "All right.", "speaker": "Andrew Pashea"}, {"start": 2510.091, "end": 2510.632, "text": "Thanks.", "speaker": "Andrew Pashea"}, {"start": 2510.652, "end": 2512.536, "text": "Just ask someone who has a.", "speaker": "Andrew Pashea"}, {"start": 2512.516, "end": 2515.059, "text": " More immediate background in this social sciences.", "speaker": "Andrew Pashea"}, {"start": 2515.72, "end": 2516.001, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 2516.141, "end": 2516.822, "text": "So, so.", "speaker": "Andrew Pashea"}, {"start": 2517.323, "end": 2521.729, "text": "And I like how the question that Mark had prior to this had to do with, like.", "speaker": "Andrew Pashea"}, {"start": 2522.33, "end": 2524.332, "text": "You know, what is it to make linear assumptions?", "speaker": "Andrew Pashea"}, {"start": 2524.853, "end": 2527.777, "text": "So, in the context of the textbook, what we've seen thus far, like.", "speaker": "Andrew Pashea"}, {"start": 2528.278, "end": 2535.908, "text": "This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2535.948, "end": 2539.774, "text": "And it's just to show how a model can be composed.", "speaker": "Andrew Pashea"}, {"start": 2540.755, "end": 2542.337, "text": "Like, how can.", "speaker": "Andrew Pashea"}, {"start": 2542.57, "end": 2551.986, "text": " In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right?", "speaker": "Andrew Pashea"}, {"start": 2552.407, "end": 2560.16, "text": "It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right?", "speaker": "Andrew Pashea"}, {"start": 2560.18, "end": 2561.422, "text": "So that's the action part.", "speaker": "Andrew Pashea"}, {"start": 2561.783, "end": 2563.245, "text": "We haven't reached that yet.", "speaker": "Andrew Pashea"}, {"start": 2563.225, "end": 2572.756, "text": " But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function?", "speaker": "Andrew Pashea"}, {"start": 2573.317, "end": 2590.597, "text": "So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right?", "speaker": "Andrew Pashea"}, {"start": 2591.198, "end": 2593.08, "text": "So that linear model we were looking at", "speaker": "Andrew Pashea"}, {"start": 2593.06, "end": 2616.688, "text": " we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people", "speaker": "Andrew Pashea"}, {"start": 2616.938, "end": 2619.842, "text": " what are the kinds of sensory information that I receive?", "speaker": "Andrew Pashea"}, {"start": 2620.522, "end": 2622.965, "text": "And what are the kinds of actions that I can do?", "speaker": "Andrew Pashea"}, {"start": 2623.005, "end": 2625.308, "text": "And what do I have beliefs about?", "speaker": "Andrew Pashea"}, {"start": 2625.408, "end": 2629.173, "text": "Or what variables do I think explains everything, right?", "speaker": "Andrew Pashea"}, {"start": 2629.774, "end": 2640.887, "text": "So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human", "speaker": "Andrew Pashea"}, {"start": 2640.867, "end": 2655.518, "text": " uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like", "speaker": "Andrew Pashea"}, {"start": 2656.021, "end": 2658.344, "text": " There's, there's a lot to track.", "speaker": "Andrew Pashea"}, {"start": 2658.364, "end": 2662.21, "text": "So there are some people who have tried to model this sort of thing.", "speaker": "Andrew Pashea"}, {"start": 2662.23, "end": 2663.572, "text": "And so I've shared this paper.", "speaker": "Andrew Pashea"}, {"start": 2664.313, "end": 2672.665, "text": "It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference.", "speaker": "Andrew Pashea"}, {"start": 2673.226, "end": 2676.19, "text": "And this kind of has to do with.", "speaker": "Andrew Pashea"}, {"start": 2676.507, "end": 2693.875, "text": " Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere.", "speaker": "Andrew Pashea"}, {"start": 2693.855, "end": 2711.429, "text": " And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing.", "speaker": "Andrew Pashea"}, {"start": 2711.489, "end": 2714.715, "text": "And that's one way of trying to look at a problem.", "speaker": "Andrew Pashea"}, {"start": 2714.695, "end": 2734.948, "text": " like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one", "speaker": "Andrew Pashea"}, {"start": 2735.333, "end": 2737.86, "text": " And a lot of agents who rally around idea two.", "speaker": "Andrew Pashea"}, {"start": 2737.9, "end": 2743.315, "text": "And as they do that, the two groups start to communicate directly with each other less and less.", "speaker": "Andrew Pashea"}, {"start": 2743.997, "end": 2749.613, "text": "The agent's actual actions that are simulated has to do with them communicating with one another and sharing their", "speaker": "Andrew Pashea"}, {"start": 2750.015, "end": 2753.178, "text": " They're kind of more verbal beliefs with one another.", "speaker": "Andrew Pashea"}, {"start": 2753.199, "end": 2764.771, "text": "And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing.", "speaker": "Andrew Pashea"}, {"start": 2764.791, "end": 2769.256, "text": "And somehow that gets all blended in with our assumptions of how science works and things like that.", "speaker": "Andrew Pashea"}, {"start": 2769.757, "end": 2771.338, "text": "It's much more complex.", "speaker": "Andrew Pashea"}, {"start": 2771.479, "end": 2779.988, "text": "And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy", "speaker": "Andrew Pashea"}, {"start": 2779.968, "end": 2783.633, "text": " But remember that free energy, based on how we've discussed it so far,", "speaker": "Andrew Pashea"}, {"start": 2783.967, "end": 2787.932, "text": " It's not some magical quantity of energy or something like that.", "speaker": "Andrew Pashea"}, {"start": 2788.914, "end": 2790.736, "text": "It's really just a proxy.", "speaker": "Andrew Pashea"}, {"start": 2790.796, "end": 2797.485, "text": "It's a word that is a proxy for something called surprisal, which is more or less uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2797.625, "end": 2799.608, "text": "So we want to minimize our uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2800.169, "end": 2813.687, "text": "So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another,", "speaker": "Andrew Pashea"}, {"start": 2813.667, "end": 2817.235, "text": " about is it idea one or is it idea two?", "speaker": "Andrew Pashea"}, {"start": 2817.716, "end": 2819.781, "text": "Is it science or is it something else?", "speaker": "Andrew Pashea"}, {"start": 2820.443, "end": 2824.993, "text": "One way to minimize uncertainty is to repeatedly tell each other", "speaker": "Andrew Pashea"}, {"start": 2825.328, "end": 2828.993, "text": " what the truth is and you all converge on what you all think the truth is.", "speaker": "Andrew Pashea"}, {"start": 2829.073, "end": 2829.293, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 2829.834, "end": 2838.545, "text": "And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much.", "speaker": "Andrew Pashea"}, {"start": 2839.085, "end": 2841.929, "text": "Because the other group is adding to your uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2842.129, "end": 2849.338, "text": "Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe.", "speaker": "Andrew Pashea"}, {"start": 2849.418, "end": 2854.845, "text": "So, so you see that there, there is a kind of like logic where the free energy principle is involved.", "speaker": "Andrew Pashea"}, {"start": 2855.129, "end": 2864.026, "text": " where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2864.467, "end": 2870.679, "text": "And that has nothing to do with the true validity of science, right?", "speaker": "Andrew Pashea"}, {"start": 2871.641, "end": 2875.348, "text": "Like I'm attempting to be a scientist explaining all this stuff,", "speaker": "Andrew Pashea"}, {"start": 2875.683, "end": 2878.748, "text": " But the people who, you see what I'm saying?", "speaker": "Andrew Pashea"}, {"start": 2878.788, "end": 2884.538, "text": "Like to take in sensory information is what we do.", "speaker": "Andrew Pashea"}, {"start": 2884.578, "end": 2888.544, "text": "It's not about like following proper logic and stuff.", "speaker": "Andrew Pashea"}, {"start": 2888.564, "end": 2892.19, "text": "It's people learn to do that, right?", "speaker": "Andrew Pashea"}, {"start": 2892.17, "end": 2894.737, "text": " Can I interject something real quick?", "speaker": "Magdalena Hurtado"}, {"start": 2894.757, "end": 2906.128, "text": "I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're", "speaker": "Magdalena Hurtado"}, {"start": 2906.581, "end": 2923.723, "text": " Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2924.364, "end": 2931.593, "text": "So what happens in humans groups is that the active inferencing that's happening in a lot of individuals,", "speaker": "Magdalena Hurtado"}, {"start": 2931.573, "end": 2937.991, "text": " is to analyze and update whether or not they believe the person who's giving the message.", "speaker": "Magdalena Hurtado"}, {"start": 2938.552, "end": 2942.864, "text": "It's not about evaluating the validity of the premises.", "speaker": "Magdalena Hurtado"}, {"start": 2943.486, "end": 2947.236, "text": "It's not about the... That gets at a very...", "speaker": "Magdalena Hurtado"}, {"start": 2947.216, "end": 2957.913, "text": " That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer.", "speaker": "Fraser Paterson"}, {"start": 2958.734, "end": 2961.178, "text": "We might imagine an active inference agent doing inference about this.", "speaker": "Fraser Paterson"}, {"start": 2961.759, "end": 2967.688, "text": "But above that is a question, how reliable is this thing that I'm hearing at all?", "speaker": "Fraser Paterson"}, {"start": 2967.668, "end": 2992.039, "text": " and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant", "speaker": "Fraser Paterson"}, {"start": 2992.019, "end": 3018.583, "text": " uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation", "speaker": "Fraser Paterson"}, {"start": 3018.563, "end": 3022.371, "text": " for what it means for one entity to be an active inference agent.", "speaker": "Fraser Paterson"}, {"start": 3022.391, "end": 3031.651, "text": "Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents.", "speaker": "Fraser Paterson"}, {"start": 3031.671, "end": 3033.595, "text": "But that's not going to be the focus of the book.", "speaker": "Fraser Paterson"}, {"start": 3034.396, "end": 3038.445, "text": "A lot of that is at the cutting edge of active inference research now.", "speaker": "Fraser Paterson"}, {"start": 3039.961, "end": 3044.874, "text": " Thanks for keeping it closer to the textbook by bringing up precision.", "speaker": "Andrew Pashea"}, {"start": 3044.934, "end": 3046.117, "text": "I very much agree with that.", "speaker": "Andrew Pashea"}, {"start": 3046.157, "end": 3052.915, "text": "And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but.", "speaker": "Andrew Pashea"}, {"start": 3054.459, "end": 3055.422, "text": "This.", "speaker": "Andrew Pashea"}, {"start": 3055.655, "end": 3065.723, "text": " Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference.", "speaker": "Andrew Pashea"}, {"start": 3065.743, "end": 3070.095, "text": "So active inferences is generally setting up these principles.", "speaker": "Andrew Pashea"}, {"start": 3070.115, "end": 3070.997, "text": "It.", "speaker": "Andrew Pashea"}, {"start": 3070.977, "end": 3092.985, "text": " definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different", "speaker": "Andrew Pashea"}, {"start": 3092.965, "end": 3093.386, "text": " Right?", "speaker": "Andrew Pashea"}, {"start": 3093.806, "end": 3097.411, "text": "Like, our variables that we're including in our respective models are different.", "speaker": "Andrew Pashea"}, {"start": 3097.772, "end": 3099.694, "text": "The precisions can definitely differ.", "speaker": "Andrew Pashea"}, {"start": 3099.714, "end": 3106.303, "text": "If I get information from someone in my group, I might have a higher precision and trust what they say more.", "speaker": "Andrew Pashea"}, {"start": 3106.764, "end": 3114.654, "text": "I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right?", "speaker": "Andrew Pashea"}, {"start": 3115.155, "end": 3118.72, "text": "So a lot of the, you know, a lot of the", "speaker": "Andrew Pashea"}, {"start": 3119.054, "end": 3131.018, "text": " political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right?", "speaker": "Andrew Pashea"}, {"start": 3131.159, "end": 3134.245, "text": "It's like, oh, I just lump them all in with political group", "speaker": "Andrew Pashea"}, {"start": 3134.225, "end": 3156.351, "text": " be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available", "speaker": "Andrew Pashea"}, {"start": 3156.331, "end": 3172.112, "text": " then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone", "speaker": "Andrew Pashea"}, {"start": 3172.092, "end": 3191.299, "text": " ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it", "speaker": "Andrew Pashea"}, {"start": 3192.308, "end": 3195.736, "text": " Do make sure, I've been absent for a little bit.", "speaker": "Fraser Paterson"}, {"start": 3196.237, "end": 3200.687, "text": "I'm gonna be much more active on the questions side of things on the CODA.", "speaker": "Fraser Paterson"}, {"start": 3200.748, "end": 3205.418, "text": "So do make sure if you've got questions, please do put them down in the questions tab here.", "speaker": "Fraser Paterson"}, {"start": 3205.458, "end": 3209.608, "text": "I've gone through and I've already answered quite a few of them for chapter five.", "speaker": "Fraser Paterson"}, {"start": 3209.588, "end": 3224.188, "text": " uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see", "speaker": "Fraser Paterson"}, {"start": 3224.573, "end": 3226.135, "text": " Are there any more live questions?", "speaker": "Fraser Paterson"}, {"start": 3227.777, "end": 3228.718, "text": "That would be nice.", "speaker": "Fraser Paterson"}, {"start": 3228.738, "end": 3233.584, "text": "We can stay around for maybe five or so minutes after the deadline if people so choose.", "speaker": "Fraser Paterson"}, {"start": 3234.645, "end": 3236.267, "text": "But we are coming up to the top of the hour.", "speaker": "Fraser Paterson"}, {"start": 3236.287, "end": 3241.994, "text": "So yeah, all of the questions in chapter five now are answered, or I've given my attempt.", "speaker": "Fraser Paterson"}, {"start": 3242.655, "end": 3246.379, "text": "As I say, I'll go through and attempt to answer all the previous ones as well.", "speaker": "Fraser Paterson"}, {"start": 3249.937, "end": 3276.737, "text": " think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry", "speaker": "Fraser Paterson"}, {"start": 3277.595, "end": 3301.122, "text": " okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian", "speaker": "Sun Xin"}, {"start": 3301.102, "end": 3308.391, "text": " inference to engage in parameter learning so that it can update beliefs about the current state.", "speaker": "Sun Xin"}, {"start": 3309.172, "end": 3319.545, "text": "If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself.", "speaker": "Sun Xin"}, {"start": 3319.985, "end": 3323.97, "text": "And I think this is what we call plasticity.", "speaker": "Sun Xin"}, {"start": 3323.95, "end": 3331.683, "text": " The ultimate goal of all this is to minimize expected free energy.", "speaker": "Sun Xin"}, {"start": 3332.404, "end": 3333.085, "text": "Is that right?", "speaker": "Sun Xin"}, {"start": 3334.548, "end": 3335.088, "text": "Wow, okay.", "speaker": "Fraser Paterson"}, {"start": 3335.129, "end": 3338.174, "text": "I mean, basically, yeah, that's substantively correct.", "speaker": "Fraser Paterson"}, {"start": 3339.115, "end": 3342.06, "text": "I will note, however, we haven't yet talked about expected free energy.", "speaker": "Fraser Paterson"}, {"start": 3342.395, "end": 3344.237, "text": " and planning and action selection.", "speaker": "Fraser Paterson"}, {"start": 3344.257, "end": 3345.358, "text": "So that is going to come later.", "speaker": "Fraser Paterson"}, {"start": 3346.379, "end": 3346.959, "text": "But yeah, you're right.", "speaker": "Fraser Paterson"}, {"start": 3346.979, "end": 3352.825, "text": "In terms of the general flavor of things, we have beliefs about hidden states in the world.", "speaker": "Fraser Paterson"}, {"start": 3353.686, "end": 3365.076, "text": "And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 3365.997, "end": 3369.921, "text": "Or we've seen that we can also recast this in terms of predictive processing.", "speaker": "Fraser Paterson"}, {"start": 3370.66, "end": 3372.943, "text": " minimizing the precision way to prediction errors.", "speaker": "Fraser Paterson"}, {"start": 3372.963, "end": 3373.924, "text": "It's the same kind of story.", "speaker": "Fraser Paterson"}, {"start": 3374.926, "end": 3386.221, "text": "And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states.", "speaker": "Fraser Paterson"}, {"start": 3387.063, "end": 3387.283, "text": "Yes.", "speaker": "Fraser Paterson"}, {"start": 3388.725, "end": 3391.328, "text": "The other thing, you mentioned parameter learning.", "speaker": "Fraser Paterson"}, {"start": 3391.388, "end": 3392.77, "text": "So yes, exactly.", "speaker": "Fraser Paterson"}, {"start": 3392.79, "end": 3393.932, "text": "There's these two problems.", "speaker": "Fraser Paterson"}, {"start": 3394.032, "end": 3397.317, "text": "We have the problem of inferring the hidden states.", "speaker": "Fraser Paterson"}, {"start": 3397.357, "end": 3398.338, "text": "What should the hidden states be?", "speaker": "Fraser Paterson"}, {"start": 3398.977, "end": 3407.788, "text": " But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference.", "speaker": "Fraser Paterson"}, {"start": 3408.529, "end": 3415.017, "text": "So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like.", "speaker": "Fraser Paterson"}, {"start": 3415.738, "end": 3419.183, "text": "But then above that, we need to deal with the problem of what should the setting of the parameters be?", "speaker": "Fraser Paterson"}, {"start": 3420.204, "end": 3426.853, "text": "And we've only really just begun to look at this in terms of prediction errors and hierarchical models.", "speaker": "Fraser Paterson"}, {"start": 3428.034, "end": 3428.815, "text": "But we saw", "speaker": "Fraser Paterson"}, {"start": 3429.149, "end": 3432.913, "text": " how to deal with that in terms of expectation maximization and earlier chapters.", "speaker": "Fraser Paterson"}, {"start": 3433.634, "end": 3443.644, "text": "We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture.", "speaker": "Fraser Paterson"}, {"start": 3443.964, "end": 3450.251, "text": "Because we generally can't do expectation maximization for the reason that we don't know the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 3450.291, "end": 3458.299, "text": "So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah.", "speaker": "Fraser Paterson"}, {"start": 3458.339, "end": 3458.599, "text": "Thank you.", "speaker": "Sun Xin"}, {"start": 3459.49, "end": 3461.011, "text": " No problem.", "speaker": "Fraser Paterson"}, {"start": 3461.032, "end": 3462.473, "text": "Do we have more questions?", "speaker": "Fraser Paterson"}, {"start": 3462.633, "end": 3463.254, "text": "More questions?", "speaker": "Fraser Paterson"}, {"start": 3464.655, "end": 3466.857, "text": "Maybe one more question, if there is one more.", "speaker": "Fraser Paterson"}, {"start": 3467.117, "end": 3469.52, "text": "And then I'll stop the recording.", "speaker": "Fraser Paterson"}, {"start": 3475.566, "end": 3475.866, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3475.886, "end": 3476.887, "text": "I might stop the recording here.", "speaker": "Fraser Paterson"}, {"start": 3478.068, "end": 3478.549, "text": "Well, Mark.", "speaker": "Fraser Paterson"}, {"start": 3478.569, "end": 3479.71, "text": "Hello, everyone, and welcome.", "speaker": "Fraser Paterson"}, {"start": 3479.91, "end": 3481.732, "text": "I've got my esteemed friend.", "speaker": "Fraser Paterson"}, {"start": 3482.753, "end": 3484.975, "text": "Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 3486.423, "end": 3489.528, "text": " That was interesting audio feedback.", "speaker": "Marc Broberg"}, {"start": 3489.548, "end": 3492.053, "text": "I thought I would jump in since nobody else asked us.", "speaker": "Marc Broberg"}, {"start": 3493.135, "end": 3499.105, "text": "But earlier you mentioned using gradient descent to learn parameters, right?", "speaker": "Marc Broberg"}, {"start": 3499.245, "end": 3501.289, "text": "Is that also used in perception?", "speaker": "Marc Broberg"}, {"start": 3501.69, "end": 3503.893, "text": "Is it the same approach?", "speaker": "Marc Broberg"}, {"start": 3503.994, "end": 3508.581, "text": "Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3508.602, "end": 3510.625, "text": "So it's a very, very general", "speaker": "Fraser Paterson"}, {"start": 3511.06, "end": 3512.522, "text": " optimization techniques.", "speaker": "Fraser Paterson"}, {"start": 3512.602, "end": 3520.19, "text": "So yes, it certainly can be used for learning, well, for perception.", "speaker": "Fraser Paterson"}, {"start": 3520.711, "end": 3522.513, "text": "I'm trying to find a nice picture here.", "speaker": "Fraser Paterson"}, {"start": 3523.173, "end": 3531.983, "text": "Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent.", "speaker": "Fraser Paterson"}, {"start": 3532.524, "end": 3533.445, "text": "Let me just make sure of that.", "speaker": "Fraser Paterson"}, {"start": 3533.565, "end": 3535.107, "text": "Yes, in other words, is the answer.", "speaker": "Fraser Paterson"}, {"start": 3535.547, "end": 3537.049, "text": "It's used for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3538.39, "end": 3539.852, "text": "Very, very general technique.", "speaker": "Fraser Paterson"}, {"start": 3540.237, "end": 3549.91, "text": " Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six.", "speaker": "Fraser Paterson"}, {"start": 3549.93, "end": 3550.39, "text": "Absolutely.", "speaker": "Fraser Paterson"}, {"start": 3550.51, "end": 3550.891, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3550.911, "end": 3552.593, "text": "We're gonna talk about phase planes and so on.", "speaker": "Fraser Paterson"}, {"start": 3553.394, "end": 3553.635, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3553.955, "end": 3556.258, "text": "There are problems with it, but yes, we're gonna see it going forward.", "speaker": "Fraser Paterson"}, {"start": 3556.278, "end": 3559.242, "text": "It's a very, it's a very foundational technique.", "speaker": "Fraser Paterson"}, {"start": 3559.382, "end": 3562.807, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3562.827, "end": 3563.027, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3563.047, "end": 3566.211, "text": "Any, any last minute questions for the YouTube recording before we, uh,", "speaker": "Fraser Paterson"}, {"start": 3569.093, "end": 3570.975, "text": " If not, I think I'm sorry.", "speaker": "Fraser Paterson"}, {"start": 3571.236, "end": 3572.377, "text": "Can I ask a quick question?", "speaker": "Dorsa"}, {"start": 3572.437, "end": 3574.2, "text": "Yeah, sure.", "speaker": "Fraser Paterson"}, {"start": 3574.74, "end": 3575.461, "text": "I do apologize.", "speaker": "Dorsa"}, {"start": 3575.702, "end": 3577.424, "text": "I have joined the group very late.", "speaker": "Dorsa"}, {"start": 3577.484, "end": 3579.767, "text": "So maybe this was addressed in like the past weeks.", "speaker": "Dorsa"}, {"start": 3579.827, "end": 3593.245, "text": "But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning?", "speaker": "Dorsa"}, {"start": 3593.478, "end": 3619.553, "text": " well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms", "speaker": "Fraser Paterson"}, {"start": 3619.955, "end": 3629.547, "text": " If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control.", "speaker": "Fraser Paterson"}, {"start": 3630.147, "end": 3643.784, "text": "There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference.", "speaker": "Fraser Paterson"}, {"start": 3643.804, "end": 3646.988, "text": "So the answer is yes, there is a profound relationship.", "speaker": "Fraser Paterson"}, {"start": 3647.008, "end": 3649.371, "text": "We probably don't have time to get into it here.", "speaker": "Fraser Paterson"}, {"start": 3649.84, "end": 3659.489, "text": " One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain.", "speaker": "Fraser Paterson"}, {"start": 3660.41, "end": 3665.735, "text": "Because we've seen with, well, we haven't yet seen with expected free energy how that works.", "speaker": "Fraser Paterson"}, {"start": 3665.755, "end": 3672.461, "text": "But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning.", "speaker": "Fraser Paterson"}, {"start": 3673.062, "end": 3679.628, "text": "And we don't just have this scale of reward, this signal in reinforcement learning.", "speaker": "Fraser Paterson"}, {"start": 3680.114, "end": 3708.67, "text": " we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um", "speaker": "Fraser Paterson"}, {"start": 3708.97, "end": 3717.107, "text": " a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3717.147, "end": 3727.688, "text": "Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning,", "speaker": "Andrew Pashea"}, {"start": 3727.668, "end": 3741.341, "text": " Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two.", "speaker": "Andrew Pashea"}, {"start": 3742.282, "end": 3752.331, "text": "But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven.", "speaker": "Andrew Pashea"}, {"start": 3753.192, "end": 3756.595, "text": "So there can be like a KL control agent or there can be a variety.", "speaker": "Andrew Pashea"}, {"start": 3756.575, "end": 3779.157, "text": " other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that", "speaker": "Andrew Pashea"}, {"start": 3779.137, "end": 3782.622, "text": " that things like curiosity exist.", "speaker": "Andrew Pashea"}, {"start": 3782.742, "end": 3787.429, "text": "Why does curiosity exist if we're actually always just driven towards a reward?", "speaker": "Andrew Pashea"}, {"start": 3787.509, "end": 3790.173, "text": "Once you know what the reward is, you should just go for that, right?", "speaker": "Andrew Pashea"}, {"start": 3790.213, "end": 3796.262, "text": "And you would have no reason to find some other strategy for going for it.", "speaker": "Andrew Pashea"}, {"start": 3796.463, "end": 3798.466, "text": "You would have no reason to go out of your way.", "speaker": "Andrew Pashea"}, {"start": 3798.606, "end": 3803.032, "text": "From what you already know, you just keep going for the reward as much as possible.", "speaker": "Andrew Pashea"}, {"start": 3803.393, "end": 3808.941, "text": "Perhaps you learn other things through happenstance along the way.", "speaker": "Andrew Pashea"}, {"start": 3809.258, "end": 3811.041, "text": " That's a big difference.", "speaker": "Fraser Paterson"}, {"start": 3811.722, "end": 3814.948, "text": "In reinforcement, you can just sort of start maximizing a reward signal.", "speaker": "Fraser Paterson"}, {"start": 3815.569, "end": 3822.601, "text": "But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards.", "speaker": "Fraser Paterson"}, {"start": 3822.621, "end": 3828.972, "text": "You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards.", "speaker": "Fraser Paterson"}, {"start": 3828.952, "end": 3856.731, "text": " yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite", "speaker": "Andrew Pashea"}, {"start": 3857.234, "end": 3868.402, "text": " how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3868.743, "end": 3872.492, "text": "Meanwhile, active inference has a much more principled way that relates to this notion of", "speaker": "Andrew Pashea"}, {"start": 3872.472, "end": 3892.793, "text": " free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually", "speaker": "Andrew Pashea"}, {"start": 3893.06, "end": 3896.992, "text": " Um, when I play a sport, I throw the ball this way.", "speaker": "Andrew Pashea"}, {"start": 3897.012, "end": 3906.681, "text": "Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person.", "speaker": "Andrew Pashea"}, {"start": 3906.897, "end": 3909.963, "text": " but maybe I will kind of curve it or change it, right?", "speaker": "Andrew Pashea"}, {"start": 3910.343, "end": 3914.03, "text": "You're not randomly throwing it in the sky or in the opposite direction.", "speaker": "Andrew Pashea"}, {"start": 3914.431, "end": 3919.481, "text": "You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right?", "speaker": "Andrew Pashea"}, {"start": 3919.541, "end": 3927.235, "text": "There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better.", "speaker": "Andrew Pashea"}, {"start": 3927.552, "end": 3953.037, "text": " based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about", "speaker": "Andrew Pashea"}, {"start": 3953.523, "end": 3956.968, "text": " How do I perform the best whenever it comes to active inference?", "speaker": "Andrew Pashea"}, {"start": 3957.129, "end": 3965.422, "text": "It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them.", "speaker": "Andrew Pashea"}, {"start": 3966.223, "end": 3975.918, "text": "But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear.", "speaker": "Andrew Pashea"}, {"start": 3976.038, "end": 3977.32, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3978.194, "end": 3983.299, "text": " I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys.", "speaker": "Fraser Paterson"}, {"start": 3983.319, "end": 3986.502, "text": "So Giancuomo, fire away.", "speaker": "Fraser Paterson"}, {"start": 3986.522, "end": 3991.827, "text": "Very quickly, I think it ties in with what was just spoken and talked about now.", "speaker": "SPEAKER_00"}, {"start": 3992.648, "end": 4007.982, "text": "Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is?", "speaker": "SPEAKER_00"}, {"start": 4008.265, "end": 4033.468, "text": " Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty.", "speaker": "SPEAKER_00"}, {"start": 4033.508, "end": 4035.029, "text": "They say, this is the answer.", "speaker": "SPEAKER_00"}, {"start": 4035.109, "end": 4037.091, "text": "And you're like, no, no, no, it's not.", "speaker": "SPEAKER_00"}, {"start": 4037.594, "end": 4057.919, "text": " And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware.", "speaker": "SPEAKER_00"}, {"start": 4057.959, "end": 4064.007, "text": "Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind?", "speaker": "SPEAKER_00"}, {"start": 4065.455, "end": 4093.251, "text": " well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions", "speaker": "Fraser Paterson"}, {"start": 4093.687, "end": 4094.989, "text": " This is built in from the start.", "speaker": "Fraser Paterson"}, {"start": 4095.029, "end": 4103.639, "text": "So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right?", "speaker": "Fraser Paterson"}, {"start": 4104.22, "end": 4105.982, "text": "And okay, we can't do Bayes' rule exactly.", "speaker": "Fraser Paterson"}, {"start": 4106.022, "end": 4107.023, "text": "We have to do it approximately.", "speaker": "Fraser Paterson"}, {"start": 4107.083, "end": 4108.365, "text": "We do variational inference.", "speaker": "Fraser Paterson"}, {"start": 4108.405, "end": 4110.407, "text": "We do prediction error and minimization, that kind of thing.", "speaker": "Fraser Paterson"}, {"start": 4110.828, "end": 4114.052, "text": "But we're always, always, always reasoning about probability distribution.", "speaker": "Fraser Paterson"}, {"start": 4114.072, "end": 4118.597, "text": "So yes, every single active implementation ever is always", "speaker": "Fraser Paterson"}, {"start": 4120.535, "end": 4147.757, "text": " predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet", "speaker": "Fraser Paterson"}, {"start": 4147.737, "end": 4149.84, "text": " And there's issues around that and how to do that.", "speaker": "Fraser Paterson"}, {"start": 4150.821, "end": 4151.722, "text": "But we haven't seen that just yet.", "speaker": "Fraser Paterson"}, {"start": 4151.742, "end": 4153.024, "text": "So, yes, absolutely.", "speaker": "Fraser Paterson"}, {"start": 4153.084, "end": 4155.808, "text": "It's part and parcel of what it is to be an active infatuation.", "speaker": "Fraser Paterson"}, {"start": 4155.828, "end": 4157.15, "text": "It's the reason about uncertainty now.", "speaker": "Fraser Paterson"}, {"start": 4158.411, "end": 4158.932, "text": "Okay, thank you.", "speaker": "Fraser Paterson"}, {"start": 4159.613, "end": 4159.913, "text": "No problem.", "speaker": "Fraser Paterson"}, {"start": 4159.933, "end": 4162.036, "text": "And then, Mark, and then I think we can do... Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 4162.337, "end": 4167.163, "text": "Maybe just really quickly, did you have a comment, Andrew?", "speaker": "Fraser Paterson"}, {"start": 4167.413, "end": 4171.819, "text": " I also really have to go, but yeah, just briefly, you basically answered it.", "speaker": "Andrew Pashea"}, {"start": 4171.839, "end": 4173.622, "text": "Sorry, I was looking at the chat.", "speaker": "Andrew Pashea"}, {"start": 4173.642, "end": 4177.928, "text": "It's just, yeah, it depends on how the question is being asked.", "speaker": "Andrew Pashea"}, {"start": 4177.948, "end": 4185.339, "text": "Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are?", "speaker": "Andrew Pashea"}, {"start": 4185.459, "end": 4191.348, "text": "Like that's going to take a little bit more than just like only looking at the simplified models.", "speaker": "Andrew Pashea"}, {"start": 4191.368, "end": 4193.05, "text": "We're looking at the textbook here.", "speaker": "Andrew Pashea"}, {"start": 4193.03, "end": 4216.928, "text": " um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization", "speaker": "Andrew Pashea"}, {"start": 4216.908, "end": 4230.176, "text": " of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots.", "speaker": "Andrew Pashea"}, {"start": 4230.256, "end": 4232.18, "text": "There are 2 possible things.", "speaker": "Andrew Pashea"}, {"start": 4232.22, "end": 4235.507, "text": "It could be is completely 50 50 is fully uncertain.", "speaker": "Andrew Pashea"}, {"start": 4235.487, "end": 4246.867, "text": " Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision.", "speaker": "Andrew Pashea"}, {"start": 4247.388, "end": 4247.609, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4247.709, "end": 4255.423, "text": "Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob.", "speaker": "Andrew Pashea"}, {"start": 4255.803, "end": 4256.044, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4256.104, "end": 4258.508, "text": "The more you turn up precision, the more you're.", "speaker": "Andrew Pashea"}, {"start": 4258.488, "end": 4263.134, "text": " you're going to take into account the prediction errors you're receiving and vice versa.", "speaker": "Andrew Pashea"}, {"start": 4263.214, "end": 4275.609, "text": "So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving.", "speaker": "Andrew Pashea"}, {"start": 4275.669, "end": 4275.89, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4275.91, "end": 4279.294, "text": "So so uncertainty is very much like throughout.", "speaker": "Andrew Pashea"}, {"start": 4279.474, "end": 4284.18, "text": "I mean, it's a very ubiquitous term through many parts of active inference.", "speaker": "Andrew Pashea"}, {"start": 4284.2, "end": 4284.34, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 4284.48, "end": 4285.742, "text": "So it's essential.", "speaker": "Andrew Pashea"}, {"start": 4287.224, "end": 4288.225, "text": "So all right.", "speaker": "Jim DeLong"}, {"start": 4288.677, "end": 4311.58, "text": " so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see", "speaker": "Jim DeLong"}, {"start": 4312.673, "end": 4337.179, "text": " right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh", "speaker": "Andrew Pashea"}, {"start": 4337.547, "end": 4338.789, "text": " who cares about priors.", "speaker": "Andrew Pashea"}, {"start": 4338.829, "end": 4342.634, "text": "And that's what we see with maximum likelihood estimation in chapter two there.", "speaker": "Andrew Pashea"}, {"start": 4343.655, "end": 4354.31, "text": "But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something.", "speaker": "Andrew Pashea"}, {"start": 4354.35, "end": 4366.086, "text": "It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like,", "speaker": "Andrew Pashea"}, {"start": 4366.403, "end": 4392.141, "text": " thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um", "speaker": "Andrew Pashea"}, {"start": 4392.661, "end": 4394.263, "text": " you know, how those connect up with each other.", "speaker": "Andrew Pashea"}, {"start": 4394.704, "end": 4403.396, "text": "So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there.", "speaker": "Andrew Pashea"}, {"start": 4403.476, "end": 4413.409, "text": "And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe.", "speaker": "Andrew Pashea"}, {"start": 4413.55, "end": 4416.193, "text": "And so I fully trust it without question.", "speaker": "Andrew Pashea"}, {"start": 4416.654, "end": 4416.854, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4416.874, "end": 4419.057, "text": "So I say all these things because, uh,", "speaker": "Andrew Pashea"}, {"start": 4419.037, "end": 4426.085, "text": " Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe.", "speaker": "Andrew Pashea"}, {"start": 4426.666, "end": 4442.303, "text": "And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things.", "speaker": "Andrew Pashea"}, {"start": 4442.343, "end": 4444.065, "text": "Yeah, it's very interesting to think about.", "speaker": "Andrew Pashea"}, {"start": 4445.294, "end": 4449.623, "text": " Yeah, do take a look at, I think I'm sharing, but figure 2.11.", "speaker": "Fraser Paterson"}, {"start": 4450.525, "end": 4459.965, "text": "This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief.", "speaker": "Fraser Paterson"}, {"start": 4459.985, "end": 4462.29, "text": "So if the precision is really tight around the prior,", "speaker": "Fraser Paterson"}, {"start": 4462.692, "end": 4472.808, "text": " the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model.", "speaker": "Fraser Paterson"}, {"start": 4472.828, "end": 4474.511, "text": "So yeah, do take a look at that.", "speaker": "Fraser Paterson"}, {"start": 4474.531, "end": 4482.584, "text": "That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have.", "speaker": "Fraser Paterson"}, {"start": 4482.684, "end": 4483.786, "text": "Yeah.", "speaker": "Jim DeLong"}, {"start": 4483.806, "end": 4484.146, "text": "Excellent.", "speaker": "Jim DeLong"}, {"start": 4484.206, "end": 4484.667, "text": "Thank you.", "speaker": "Jim DeLong"}, {"start": 4485.372, "end": 4485.512, "text": " Cool.", "speaker": "Fraser Paterson"}, {"start": 4485.532, "end": 4487.255, "text": "All right, guys, we're going to have to call it there.", "speaker": "Fraser Paterson"}, {"start": 4487.756, "end": 4491.963, "text": "Do put questions in the chats on the page, the code page.", "speaker": "Fraser Paterson"}, {"start": 4492.003, "end": 4493.606, "text": "I'll be a bit more attentive going forward.", "speaker": "Fraser Paterson"}, {"start": 4494.267, "end": 4495.489, "text": "I look forward to next week.", "speaker": "Fraser Paterson"}, {"start": 4495.509, "end": 4499.796, "text": "I'll be doing another session of this kind on Friday for the people who usually attend that session.", "speaker": "Fraser Paterson"}, {"start": 4499.816, "end": 4500.557, "text": "So thank you very much.", "speaker": "Fraser Paterson"}, {"start": 4500.918, "end": 4502.22, "text": "Stop sharing and stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4503.522, "end": 4504.704, "text": "Have we got a recording here?", "speaker": "Fraser Paterson"}, {"start": 4506.247, "end": 4507.269, "text": "Okay, stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4507.749, "end": 4508.551, "text": "Goodbye, YouTube people.", "speaker": "Fraser Paterson"}, {"start": 4508.751, "end": 4509.372, "text": "Until next time.", "speaker": "Fraser Paterson"}]}] \ No newline at end of file +[{"video_id": "MjeQeWeyYhE", "segments": [{"start": 3.727, "end": 29.498, "text": " all right hello everyone so we're here to it's june the 30th uh 2026 we're um in a bit of a sort of intermediary session for the fundamentals of active active infants textbook uh we have officially finished part one now yay so chapters one to four one to five sorry um and that marks quite a quite a milestone for for what we've done so we've kind of gone through the the fundamentals as it were the fundamentals of the fundamentals in some sense", "speaker": "Fraser Paterson"}, {"start": 29.798, "end": 40.317, "text": " And we're now going to be moving into part two, where we don't initially start making contact with action, but we're going to bring in action in part two.", "speaker": "Fraser Paterson"}, {"start": 40.337, "end": 43.923, "text": "We're going to do active inputs proper, which is going to be very, very exciting going forward.", "speaker": "Fraser Paterson"}, {"start": 43.943, "end": 50.575, "text": "So we've seen a huge amount of background, appreciated a lot of where things come from, general ideas, general concepts.", "speaker": "Fraser Paterson"}, {"start": 51.027, "end": 56.833, "text": " that we're going to start doing actual, quote unquote, active inference in the second part.", "speaker": "Fraser Paterson"}, {"start": 56.853, "end": 67.444, "text": "But this session, and indeed the session on Friday, which I will also host, at least going forward, that's the plan, this is meant to be just kind of reflection.", "speaker": "Fraser Paterson"}, {"start": 68.305, "end": 72.409, "text": "We're going to sort of slow down and think about part one in general, chapters one to five.", "speaker": "Fraser Paterson"}, {"start": 73.531, "end": 80.578, "text": "This will be an opportunity for people to ask questions to myself and Andrew live or otherwise about anything to do with part one.", "speaker": "Fraser Paterson"}, {"start": 80.812, "end": 85.079, "text": " you know, going so that we can be adequately grounded going forward.", "speaker": "Fraser Paterson"}, {"start": 85.981, "end": 94.696, "text": "I will say we have one one question is, you know, we've done each chapter across two weeks.", "speaker": "Fraser Paterson"}, {"start": 95.658, "end": 100.947, "text": "Historically, you know, chapter one, we've had two weeks going forward.", "speaker": "Fraser Paterson"}, {"start": 101.72, "end": 128.115, "text": " depending on what people might like we might like to do a review yet another week of review uh so you know this week and then next week as well and then go into chapter six from part two i know that there are there are a lot of people who are excited to have this review session basically um for you know part one not immediately jumping into part two so we're definitely gonna have this week but i would be interested to see if what what people's kind of thoughts are around maybe having an additional week", "speaker": "Fraser Paterson"}, {"start": 128.095, "end": 152.685, "text": " of getting more kind of review in part one where we can really consolidate really get our teeth around right you know sink our teeth into to the ideas that might be too much for some people we might lose momentum um maybe some people are eager to get the part two so it's an open question and we we can sort of do what we want we can decide you know do we want a second week or not um i see a lot of people in the chat are saying yes please so maybe", "speaker": "Fraser Paterson"}, {"start": 152.935, "end": 159.784, "text": " It would be a good idea to reach out through the blankets email.", "speaker": "Fraser Paterson"}, {"start": 160.425, "end": 166.993, "text": "Maybe if you are interested in a second week or on the Discord as well, that would be a very excellent place.", "speaker": "Fraser Paterson"}, {"start": 168.415, "end": 169.296, "text": "That would be amazing.", "speaker": "Fraser Paterson"}, {"start": 169.416, "end": 170.057, "text": "Good idea.", "speaker": "Fraser Paterson"}, {"start": 170.698, "end": 171.198, "text": "Yes, please.", "speaker": "Fraser Paterson"}, {"start": 171.238, "end": 176.445, "text": "I see a lot of people are saying yes.", "speaker": "Fraser Paterson"}, {"start": 176.695, "end": 182.461, "text": " In this chat here, it would be very helpful if you could give your yay or nay to that as well.", "speaker": "Fraser Paterson"}, {"start": 182.581, "end": 185.384, "text": "So we can actually look at that just directly through the chat here.", "speaker": "Fraser Paterson"}, {"start": 185.464, "end": 188.848, "text": "So if you want a second week review, yes.", "speaker": "Fraser Paterson"}, {"start": 189.108, "end": 190.089, "text": "If you don't, no.", "speaker": "Fraser Paterson"}, {"start": 190.71, "end": 192.252, "text": "And then we can go forward.", "speaker": "Fraser Paterson"}, {"start": 193.553, "end": 193.893, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 195.495, "end": 196.096, "text": "Enough of that.", "speaker": "Fraser Paterson"}, {"start": 196.616, "end": 197.898, "text": "I'll start sharing my screen here.", "speaker": "Fraser Paterson"}, {"start": 199.259, "end": 200.28, "text": "Probably my entire screen.", "speaker": "Fraser Paterson"}, {"start": 204.084, "end": 204.505, "text": "Okie dokie.", "speaker": "Fraser Paterson"}, {"start": 204.525, "end": 205.446, "text": "So people should be able to see.", "speaker": "Fraser Paterson"}, {"start": 205.486, "end": 206.567, "text": "Maybe just get rid of...", "speaker": "Fraser Paterson"}, {"start": 207.357, "end": 214.009, "text": " The session so people should be able to see chapter five I can't currently see you guys, so if you can't do do make some noise.", "speaker": "Fraser Paterson"}, {"start": 215.011, "end": 227.172, "text": "about what you are not seeing right now i'll just say so, you know we did chapter five last week in the week before what i've done is down here.", "speaker": "Fraser Paterson"}, {"start": 227.928, "end": 229.152, "text": " So you've got your overview.", "speaker": "Fraser Paterson"}, {"start": 229.935, "end": 232.945, "text": "I've added one or two additional resources.", "speaker": "Fraser Paterson"}, {"start": 232.965, "end": 233.827, "text": "So these actually came up.", "speaker": "Fraser Paterson"}, {"start": 233.908, "end": 235.593, "text": "Andrew shared the first of these.", "speaker": "Fraser Paterson"}, {"start": 235.613, "end": 236.436, "text": "These are all videos.", "speaker": "Fraser Paterson"}, {"start": 237.239, "end": 238.242, "text": "One of them is an active inference", "speaker": "Fraser Paterson"}, {"start": 240.668, "end": 241.791, "text": " coding and active inference.", "speaker": "Fraser Paterson"}, {"start": 243.094, "end": 245.139, "text": "So that's Ryan Smith and his team.", "speaker": "Fraser Paterson"}, {"start": 245.841, "end": 248.227, "text": "They've done excellent work on predictive coding.", "speaker": "Fraser Paterson"}, {"start": 248.708, "end": 249.45, "text": "I'll play that just now.", "speaker": "Fraser Paterson"}, {"start": 249.991, "end": 258.171, "text": "This is a great stream just about predictive coding more generally and its relation to active inference and its relation to", "speaker": "Fraser Paterson"}, {"start": 258.151, "end": 260.855, "text": " you know, its status in actual brains.", "speaker": "Fraser Paterson"}, {"start": 261.616, "end": 265.781, "text": "So if you're interested in more context, this would be excellent to go through.", "speaker": "Fraser Paterson"}, {"start": 265.801, "end": 267.003, "text": "And there's two more links there as well.", "speaker": "Fraser Paterson"}, {"start": 267.043, "end": 271.99, "text": "One by Jakob Howie, he's a philosopher over in Australia in the University of Monash.", "speaker": "Fraser Paterson"}, {"start": 272.01, "end": 274.013, "text": "What is predictive processing and what is it good for?", "speaker": "Fraser Paterson"}, {"start": 274.133, "end": 274.914, "text": "Another excellent talk.", "speaker": "Fraser Paterson"}, {"start": 275.495, "end": 280.041, "text": "And then a very, very good talk, which I absolutely love and adore, by Dr. John Verveke.", "speaker": "Fraser Paterson"}, {"start": 280.325, "end": 281.426, "text": " from the University of Toronto.", "speaker": "Fraser Paterson"}, {"start": 281.446, "end": 286.772, "text": "He's actually speaking about Neoplatonism and mystical experiences, believe it or not.", "speaker": "Fraser Paterson"}, {"start": 287.513, "end": 289.135, "text": "And I know it sounds weird.", "speaker": "Fraser Paterson"}, {"start": 289.575, "end": 292.799, "text": "Predictive processing actually shows up in here in a very significant way.", "speaker": "Fraser Paterson"}, {"start": 292.819, "end": 297.084, "text": "So that might be a nice mood setter, as it were.", "speaker": "Fraser Paterson"}, {"start": 297.164, "end": 298.365, "text": "So there's some additional resources there.", "speaker": "Fraser Paterson"}, {"start": 298.686, "end": 301.128, "text": "I've begun filling in the map for chapter five.", "speaker": "Fraser Paterson"}, {"start": 301.949, "end": 307.075, "text": "So you see chapters, well, section 5.1 and 5.2, there's a still lying fellow, I'm afraid.", "speaker": "Fraser Paterson"}, {"start": 307.73, "end": 334.446, "text": " the past week i've been horribly ill and i had a lot of deadlines to attend to so i've not been as attentive as i should be the sections 5.3 and 5.5 the the map is now there for those sections what i've tried to do i'm i'm going back and forth about this i'm of two minds i tend to express the mathematical equations in terms of latex formatting i don't expect that", "speaker": "Fraser Paterson"}, {"start": 334.814, "end": 338.84, "text": " A lot of you will be fluent in LaTeX, but this is a way to express mathematical notation.", "speaker": "Fraser Paterson"}, {"start": 339.401, "end": 353.881, "text": "What I'm going to do is I'm going to come back and I think I have a link to the existing equations in the actual equations tab in the coder so that you don't have to either directly read the LaTeX or try and put it into some place that will render it for you.", "speaker": "Fraser Paterson"}, {"start": 354.322, "end": 359.269, "text": "You can maybe just put this directly into an LLM that will explain what it is and maybe even render the text for you.", "speaker": "Fraser Paterson"}, {"start": 359.249, "end": 363.355, "text": " So it's not ideal, but the content is there at least.", "speaker": "Fraser Paterson"}, {"start": 363.736, "end": 367.081, "text": "And especially for those who don't have the book, I think this is quite useful.", "speaker": "Fraser Paterson"}, {"start": 367.101, "end": 374.753, "text": "So the idea with the chapter maps is that we do the same kind of thing we did for the overall content, so out here in the full chapter.", "speaker": "Fraser Paterson"}, {"start": 375.274, "end": 382.405, "text": "I try and break things down into what's the core idea, what are the core shifts in understanding, and then what's the core concepts", "speaker": "Fraser Paterson"}, {"start": 382.975, "end": 403.431, "text": " uh i don't know if i have that here what's previewed what's deferred and what's optional and then maybe some minimal takeaways and then i do that for each section in the book as well so hopefully that's somewhat useful especially for people who don't have the uh the book um but yeah as i say check i i need to uh let me get going with type of 105.2 so are there", "speaker": "Fraser Paterson"}, {"start": 404.373, "end": 404.954, "text": " Any questions?", "speaker": "Fraser Paterson"}, {"start": 404.994, "end": 416.831, "text": "Before we do, I'll just give, I think, a very brief 10-minute recap of chapters one to five, and then we can maybe get into some questions, live questions and written questions if people are wanting to do that.", "speaker": "Fraser Paterson"}, {"start": 418.033, "end": 419.795, "text": "I'll start my share temporarily.", "speaker": "Fraser Paterson"}, {"start": 420.756, "end": 421.858, "text": "Just coming back.", "speaker": "Fraser Paterson"}, {"start": 423.863, "end": 424.404, "text": " Very good.", "speaker": "Fraser Paterson"}, {"start": 425.745, "end": 426.386, "text": "Yes, but it's limited.", "speaker": "Fraser Paterson"}, {"start": 428.008, "end": 430.991, "text": "Formatting in LaTeX is a bit strange in Coda.", "speaker": "Fraser Paterson"}, {"start": 432.032, "end": 442.664, "text": "We have looked into that, but we'll be hopefully trying to make things a little bit easier in terms of the ability to read the equations that are displayed in Coda.", "speaker": "Fraser Paterson"}, {"start": 444.026, "end": 445.928, "text": "OK, so I'll share, come back.", "speaker": "Fraser Paterson"}, {"start": 447.89, "end": 452.976, "text": "Let's go all the way back, hopefully people can see my screen again, to the introduction.", "speaker": "Fraser Paterson"}, {"start": 454.441, "end": 456.885, "text": " So everyone's seen this figure here.", "speaker": "Fraser Paterson"}, {"start": 456.905, "end": 458.408, "text": "This is essentially the first figure in the book.", "speaker": "Fraser Paterson"}, {"start": 458.849, "end": 461.814, "text": "This is the breakdown of the partitions of the book, so part one, part two, part three.", "speaker": "Fraser Paterson"}, {"start": 462.396, "end": 468.046, "text": "We finished part one, so we've done chapters one to five, hypothesis, testing, brain, all the way to predictive coding.", "speaker": "Fraser Paterson"}, {"start": 469.649, "end": 471.232, "text": "Part two is active inference core.", "speaker": "Fraser Paterson"}, {"start": 471.252, "end": 473.215, "text": "So this is the heart of the book, really.", "speaker": "Fraser Paterson"}, {"start": 473.255, "end": 476.922, "text": "This is the fundamentals of active inference per se.", "speaker": "Fraser Paterson"}, {"start": 476.902, "end": 498.64, "text": " what we've done is we've looked at the kind of mathematical constituents and the sort of surrounding set of ideas in which active inference lives in terms of the bayesian brain hypothesis in terms of bayesian updating more generally approximate bayesian inference variational inference and then we've seen a slight twist on those ideas with predictive coding right at the very end", "speaker": "Fraser Paterson"}, {"start": 499.328, "end": 504.875, "text": " But as has been the case, everything has been just perception only.", "speaker": "Fraser Paterson"}, {"start": 504.895, "end": 506.057, "text": "So we haven't actually dealt with actions.", "speaker": "Fraser Paterson"}, {"start": 506.437, "end": 508.24, "text": "And we're going to be moving into that with part two.", "speaker": "Fraser Paterson"}, {"start": 509.221, "end": 511.564, "text": "But even more fundamentally, everything has been static.", "speaker": "Fraser Paterson"}, {"start": 511.824, "end": 516.651, "text": "So all of the models we've been looking at, I don't know if I can get a nice example.", "speaker": "Fraser Paterson"}, {"start": 517.592, "end": 519.014, "text": "All of the models, so this is chapter one.", "speaker": "Fraser Paterson"}, {"start": 519.995, "end": 523.159, "text": "Let's maybe go down to some of the equations.", "speaker": "Fraser Paterson"}, {"start": 523.443, "end": 524.244, "text": " They've all been static.", "speaker": "Fraser Paterson"}, {"start": 524.264, "end": 527.387, "text": "So the environment hasn't really been changing at all.", "speaker": "Fraser Paterson"}, {"start": 527.407, "end": 530.371, "text": "And this is very simple.", "speaker": "Fraser Paterson"}, {"start": 531.372, "end": 534.235, "text": "So part one, really going to chapter two, I think.", "speaker": "Fraser Paterson"}, {"start": 536.858, "end": 538.88, "text": "So we have an environment.", "speaker": "Fraser Paterson"}, {"start": 539.501, "end": 541.503, "text": "This is one we've seen this many times now.", "speaker": "Fraser Paterson"}, {"start": 541.523, "end": 548.691, "text": "This is a representation of the generative process, the real process in and of itself out there.", "speaker": "Fraser Paterson"}, {"start": 549.262, "end": 553.407, "text": " This is fairly static, so nothing is really changing.", "speaker": "Fraser Paterson"}, {"start": 553.467, "end": 555.609, "text": "The hidden states aren't really changing from moment to moment.", "speaker": "Fraser Paterson"}, {"start": 555.629, "end": 562.397, "text": "We're just presented with a situation and we have to update our beliefs about what hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 562.437, "end": 564.199, "text": "That's been constant throughout all of Chapter 1.", "speaker": "Fraser Paterson"}, {"start": 564.879, "end": 566.721, "text": "Things have gotten progressively more complicated.", "speaker": "Fraser Paterson"}, {"start": 566.741, "end": 570.906, "text": "We've looked at univariate hidden states where there's literally just one hidden state.", "speaker": "Fraser Paterson"}, {"start": 571.567, "end": 577.193, "text": "We've looked at multivariate hidden states that came up in Chapter 3 where suddenly we're dealing with vectors of things.", "speaker": "Fraser Paterson"}, {"start": 577.662, "end": 579.304, "text": " But in every case, it's been static.", "speaker": "Fraser Paterson"}, {"start": 579.884, "end": 585.331, "text": "The dynamics of the hidden state have been non-existent.", "speaker": "Fraser Paterson"}, {"start": 585.351, "end": 586.992, "text": "That is going to change going into part two.", "speaker": "Fraser Paterson"}, {"start": 587.052, "end": 591.918, "text": "We're going to start looking at generative processes which change over time.", "speaker": "Fraser Paterson"}, {"start": 592.579, "end": 593.94, "text": "And they're the interesting ones.", "speaker": "Fraser Paterson"}, {"start": 593.98, "end": 596.283, "text": "They're the ones that are actually useful.", "speaker": "Fraser Paterson"}, {"start": 596.343, "end": 602.55, "text": "And they're the ones that allow us to actually begin to have a reason to act in the environment.", "speaker": "Fraser Paterson"}, {"start": 602.61, "end": 605.593, "text": "We haven't really had any reason to perform actions yet.", "speaker": "Fraser Paterson"}, {"start": 606.805, "end": 611.032, "text": " So that's a big move that's going to take place and that we're going to have to deal with.", "speaker": "Fraser Paterson"}, {"start": 612.074, "end": 613.517, "text": "So maybe just coming back.", "speaker": "Fraser Paterson"}, {"start": 613.557, "end": 618.325, "text": "So chapter one was about, what was this about?", "speaker": "Fraser Paterson"}, {"start": 618.345, "end": 619.187, "text": "It was about perception.", "speaker": "Fraser Paterson"}, {"start": 619.427, "end": 619.588, "text": "Okay.", "speaker": "Fraser Paterson"}, {"start": 619.628, "end": 624.176, "text": "So like in terms of how we're going to think about perception in active inference.", "speaker": "Fraser Paterson"}, {"start": 624.757, "end": 627.261, "text": "So in chapter one, we cast or we framed", "speaker": "Fraser Paterson"}, {"start": 630.565, "end": 654.058, "text": " of perception as bayesian inference okay that's kind of the stance that active difference takes to the question what is perception what is perception it's bayesian inference or rather approximate bayesian inference so that was that was chapter one chapter two was then kind of an unpacking of this in terms of the mathematics um so we're still doing just perception we looked at", "speaker": "Fraser Paterson"}, {"start": 655.118, "end": 661.187, "text": " you know, how to do like exact Bayesian inference, okay, if we were able to do everything.", "speaker": "Fraser Paterson"}, {"start": 663.11, "end": 665.213, "text": "We saw how to use Bayes' rule basically.", "speaker": "Fraser Paterson"}, {"start": 665.233, "end": 675.187, "text": "So we cemented the distinction personally before that around the difference between the generative process and the generative model.", "speaker": "Fraser Paterson"}, {"start": 675.989, "end": 679.354, "text": "So I come all the way down to the figures here.", "speaker": "Fraser Paterson"}, {"start": 679.374, "end": 681.076, "text": "I don't have a particularly good figure for that.", "speaker": "Fraser Paterson"}, {"start": 681.096, "end": 682.358, "text": "Here we go, 2.4.", "speaker": "Fraser Paterson"}, {"start": 682.793, "end": 691.766, "text": " The really crucial distinction was this distinction between the environments, the generative model, the real world, and the agents, and its model of the real world.", "speaker": "Fraser Paterson"}, {"start": 692.307, "end": 695.191, "text": "This was the crucial thing, I think, in chapter two, basically.", "speaker": "Fraser Paterson"}, {"start": 695.992, "end": 699.978, "text": "And we have a way of notating this convention.", "speaker": "Fraser Paterson"}, {"start": 699.998, "end": 708.531, "text": "We have a convention for notating these differences that is used in the book where we say starred variables, so theta star, x star, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 708.971, "end": 712.176, "text": "These belong to the generative process, the real world out there.", "speaker": "Fraser Paterson"}, {"start": 712.797, "end": 716.822, "text": " Okay, and then modeled variables are without a star.", "speaker": "Fraser Paterson"}, {"start": 717.142, "end": 719.184, "text": "So they correspond to our model.", "speaker": "Fraser Paterson"}, {"start": 720.746, "end": 721.467, "text": "And what is our model?", "speaker": "Fraser Paterson"}, {"start": 722.668, "end": 730.037, "text": "It is quite literally a joint probability distribution over hidden states, latent states, and observations.", "speaker": "Fraser Paterson"}, {"start": 730.978, "end": 733.421, "text": "We use x for latent states and y for observations.", "speaker": "Fraser Paterson"}, {"start": 734.722, "end": 742.551, "text": "But you can see, of course, the x in here is our model of the real latent state outside of here, x star.", "speaker": "Fraser Paterson"}, {"start": 743.172, "end": 744.755, "text": " And they need not agree with each other.", "speaker": "Fraser Paterson"}, {"start": 744.795, "end": 745.737, "text": "They need not be the same.", "speaker": "Fraser Paterson"}, {"start": 746.839, "end": 752.69, "text": "But what we need is we need a generative model that allows us to make counterfactual predictions about what the hidden states might be.", "speaker": "Fraser Paterson"}, {"start": 752.731, "end": 754.795, "text": "That is p of x and y.", "speaker": "Fraser Paterson"}, {"start": 755.817, "end": 760.185, "text": "And then if we have the model evidence, the probability of observing anything,", "speaker": "Fraser Paterson"}, {"start": 760.722, "end": 769.453, "text": " or some particular observation given any hidden state, we can then do exact Bayesian inference to get our posterior belief about the hidden states given our observation.", "speaker": "Fraser Paterson"}, {"start": 770.174, "end": 772.037, "text": "And that was kind of chapter two.", "speaker": "Fraser Paterson"}, {"start": 772.057, "end": 779.526, "text": "And we saw ways of talking about this by talking about specific kinds of probability distributions, namely normal distributions or Gaussian distributions.", "speaker": "Fraser Paterson"}, {"start": 779.546, "end": 780.648, "text": "And these are very ubiquitous.", "speaker": "Fraser Paterson"}, {"start": 780.668, "end": 781.369, "text": "They show up everywhere.", "speaker": "Fraser Paterson"}, {"start": 782.43, "end": 783.331, "text": "We're going to continue that.", "speaker": "Fraser Paterson"}, {"start": 783.351, "end": 790.24, "text": "In fact, we're going to really intensify our use of normal distributions going into chapter six with continuous active inference.", "speaker": "Fraser Paterson"}, {"start": 790.642, "end": 791.784, "text": " There are other distributions as well.", "speaker": "Fraser Paterson"}, {"start": 791.804, "end": 792.725, "text": "We haven't really looked at those yet.", "speaker": "Fraser Paterson"}, {"start": 793.967, "end": 799.554, "text": "But the real sort of, there's a problem with this, which is that, yes, we can do inference.", "speaker": "Fraser Paterson"}, {"start": 799.574, "end": 803.079, "text": "Yes, we might even be able to do exact inference for really, really simple problems.", "speaker": "Fraser Paterson"}, {"start": 804.14, "end": 812.952, "text": "But in chapter two, we said, all right, we've got parameters, which are sort of like the settings on dials on our generative model.", "speaker": "Fraser Paterson"}, {"start": 813.793, "end": 815.075, "text": "And then we're going to do inference.", "speaker": "Fraser Paterson"}, {"start": 815.095, "end": 820.142, "text": "And inference is like, I set the dials, and then my machine does some inference stuff, right?", "speaker": "Fraser Paterson"}, {"start": 820.983, "end": 847.716, "text": " that was cool but the question about how to set the dials was completely skipped in in chapter two we just sort of assumed that the dials were set to be good and then we could do inference but really we need to figure out how do we actually set the dials on the inference process itself and that was chapter three so chapter three is kind of chapter two redux we were doing everything we were doing in chapter two we assumed we could do exact bayesian inference with grid approximation", "speaker": "Fraser Paterson"}, {"start": 848.303, "end": 850.966, "text": " And we're doing good old fashioned Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 851.346, "end": 854.99, "text": "But now we're having to do parameter learning and estimation as well.", "speaker": "Fraser Paterson"}, {"start": 856.692, "end": 867.903, "text": "And we saw for the first time that this is the first flavor that we have of what it means to be doing learning in active inference.", "speaker": "Fraser Paterson"}, {"start": 867.923, "end": 872.448, "text": "So we have these two processes, inference and learning parameter estimation.", "speaker": "Fraser Paterson"}, {"start": 873.109, "end": 877.213, "text": "And they're both able to be done by means of Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 878.56, "end": 879.421, "text": " That's the crucial thing.", "speaker": "Fraser Paterson"}, {"start": 879.441, "end": 889.438, "text": "So really a lot of this is just chapter two again, but we expressed, let me, uh, so yeah.", "speaker": "Fraser Paterson"}, {"start": 889.458, "end": 904.163, "text": "And look, one of the very, very crucial and central, um, ideas or mechanisms that is used for doing learning where we need to figure out what should the setting of the parameters be this idea of gradient descent.", "speaker": "Fraser Paterson"}, {"start": 904.868, "end": 908.953, "text": " And we saw many versions of this across the chapter and indeed in other chapters.", "speaker": "Fraser Paterson"}, {"start": 908.993, "end": 920.085, "text": "It's a very ubiquitous strategy in machine learning and artificial intelligence more generally, where you can imagine you've got some surface that corresponds to a cost of a certain setting of parameters.", "speaker": "Fraser Paterson"}, {"start": 920.185, "end": 926.172, "text": "So let's imagine we have our two parameters, I don't know, beta 1, beta 0.", "speaker": "Fraser Paterson"}, {"start": 926.232, "end": 934.521, "text": "And what you do to set the parameters to be good parameters is you have a notion about how costly each point is in this space.", "speaker": "Fraser Paterson"}, {"start": 934.872, "end": 935.733, "text": " That's this surface.", "speaker": "Fraser Paterson"}, {"start": 936.754, "end": 944.063, "text": "And then finding good parameters corresponds to descending this surface in such a way to get to the lowest possible point.", "speaker": "Fraser Paterson"}, {"start": 945.064, "end": 950.271, "text": "And then those, the setting of the parameters, those parameters at the lowest point, they're going to be your best parameters.", "speaker": "Fraser Paterson"}, {"start": 950.291, "end": 952.654, "text": "Very general way of doing optimization.", "speaker": "Fraser Paterson"}, {"start": 954.015, "end": 964.288, "text": "However, we saw that we could actually do that process can itself correspond to a process of inference where we're inferring what the parameters should be.", "speaker": "Fraser Paterson"}, {"start": 965.72, "end": 973.231, "text": " And that, so, you know, we saw that up here in 3.5 when we began to do expectation maximization.", "speaker": "Fraser Paterson"}, {"start": 975.574, "end": 985.308, "text": "So the idea is that, and this, believe it or not, is actually relevant to what we saw in chapter five, although we haven't really, we hadn't been able to appreciate that until now.", "speaker": "Fraser Paterson"}, {"start": 986.25, "end": 993.921, "text": "A lot of the time, because chapter five is about predictive coding and hierarchical models, hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 994.357, "end": 1021.09, "text": " um and we'll get to that we haven't got to that just now and a lot of the time with hierarchical models uh we we use them one of the reasons why we like to use them is because we imagine that there's multiple different kinds of processes that are happening at the same time and they might be happening at different time scales okay and indeed it is very very very common that when we're trying to solve the problem as to what should the good parameters be for our model", "speaker": "Fraser Paterson"}, {"start": 1021.525, "end": 1025.652, "text": " And also, how should I use those settings to do good inference?", "speaker": "Fraser Paterson"}, {"start": 1026.934, "end": 1040.558, "text": "It's very common to set the parameters and do inference at one time scale, and then at another time scale that's ticking along at a slower pace to do learning when we update our model parameters.", "speaker": "Fraser Paterson"}, {"start": 1040.578, "end": 1042.442, "text": "So imagine you've got learning up here.", "speaker": "Fraser Paterson"}, {"start": 1042.582, "end": 1043.744, "text": "What should the model parameters be?", "speaker": "Fraser Paterson"}, {"start": 1043.764, "end": 1046.048, "text": "And you can do inference on model parameters.", "speaker": "Fraser Paterson"}, {"start": 1046.45, "end": 1050.538, "text": " And then that can inform how you do inference at the low-level hidden states.", "speaker": "Fraser Paterson"}, {"start": 1050.558, "end": 1053.805, "text": "So there's kind of these two processes that are happening in two different timescales.", "speaker": "Fraser Paterson"}, {"start": 1053.825, "end": 1062.062, "text": "That's a very common framing for the problem of Bayesian inference and indeed machine learning more generally.", "speaker": "Fraser Paterson"}, {"start": 1063.122, "end": 1064.023, "text": " So that's kind of chapter three.", "speaker": "Fraser Paterson"}, {"start": 1064.143, "end": 1065.264, "text": "I'm going to race through this.", "speaker": "Fraser Paterson"}, {"start": 1065.925, "end": 1068.828, "text": "And we'll get to chapter five and then to some actual questions from you guys.", "speaker": "Fraser Paterson"}, {"start": 1069.749, "end": 1081.062, "text": "Chapter four, then, was a crucial turning point for us in terms of our understanding of the problem that needs to be solved in active inference.", "speaker": "Fraser Paterson"}, {"start": 1081.082, "end": 1084.966, "text": "So chapter three, we're still doing exact inference.", "speaker": "Fraser Paterson"}, {"start": 1084.986, "end": 1090.632, "text": "But we saw that really we culminated in an algorithm", "speaker": "Fraser Paterson"}, {"start": 1091.658, "end": 1113.632, "text": " allows us in so in chapter three we didn't know about hierarchical models yet okay we didn't really know about that language and it culminated in a very very important algorithm which is the expectation maximization algorithm which is to say a way to solve the problem of what should my model parameters be and then also what should my infinite hidden states be", "speaker": "Fraser Paterson"}, {"start": 1113.95, "end": 1115.672, "text": " Those two problems are related to one another.", "speaker": "Fraser Paterson"}, {"start": 1115.752, "end": 1118.615, "text": "To know the hidden states, you need to know what the good parameters are.", "speaker": "Fraser Paterson"}, {"start": 1118.935, "end": 1121.378, "text": "But to know what the good parameters are, you have to know how to infer hidden states well.", "speaker": "Fraser Paterson"}, {"start": 1121.958, "end": 1131.588, "text": "So to separate this problem, we came up with and we saw the solution to the chicken and egg problem, which was the expectation maximization algorithm.", "speaker": "Fraser Paterson"}, {"start": 1131.608, "end": 1132.689, "text": "I won't go over that again here.", "speaker": "Fraser Paterson"}, {"start": 1132.73, "end": 1134.131, "text": "Maybe we will if people want me to.", "speaker": "Fraser Paterson"}, {"start": 1135.072, "end": 1139.957, "text": "But that was the solution to the problem of chapter 3, that chicken and egg problem.", "speaker": "Fraser Paterson"}, {"start": 1140.781, "end": 1151.908, "text": " The expectation maximization algorithm, as it was presented there, assumes that we can do exact inference to find our exact Bayesian posterior of hidden state skewed observations.", "speaker": "Fraser Paterson"}, {"start": 1151.928, "end": 1155.938, "text": "And that, unfortunately, is usually not something we can do.", "speaker": "Fraser Paterson"}, {"start": 1156.542, "end": 1161.508, "text": " Being able to find the exact posterior is usually impossible, for reasons that we saw.", "speaker": "Fraser Paterson"}, {"start": 1161.708, "end": 1162.75, "text": "We can review again if you want.", "speaker": "Fraser Paterson"}, {"start": 1163.33, "end": 1166.895, "text": "So then chapter 4 says, all right, well, we can't do exact Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1167.536, "end": 1168.096, "text": "What are we going to do?", "speaker": "Fraser Paterson"}, {"start": 1168.116, "end": 1177.107, "text": "We're going to have to approximate the exact Bayesian inference somehow, because we'd still like to be able to do something expectation maximization-like.", "speaker": "Fraser Paterson"}, {"start": 1178.029, "end": 1180.812, "text": "So the idea with chapter 4 is, ah, OK, we're going to now", "speaker": "Fraser Paterson"}, {"start": 1181.804, "end": 1193.779, "text": " have a look at one way of doing approximate Bayesian inference, which is to say variational Bayesian inference, where we say we're not going to try and find the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 1194.7, "end": 1198.304, "text": "We're going to try and find a posterior that's close enough to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1198.705, "end": 1208.537, "text": "And this then motivated the idea of variational free energy as a quantity that is a measurement of how well", "speaker": "Fraser Paterson"}, {"start": 1208.838, "end": 1213.784, "text": " an approximate posterior fits or how close it is to the true posterior.", "speaker": "Fraser Paterson"}, {"start": 1214.885, "end": 1222.794, "text": "And that's kind of where we saw the idea that minimizing variational free energy is a bound, an upper bound on surprisal.", "speaker": "Fraser Paterson"}, {"start": 1223.055, "end": 1225.858, "text": "We saw from chapter three that surprisal is something we want to make very small.", "speaker": "Fraser Paterson"}, {"start": 1226.719, "end": 1237.191, "text": "So by minimizing this tractable quantity variational free energy, we can get something approximately, well, something that's approximately good enough to minimizing surprisal,", "speaker": "Fraser Paterson"}, {"start": 1237.778, "end": 1242.586, "text": " which is something we can't do directly, and therefore we can do approximate Bayesian inference.", "speaker": "Fraser Paterson"}, {"start": 1242.606, "end": 1248.976, "text": "So in a lot of ways, that really kind of spiritually is the end of part one, I would say.", "speaker": "Fraser Paterson"}, {"start": 1249.958, "end": 1265.102, "text": "The motivation as to where variational free energy comes from, its relation to surprisal, its relation to the minimization of surprisal, and the various forms of variational free energy, we saw that there's at least four kind of canonical ways to express the VFE.", "speaker": "Fraser Paterson"}, {"start": 1265.842, "end": 1269.509, "text": " That's, in my mind, really kind of the end of part one.", "speaker": "Fraser Paterson"}, {"start": 1270.311, "end": 1278.908, "text": "Chapter five was a nice kind of detour into a different way of thinking about what the VFE is.", "speaker": "Fraser Paterson"}, {"start": 1280.451, "end": 1287.445, "text": "So, you know, it's a different kind of a specialization about what it means to be doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 1288.387, "end": 1314.799, "text": " we saw we could express the vfe in terms of prediction errors and specifically precision weighted prediction errors and there's interesting notions about how this relates to things like attention um and various disorders of attention and so on a lot of this you know in terms of the historical development came from uh you know an independent line of inquiry to um the hardcore statistical", "speaker": "Fraser Paterson"}, {"start": 1314.965, "end": 1334.813, "text": " know physical methods from which variational free energy free energy minimization came from a lot of predictive coding and predictive processing this came from you know we're doing sort of studies on neurobiology and neuropsychology and we're looking at how literally how neurons do things and so on but then later it was", "speaker": "Fraser Paterson"}, {"start": 1335.535, "end": 1349.242, "text": " lots of crosses and bridges were observed to be possible to make between these two different ways of thinking about how intelligent things like brains do their intelligent things like inference and learning.", "speaker": "Fraser Paterson"}, {"start": 1350.083, "end": 1357.257, "text": "It turns out that we can very fruitfully re-express all those kind of same ideas that we saw in chapter four with variational free energy minimization", "speaker": "Fraser Paterson"}, {"start": 1357.574, "end": 1369.356, "text": " in terms of this you know precision weighted prediction error machinery okay and this is a very big sub uh discipline or sub you know a different way of thinking about that whole procedure", "speaker": "Fraser Paterson"}, {"start": 1370.163, "end": 1387.641, "text": " And indeed, there is an entire theory called predictive coding that kind of swings alongside active inference as a slightly different take on the Bayesian brain hypothesis, the idea that the brain is doing some kind of Bayesian inference somehow.", "speaker": "Fraser Paterson"}, {"start": 1388.962, "end": 1391.865, "text": "And that's kind of where we left off with part one.", "speaker": "Fraser Paterson"}, {"start": 1391.885, "end": 1396.39, "text": "So that is, in effect, what we've done, where we've gone.", "speaker": "Fraser Paterson"}, {"start": 1396.943, "end": 1400.268, "text": " That's a lot of stuff to cover, but again, we haven't looked at actions yet.", "speaker": "Fraser Paterson"}, {"start": 1400.829, "end": 1407.12, "text": "We're gonna be doing that in part two, and that will bring us full circle to active inference.", "speaker": "Fraser Paterson"}, {"start": 1407.14, "end": 1420.021, "text": "And we're gonna see there's all kinds of problems associated with how to do, how to select actions, how that relates to the problem of inference, sorry, just perception.", "speaker": "Fraser Paterson"}, {"start": 1420.922, "end": 1423.086, "text": "Okay, we're gonna see that these are deeply related to each other.", "speaker": "Fraser Paterson"}, {"start": 1423.623, "end": 1426.045, "text": " But that's the ground that we've covered.", "speaker": "Fraser Paterson"}, {"start": 1426.846, "end": 1429.829, "text": "I'd be interested if people have questions related to that.", "speaker": "Fraser Paterson"}, {"start": 1429.849, "end": 1431.29, "text": "I see that there's lots of questions.", "speaker": "Fraser Paterson"}, {"start": 1432.371, "end": 1432.912, "text": "Stop sharing.", "speaker": "Fraser Paterson"}, {"start": 1434.533, "end": 1437.476, "text": "And I see Mark has a question.", "speaker": "Fraser Paterson"}, {"start": 1437.496, "end": 1438.397, "text": "Please fire away, Mark.", "speaker": "Fraser Paterson"}, {"start": 1440.039, "end": 1440.519, "text": "As usual.", "speaker": "Marc Broberg"}, {"start": 1441.16, "end": 1443.041, "text": "Thank you so much for this review.", "speaker": "Marc Broberg"}, {"start": 1443.061, "end": 1443.602, "text": "This is great.", "speaker": "Marc Broberg"}, {"start": 1445.724, "end": 1451.389, "text": "I was wondering if you could pull up that slide that showed the generative process and the generative model.", "speaker": "Marc Broberg"}, {"start": 1451.429, "end": 1453.191, "text": "That might be helpful as a reference.", "speaker": "Marc Broberg"}, {"start": 1453.964, "end": 1455.166, "text": " Oh, yeah, okay.", "speaker": "Fraser Paterson"}, {"start": 1455.186, "end": 1456.187, "text": "So let me come back here.", "speaker": "Fraser Paterson"}, {"start": 1458.43, "end": 1462.415, "text": "That's the one in chapter two, I assume you're referring to?", "speaker": "Fraser Paterson"}, {"start": 1462.435, "end": 1462.836, "text": "I think so.", "speaker": "Marc Broberg"}, {"start": 1463.697, "end": 1464.138, "text": "This one here?", "speaker": "Fraser Paterson"}, {"start": 1464.618, "end": 1465.479, "text": "Yes, thank you.", "speaker": "Marc Broberg"}, {"start": 1466.1, "end": 1470.126, "text": "Yeah, I really just kind of appreciate what this textbook is trying to do.", "speaker": "Marc Broberg"}, {"start": 1470.146, "end": 1473.57, "text": "And I think it took me a while to appreciate that.", "speaker": "Marc Broberg"}, {"start": 1473.69, "end": 1482.402, "text": "But nonetheless, so we've got this model of the we've got this in the generative process, right, which is happening.", "speaker": "Marc Broberg"}, {"start": 1482.382, "end": 1505.689, "text": " in the environment and we we kind of make the commitment that this is the ground truth right in a sense and the only noise comes from our observations is that is that a fair uh summary of that well yeah so so far uh given that we've only conceived of things in terms of perception the only sort of source of noise", "speaker": "Marc Broberg"}, {"start": 1505.905, "end": 1532.082, "text": " in terms of our observation channel yeah so um it's maybe not as expressed here or as easily expressed here as one might like but yes so we can take in observations and we can emit actions uh and the only source of that noise is over the observation channel yeah precisely right okay my overall question uh is that i guess it has to start with the background of", "speaker": "Fraser Paterson"}, {"start": 1532.973, "end": 1538.437, "text": " It seems like chapter four kind of serves as a background where the story gets started for the most part.", "speaker": "Marc Broberg"}, {"start": 1538.458, "end": 1542.375, "text": "And this is kind of more of a ground level", "speaker": "Marc Broberg"}, {"start": 1542.457, "end": 1544.78, "text": " view, which is nice and helpful.", "speaker": "Marc Broberg"}, {"start": 1546.603, "end": 1554.433, "text": "What was new to me, and I'm still trying to wrap my head around, is this idea of the linear generating function.", "speaker": "Marc Broberg"}, {"start": 1554.453, "end": 1559.68, "text": "Because in my mind, I'm kind of conditioned to seeing these Gaussians, right?", "speaker": "Marc Broberg"}, {"start": 1559.7, "end": 1566.089, "text": "You've got your prior and your likelihood, and these are all affecting each other when you get to the posterior.", "speaker": "Marc Broberg"}, {"start": 1566.75, "end": 1571.316, "text": "However, here we have this linear function, which I guess is", "speaker": "Marc Broberg"}, {"start": 1572.207, "end": 1575.451, "text": " then becomes part of, say, a likelihood.", "speaker": "Marc Broberg"}, {"start": 1575.491, "end": 1588.409, "text": "And I'm just not quite able to make that bridge of reconciliation between the linear function and these Gaussian curves that we're seeing, if that makes sense.", "speaker": "Marc Broberg"}, {"start": 1588.429, "end": 1588.83, "text": "Yeah, yeah.", "speaker": "Fraser Paterson"}, {"start": 1589.651, "end": 1590.212, "text": "No, absolutely.", "speaker": "Fraser Paterson"}, {"start": 1590.252, "end": 1600.746, "text": "I mean, so that's a good point, because for a lot of part one, we have assumed this linear relationship between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1601.131, "end": 1628.91, "text": " um so well in terms of the generative process and we notate this here i think this is in terms of this is um i want to give you an equation in terms of where it is so 2.10 okay that i'm not sure which pages is on just yet this is our representation of the generative process the real thing in itself the real environment um for the for our purposes we're going to assume the real environment is constituted like this okay so", "speaker": "Fraser Paterson"}, {"start": 1629.464, "end": 1638.597, "text": " This is, in some sense, still our assumption about what the relationship is between observations and the hidden states.", "speaker": "Fraser Paterson"}, {"start": 1639.066, "end": 1646.86, "text": " For our purpose, we've assumed that the real world, we're sort of creating a world here.", "speaker": "Fraser Paterson"}, {"start": 1646.94, "end": 1651.348, "text": "This is the example of the food size and light intensity world.", "speaker": "Fraser Paterson"}, {"start": 1651.468, "end": 1653.572, "text": "This is the really simple world that we've seen for all of part one.", "speaker": "Fraser Paterson"}, {"start": 1654.153, "end": 1655.475, "text": "So what are the hidden states?", "speaker": "Fraser Paterson"}, {"start": 1656.076, "end": 1658.04, "text": "They are the sizes of the food.", "speaker": "Fraser Paterson"}, {"start": 1658.46, "end": 1661.666, "text": "And we can see that there's five different sizes, one to five.", "speaker": "Fraser Paterson"}, {"start": 1661.907, "end": 1662.728, "text": "That's the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1663.552, "end": 1665.214, "text": " And the observations are light intensities.", "speaker": "Fraser Paterson"}, {"start": 1665.354, "end": 1668.318, "text": "So light comes in, hits the food, and you get some sort of light intensity.", "speaker": "Fraser Paterson"}, {"start": 1668.758, "end": 1672.903, "text": "And depending on the size of the food, you get some specific light intensity.", "speaker": "Fraser Paterson"}, {"start": 1672.924, "end": 1685.619, "text": "So as the food size is larger, what happens in the real world is that the light intensity grows linearly with the food size.", "speaker": "Fraser Paterson"}, {"start": 1685.639, "end": 1688.783, "text": "So we're saying that the relationship between the observation you get in", "speaker": "Fraser Paterson"}, {"start": 1689.235, "end": 1709.573, "text": " is just equal to the size of the food times some number plus some number and this relationship is a line it's linear so that's the relationship between the hidden states and the observations but the question about the belief that we have about the hidden state that's where the Gaussian stuff enters", "speaker": "Fraser Paterson"}, {"start": 1710.548, "end": 1714.279, "text": " So again, this is the environment regenerative process.", "speaker": "Fraser Paterson"}, {"start": 1714.941, "end": 1721.942, "text": "The corresponding model is this one, 2.11, equation 2.11, I believe.", "speaker": "Fraser Paterson"}, {"start": 1722.513, "end": 1751.187, "text": " now here initially uh sanjeev is motivating things in terms of you know we need to have our likelihood about you know what we think of what we think we will get in terms of observations depending on the hidden state and a prior belief about what we think the hidden state is really like and those two things correspond to our generative model but precisely you know those beliefs", "speaker": "Fraser Paterson"}, {"start": 1751.555, "end": 1762.486, "text": " are usually expressed, at least in this stage, in terms of Gaussian distributions, normal distributions, Gaussian distributions, same thing, really, or uniform distribution.", "speaker": "Fraser Paterson"}, {"start": 1762.526, "end": 1780.684, "text": "So what this is saying is that, for our likelihood, we're assuming that our observation will be given according to a normal distribution centered about the model that we have of observations.", "speaker": "Fraser Paterson"}, {"start": 1781.204, "end": 1783.246, "text": " And the thing is, we're uncertain.", "speaker": "Fraser Paterson"}, {"start": 1784.268, "end": 1789.194, "text": "So the agent doesn't get to see what the real relationship is between food size and light intensity.", "speaker": "Fraser Paterson"}, {"start": 1789.334, "end": 1789.995, "text": "It has no idea.", "speaker": "Fraser Paterson"}, {"start": 1791.136, "end": 1793.679, "text": "It has a model about what it thinks the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1794.24, "end": 1796.443, "text": "Now, we know that the relationship is linear.", "speaker": "Fraser Paterson"}, {"start": 1796.483, "end": 1805.994, "text": "We know that the relationship between light intensity and food size is you take the food size, you multiply it by a number, and you add another number, and that gives you the light intensity.", "speaker": "Fraser Paterson"}, {"start": 1806.014, "end": 1809.118, "text": "But in the real world, we have no idea what the relationship is.", "speaker": "Fraser Paterson"}, {"start": 1809.385, "end": 1813.791, "text": " It turns out that in this example, we've specified exactly the same relationship.", "speaker": "Fraser Paterson"}, {"start": 1814.752, "end": 1816.975, "text": "So that's quite a nice scenario to be in.", "speaker": "Fraser Paterson"}, {"start": 1817.676, "end": 1826.968, "text": "You might imagine a different model that the agent has, where it says, you know what, I think the relationship between light intensity and the food size is beta 1 times the food size.", "speaker": "Fraser Paterson"}, {"start": 1828.429, "end": 1834.397, "text": "Now that's wrong, but it's OK, maybe, for some settings.", "speaker": "Fraser Paterson"}, {"start": 1834.985, "end": 1843.699, "text": " That would be an example of a likelihood mapping that would be incorrect with respect to its generating function.", "speaker": "Fraser Paterson"}, {"start": 1844.721, "end": 1861.448, "text": "So the uncertainty about the mapping between observations and hidden states is encoded by the fact that we are reasoning about a probability distribution, which is this normal thing spread out by some amount.", "speaker": "Fraser Paterson"}, {"start": 1862.103, "end": 1869.693, "text": " And it's centered about the relationship we think is the case between the hidden states and the observations.", "speaker": "Fraser Paterson"}, {"start": 1870.234, "end": 1873.658, "text": "So the reason why we have this probability distribution is because we don't know.", "speaker": "Fraser Paterson"}, {"start": 1873.678, "end": 1877.924, "text": "We do not actually ever get to know what the real relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1877.944, "end": 1881.208, "text": "So does that answer your question, Mark, about why", "speaker": "Fraser Paterson"}, {"start": 1881.458, "end": 1887.967, "text": " why probability distributions, maybe why normal distributions, or did that help at all?", "speaker": "Fraser Paterson"}, {"start": 1888.007, "end": 1889.469, "text": "Yeah, that helped a lot.", "speaker": "Marc Broberg"}, {"start": 1890.511, "end": 1890.951, "text": "Thank you.", "speaker": "Marc Broberg"}, {"start": 1891.011, "end": 1900.164, "text": "Yeah, it's just recognizing, okay, there's a linear process, and then we have beliefs about said linear process, which are going to be probabilistic, right?", "speaker": "Marc Broberg"}, {"start": 1900.464, "end": 1908.175, "text": "And so it's linear in a sense, because what we're assuming about the generating process is that it's deterministic, is that right?", "speaker": "Marc Broberg"}, {"start": 1909.235, "end": 1934.061, "text": " uh well we're assuming that yeah like if we go back let's go back to the actual equation so our assumption about the relationship between hidden states and observations is that the hidden states the the the observation is a function of the hidden state and what is the function it's this thing and that just happens to be the equation for a line and so it's linear um we could plug anything in there at all", "speaker": "Fraser Paterson"}, {"start": 1934.648, "end": 1938.613, "text": " We could plug, we could say, okay, the relationship is just equal to the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1938.813, "end": 1942.738, "text": "I think that the observation is exactly the hidden state, right?", "speaker": "Fraser Paterson"}, {"start": 1942.758, "end": 1943.438, "text": "That's an assumption.", "speaker": "Fraser Paterson"}, {"start": 1944.86, "end": 1952.529, "text": "Or we can have this assumption where we say, all right, we think the observations are given by some number plus some other number times the hidden state.", "speaker": "Fraser Paterson"}, {"start": 1953.49, "end": 1955.653, "text": "Or we can have a quadratic relationship.", "speaker": "Fraser Paterson"}, {"start": 1955.673, "end": 1957.635, "text": "We can have anything we want in there at all, actually.", "speaker": "Fraser Paterson"}, {"start": 1957.856, "end": 1963.843, "text": "And it's up to us as modelers to plug in what we think the relationship is between hidden states and observations.", "speaker": "Fraser Paterson"}, {"start": 1964.083, "end": 1964.183, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 1966.036, "end": 1983.758, "text": " very good, thank you so much, no problem, yeah so, that's not going to happen, beta one, let's have an A, let's have AX plus B,", "speaker": "Fraser Paterson"}, {"start": 1986.978, "end": 2006.903, "text": " Michael Morehead , Second time yeah so if I have a X plus B, and I can change these numbers around here so a right the the thing that I get is a line, so this relationship is literally a line it's linear so yeah that's that's where that comes from.", "speaker": "Fraser Paterson"}, {"start": 2006.923, "end": 2009.526, "text": "Michael Morehead , Other questions other questions Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2011.582, "end": 2015.987, "text": " Okay, my question is this.", "speaker": "Magdalena Hurtado"}, {"start": 2016.608, "end": 2019.672, "text": "So could we do a thought experiment for a minute?", "speaker": "Magdalena Hurtado"}, {"start": 2019.712, "end": 2025.319, "text": "Because I think if I set it up as a thought experiment, I'll get the answer.", "speaker": "Magdalena Hurtado"}, {"start": 2026.821, "end": 2027.121, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2027.221, "end": 2027.942, "text": "Oh, so I'm inside.", "speaker": "Magdalena Hurtado"}, {"start": 2027.962, "end": 2030.966, "text": "Okay, so let's imagine the following.", "speaker": "Magdalena Hurtado"}, {"start": 2031.907, "end": 2040.357, "text": "In present-day societies, we have this societal", "speaker": "Magdalena Hurtado"}, {"start": 2040.775, "end": 2050.769, "text": " back and forth that's debated across nations, groups, populations, etc., where some people have the belief that God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2052.391, "end": 2060.623, "text": "And there's a tension in societies to update that belief by some group to science.", "speaker": "Magdalena Hurtado"}, {"start": 2060.89, "end": 2063.737, "text": " Exists a science so God exists.", "speaker": "Magdalena Hurtado"}, {"start": 2063.837, "end": 2071.114, "text": "So science tells us that God may not exist or something to that effect or quantum physics explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2071.214, "end": 2071.735, "text": "Not God.", "speaker": "Magdalena Hurtado"}, {"start": 2072.016, "end": 2072.597, "text": "Okay.", "speaker": "Magdalena Hurtado"}, {"start": 2072.617, "end": 2074.682, "text": "So there's a process, right?", "speaker": "Magdalena Hurtado"}, {"start": 2074.782, "end": 2077.228, "text": "That's in in the environment.", "speaker": "Magdalena Hurtado"}, {"start": 2077.208, "end": 2096.28, "text": " that in the individual minds are listening to and have to decide am i going to update to which generative model okay having said that in human societies what happens is that you have the per you have the individual mind right so that's markov blanketed", "speaker": "Magdalena Hurtado"}, {"start": 2096.733, "end": 2118.994, "text": " And then you have one plus minds, which can be, you know, if you look at the anthropological literature, it appears that if there's like a 30 individuals in a hunter-gatherer band across most of our evolution, it's like 30 individuals are kind of like a distributed system where minds are interacting with each other and it has implications.", "speaker": "Magdalena Hurtado"}, {"start": 2118.974, "end": 2148.88, "text": " for that's kind of like the done by number in terms of what we're efficacious yeah so we yeah we can play with it and say okay there's some some number of a small group coordinating and being able to persist and then um they then then we get to the level of 30 plus mines which which is like a public level right so where you have a lot more mines right so my question is", "speaker": "Magdalena Hurtado"}, {"start": 2148.86, "end": 2163.055, "text": " Are these are the active inference models that we have looked at so far in the first few chapters agnostic with respect to the information processing unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2165.296, "end": 2191.017, "text": " are they agnostic with respect to the information processing unit of analysis well first of all so so let me so fraser let me just say this just for my clarity so one mind is one unit of analysis of 30 individuals can be in human history another important unit of analysis in our social cultural systems and then 30 plus minds", "speaker": "Magdalena Hurtado"}, {"start": 2191.25, "end": 2195.815, "text": " Okay, as a third level of information processing.", "speaker": "Magdalena Hurtado"}, {"start": 2196.296, "end": 2217.46, "text": "And so my question is like, as someone who doesn't know the math, right, who's trying to struggle with it, where would I see in this entire book an appreciation or discussion of what is the informational process unit of analysis?", "speaker": "Magdalena Hurtado"}, {"start": 2217.44, "end": 2219.723, "text": " Yeah, no, excellent, excellent question, Magdalena.", "speaker": "Fraser Paterson"}, {"start": 2219.964, "end": 2242.857, "text": "This is actually, if I understand you correctly, this is actually the subject of my PhD, which is, yeah, so, and this, first of all, I pull up a paper here by one of my good friends, Peter, Peter Wade, as one and many relating individual and emergent group level generative models and active inference, which, and this whole paper is about, all right, let's say I have a bunch of active inference agents.", "speaker": "Fraser Paterson"}, {"start": 2243.748, "end": 2250.721, "text": " Under what conditions do they interact so as to form a larger composite active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2251.462, "end": 2257.513, "text": "So it's kind of a question like, how many active inference agents are there really in any given picture?", "speaker": "Fraser Paterson"}, {"start": 2260.018, "end": 2266.87, "text": "So your question, I think, insofar as I understand it, is, look, we have some notion about agency.", "speaker": "Fraser Paterson"}, {"start": 2267.423, "end": 2268.825, "text": " inactive inference.", "speaker": "Fraser Paterson"}, {"start": 2268.845, "end": 2274.433, "text": "We have some notion about something interacting with its environment, whatever the environment is, right?", "speaker": "Fraser Paterson"}, {"start": 2274.453, "end": 2278.098, "text": "So on the right hand side, we've got an agent, and on the left hand side, we have an environment.", "speaker": "Fraser Paterson"}, {"start": 2279.36, "end": 2293.699, "text": "But and thus far, that's the story that we have, we don't have any notion about the relationship between agents, and whether or not that relationship can constitute an overall agent.", "speaker": "Fraser Paterson"}, {"start": 2293.84, "end": 2295.642, "text": "Okay, we haven't been able to tell that story yet.", "speaker": "Fraser Paterson"}, {"start": 2296.837, "end": 2311.04, "text": " Except for the very, very end of chapter five, where we began to look at hierarchical predictive coding and hierarchical active inference, you might, and indeed some people have interpreted various layers.", "speaker": "Fraser Paterson"}, {"start": 2311.12, "end": 2314.385, "text": "Let me see if I can find that figure.", "speaker": "Fraser Paterson"}, {"start": 2315.206, "end": 2317.83, "text": "Various layers within the predictive coding hierarchy.", "speaker": "Fraser Paterson"}, {"start": 2318.672, "end": 2319.433, "text": "Which one is it here?", "speaker": "Fraser Paterson"}, {"start": 2319.473, "end": 2320.074, "text": "This one?", "speaker": "Fraser Paterson"}, {"start": 2320.354, "end": 2320.735, "text": "No, this one.", "speaker": "Fraser Paterson"}, {"start": 2322.487, "end": 2329.997, "text": " You might interpret the various layers here as individual active implementations that are just doing local free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 2330.017, "end": 2333.682, "text": "So maybe this guy is an active implementation in some sense.", "speaker": "Fraser Paterson"}, {"start": 2333.702, "end": 2335.825, "text": "And then he's passing messages, blah, blah, blah.", "speaker": "Fraser Paterson"}, {"start": 2336.445, "end": 2338.989, "text": "And then the overall thing can be regarded as an active implementation.", "speaker": "Fraser Paterson"}, {"start": 2339.029, "end": 2342.393, "text": "That's the only inkling that we've seen thus far of this idea yet.", "speaker": "Fraser Paterson"}, {"start": 2343.234, "end": 2347.6, "text": "We're also not really going to see a lot of it in the rest of the book.", "speaker": "Fraser Paterson"}, {"start": 2348.052, "end": 2349.956, "text": " because it is a very open question.", "speaker": "Fraser Paterson"}, {"start": 2350.017, "end": 2351.38, "text": "It's a very difficult question.", "speaker": "Fraser Paterson"}, {"start": 2352.302, "end": 2360.36, "text": "And it gets at the heart of what agency even is at all, which is a question over and above active inference per se.", "speaker": "Fraser Paterson"}, {"start": 2360.781, "end": 2367.276, "text": "However, my personal interpretation or my personal thoughts about this issue", "speaker": "Fraser Paterson"}, {"start": 2367.678, "end": 2374.214, "text": " is that the thing that is currently missing from active inference is exactly this idea of compositionality.", "speaker": "Fraser Paterson"}, {"start": 2374.234, "end": 2375.938, "text": "So I've got agent here, agent here.", "speaker": "Fraser Paterson"}, {"start": 2376.219, "end": 2376.841, "text": "They interact.", "speaker": "Fraser Paterson"}, {"start": 2377.482, "end": 2381.873, "text": "Under what conditions is that whole thing one active inference agent?", "speaker": "Fraser Paterson"}, {"start": 2382.275, "end": 2388.688, "text": " That story has not really been told yet in Active Inference, and I'm actually hoping to tell it as far as I can in my own research.", "speaker": "Fraser Paterson"}, {"start": 2388.709, "end": 2390.352, "text": "So I hope that answers your question in some respect.", "speaker": "Fraser Paterson"}, {"start": 2390.532, "end": 2391.494, "text": "Very, very important question.", "speaker": "Fraser Paterson"}, {"start": 2392.015, "end": 2397.828, "text": "But yeah, as far as we've seen at all, it's at the very end of chapter five, really, with this idea of hierarchical predictive coding.", "speaker": "Fraser Paterson"}, {"start": 2398.81, "end": 2399.912, "text": "OK.", "speaker": "Magdalena Hurtado"}, {"start": 2400.23, "end": 2425.562, "text": " really like your answer and I would like I would like to see your dissertation and I would also like if you can share in the chat your email or something because I would like to talk to you more about this this is absolutely crucial to me because they generate because looked at from a cultural anthropological perspective when you look at human groups", "speaker": "Magdalena Hurtado"}, {"start": 2425.542, "end": 2430.369, "text": " What I see and I'm going to use this language, I'm not a mathematician.", "speaker": "Magdalena Hurtado"}, {"start": 2430.429, "end": 2435.577, "text": "OK, so forgive me, but but I kind of I kind of get some things in math.", "speaker": "Magdalena Hurtado"}, {"start": 2435.597, "end": 2442.167, "text": "OK, so topologically generative models really work as huge attractors.", "speaker": "Magdalena Hurtado"}, {"start": 2442.147, "end": 2444.451, "text": " in informational systems in humans.", "speaker": "Magdalena Hurtado"}, {"start": 2444.912, "end": 2458.638, "text": "So if you shift your folk, so when you ask in a human group, what's really interesting about Homo sapiens, it's fascinating, is that you can take a generative model that is spoken, right?", "speaker": "Magdalena Hurtado"}, {"start": 2458.718, "end": 2462.265, "text": "So your language is a technology in human systems.", "speaker": "Magdalena Hurtado"}, {"start": 2462.245, "end": 2490.784, "text": " so take a generative model you present it to to in in some context social context and the generative model determines the the importance of the generative model determines whether or not you're going to have one mind or 20 minds or 100 000 minds um marco blanketed around the process of updating a generative model", "speaker": "Magdalena Hurtado"}, {"start": 2491.692, "end": 2504.479, "text": " And so you're saying that that math doesn't is is being debated, navigated in active inference and I would really like to be able to to see what's being done there as I'm learning the math.", "speaker": "Magdalena Hurtado"}, {"start": 2505.702, "end": 2509.55, "text": "If I just kind of sorry, is that okay?", "speaker": "Andrew Pashea"}, {"start": 2509.63, "end": 2509.951, "text": "All right.", "speaker": "Andrew Pashea"}, {"start": 2510.091, "end": 2510.632, "text": "Thanks.", "speaker": "Andrew Pashea"}, {"start": 2510.652, "end": 2512.536, "text": "Just ask someone who has a.", "speaker": "Andrew Pashea"}, {"start": 2512.516, "end": 2515.059, "text": " More immediate background in this social sciences.", "speaker": "Andrew Pashea"}, {"start": 2515.72, "end": 2516.001, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 2516.141, "end": 2516.822, "text": "So, so.", "speaker": "Andrew Pashea"}, {"start": 2517.323, "end": 2521.729, "text": "And I like how the question that Mark had prior to this had to do with, like.", "speaker": "Andrew Pashea"}, {"start": 2522.33, "end": 2524.332, "text": "You know, what is it to make linear assumptions?", "speaker": "Andrew Pashea"}, {"start": 2524.853, "end": 2527.777, "text": "So, in the context of the textbook, what we've seen thus far, like.", "speaker": "Andrew Pashea"}, {"start": 2528.278, "end": 2535.908, "text": "This linear model is essentially 1 of the simplest kinds of models that we would ever find in statistics machine learning or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2535.948, "end": 2539.774, "text": "And it's just to show how a model can be composed.", "speaker": "Andrew Pashea"}, {"start": 2540.755, "end": 2542.337, "text": "Like, how can.", "speaker": "Andrew Pashea"}, {"start": 2542.57, "end": 2551.986, "text": " In the context we're talking about now, like an agent such as a person, how can it receive sensory information, update its beliefs, right?", "speaker": "Andrew Pashea"}, {"start": 2552.407, "end": 2560.16, "text": "It's not until part two that we'll also see not only is it going to update its beliefs, but then also choose to do things, right?", "speaker": "Andrew Pashea"}, {"start": 2560.18, "end": 2561.422, "text": "So that's the action part.", "speaker": "Andrew Pashea"}, {"start": 2561.783, "end": 2563.245, "text": "We haven't reached that yet.", "speaker": "Andrew Pashea"}, {"start": 2563.225, "end": 2572.756, "text": " But that said, the general idea here is that we're just building up sort of from scratch from simpler examples, how does an agent sort of function?", "speaker": "Andrew Pashea"}, {"start": 2573.317, "end": 2590.597, "text": "So the trick with social science and any kind of science that tries to do things like computational modeling is that you're going to have to determine whenever you model something, like what are the variables or the factors that are involved, right?", "speaker": "Andrew Pashea"}, {"start": 2591.198, "end": 2593.08, "text": "So that linear model we were looking at", "speaker": "Andrew Pashea"}, {"start": 2593.06, "end": 2616.688, "text": " we have a beta zero and a beta one and together with the hidden state we end up with this like line and that's what the model looks like so those are two variables that we include in our model it's beta one and beta zero and we're thinking about a person and we want to use a little bit more colloquial or everyday terms we could say like well if i'm interacting with my environment including a whole community of people", "speaker": "Andrew Pashea"}, {"start": 2616.938, "end": 2619.842, "text": " what are the kinds of sensory information that I receive?", "speaker": "Andrew Pashea"}, {"start": 2620.522, "end": 2622.965, "text": "And what are the kinds of actions that I can do?", "speaker": "Andrew Pashea"}, {"start": 2623.005, "end": 2625.308, "text": "And what do I have beliefs about?", "speaker": "Andrew Pashea"}, {"start": 2625.408, "end": 2629.173, "text": "Or what variables do I think explains everything, right?", "speaker": "Andrew Pashea"}, {"start": 2629.774, "end": 2640.887, "text": "So those are really important questions because you could imagine very quickly how much the number of variables you include in a model, especially whenever we look at this sort of like MESO or macro level of entire human", "speaker": "Andrew Pashea"}, {"start": 2640.867, "end": 2655.518, "text": " uh societies or cultures you started with this question about you know some more fundamental questions about um um you know beliefs about the the universe and what defines it and things like that spirituality or otherwise um yeah like", "speaker": "Andrew Pashea"}, {"start": 2656.021, "end": 2658.344, "text": " There's, there's a lot to track.", "speaker": "Andrew Pashea"}, {"start": 2658.364, "end": 2662.21, "text": "So there are some people who have tried to model this sort of thing.", "speaker": "Andrew Pashea"}, {"start": 2662.23, "end": 2663.572, "text": "And so I've shared this paper.", "speaker": "Andrew Pashea"}, {"start": 2664.313, "end": 2672.665, "text": "It's a few years old now, but it was probably 1 of the best called epistemic communities after, excuse me, under active inference.", "speaker": "Andrew Pashea"}, {"start": 2673.226, "end": 2676.19, "text": "And this kind of has to do with.", "speaker": "Andrew Pashea"}, {"start": 2676.507, "end": 2693.875, "text": " Because they actually ran a simulation, like they didn't just do a theoretical thing where they not disparaged purely theoretical papers, but like they actually did a simulation and designed agents they thought matched a lot of literature that we find in psychology and anthropology and elsewhere.", "speaker": "Andrew Pashea"}, {"start": 2693.855, "end": 2711.429, "text": " And so what happens is that the agents all can have contradicting beliefs from each other, whether you phrased it as something about a particular presumably monotheistic God versus some kind of science that says there is no such thing.", "speaker": "Andrew Pashea"}, {"start": 2711.489, "end": 2714.715, "text": "And that's one way of trying to look at a problem.", "speaker": "Andrew Pashea"}, {"start": 2714.695, "end": 2734.948, "text": " like that so it's like idea one and idea two and the assumptions that idea one and idea two conflict with epistemic communities the main takeaway is that we do end up seeing through the simulation they run this kind of polarization of communities where you have a lot of agents who rally around one idea one", "speaker": "Andrew Pashea"}, {"start": 2735.333, "end": 2737.86, "text": " And a lot of agents who rally around idea two.", "speaker": "Andrew Pashea"}, {"start": 2737.9, "end": 2743.315, "text": "And as they do that, the two groups start to communicate directly with each other less and less.", "speaker": "Andrew Pashea"}, {"start": 2743.997, "end": 2749.613, "text": "The agent's actual actions that are simulated has to do with them communicating with one another and sharing their", "speaker": "Andrew Pashea"}, {"start": 2750.015, "end": 2753.178, "text": " They're kind of more verbal beliefs with one another.", "speaker": "Andrew Pashea"}, {"start": 2753.199, "end": 2764.771, "text": "And so it's interesting because, you know, as opposed to saying like, oh, humans, of course, they will do the Bayesian optimal rational thing.", "speaker": "Andrew Pashea"}, {"start": 2764.791, "end": 2769.256, "text": "And somehow that gets all blended in with our assumptions of how science works and things like that.", "speaker": "Andrew Pashea"}, {"start": 2769.757, "end": 2771.338, "text": "It's much more complex.", "speaker": "Andrew Pashea"}, {"start": 2771.479, "end": 2779.988, "text": "And so one of active inference's answers to like, why do we see these kinds of dynamics is that we're trying to minimize free energy", "speaker": "Andrew Pashea"}, {"start": 2779.968, "end": 2783.633, "text": " But remember that free energy, based on how we've discussed it so far,", "speaker": "Andrew Pashea"}, {"start": 2783.967, "end": 2787.932, "text": " It's not some magical quantity of energy or something like that.", "speaker": "Andrew Pashea"}, {"start": 2788.914, "end": 2790.736, "text": "It's really just a proxy.", "speaker": "Andrew Pashea"}, {"start": 2790.796, "end": 2797.485, "text": "It's a word that is a proxy for something called surprisal, which is more or less uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2797.625, "end": 2799.608, "text": "So we want to minimize our uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2800.169, "end": 2813.687, "text": "So if you imagine a group, that unit, that more macro unit of 30 plus agents communicating with one another, we could say that, well, if they all repeatedly agree with one another,", "speaker": "Andrew Pashea"}, {"start": 2813.667, "end": 2817.235, "text": " about is it idea one or is it idea two?", "speaker": "Andrew Pashea"}, {"start": 2817.716, "end": 2819.781, "text": "Is it science or is it something else?", "speaker": "Andrew Pashea"}, {"start": 2820.443, "end": 2824.993, "text": "One way to minimize uncertainty is to repeatedly tell each other", "speaker": "Andrew Pashea"}, {"start": 2825.328, "end": 2828.993, "text": " what the truth is and you all converge on what you all think the truth is.", "speaker": "Andrew Pashea"}, {"start": 2829.073, "end": 2829.293, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 2829.834, "end": 2838.545, "text": "And then if your truth diverges from another community's truth, then you very well might start to like stop interacting with each other as much.", "speaker": "Andrew Pashea"}, {"start": 2839.085, "end": 2841.929, "text": "Because the other group is adding to your uncertainty.", "speaker": "Andrew Pashea"}, {"start": 2842.129, "end": 2849.338, "text": "Well, I thought it was science, but the people over here say it's not, and I don't know what to believe, but my community believes in science and that's what I believe.", "speaker": "Andrew Pashea"}, {"start": 2849.418, "end": 2854.845, "text": "So, so you see that there, there is a kind of like logic where the free energy principle is involved.", "speaker": "Andrew Pashea"}, {"start": 2855.129, "end": 2864.026, "text": " where we do have this kind of rallying and polarization around particular opinions or beliefs or otherwise.", "speaker": "Andrew Pashea"}, {"start": 2864.467, "end": 2870.679, "text": "And that has nothing to do with the true validity of science, right?", "speaker": "Andrew Pashea"}, {"start": 2871.641, "end": 2875.348, "text": "Like I'm attempting to be a scientist explaining all this stuff,", "speaker": "Andrew Pashea"}, {"start": 2875.683, "end": 2878.748, "text": " But the people who, you see what I'm saying?", "speaker": "Andrew Pashea"}, {"start": 2878.788, "end": 2884.538, "text": "Like to take in sensory information is what we do.", "speaker": "Andrew Pashea"}, {"start": 2884.578, "end": 2888.544, "text": "It's not about like following proper logic and stuff.", "speaker": "Andrew Pashea"}, {"start": 2888.564, "end": 2892.19, "text": "It's people learn to do that, right?", "speaker": "Andrew Pashea"}, {"start": 2892.17, "end": 2894.737, "text": " Can I interject something real quick?", "speaker": "Magdalena Hurtado"}, {"start": 2894.757, "end": 2906.128, "text": "I know we don't have a lot of time, but can I just because I think this is a subtle but important nuance point, which is that what you're", "speaker": "Magdalena Hurtado"}, {"start": 2906.581, "end": 2923.723, "text": " Every individual in a group is not going through the active inferencing process of trying to attempt to validate whether or not they have enough evidence to update their prior about whether it's God or quantum physics that explains reality.", "speaker": "Magdalena Hurtado"}, {"start": 2924.364, "end": 2931.593, "text": "So what happens in humans groups is that the active inferencing that's happening in a lot of individuals,", "speaker": "Magdalena Hurtado"}, {"start": 2931.573, "end": 2937.991, "text": " is to analyze and update whether or not they believe the person who's giving the message.", "speaker": "Magdalena Hurtado"}, {"start": 2938.552, "end": 2942.864, "text": "It's not about evaluating the validity of the premises.", "speaker": "Magdalena Hurtado"}, {"start": 2943.486, "end": 2947.236, "text": "It's not about the... That gets at a very...", "speaker": "Magdalena Hurtado"}, {"start": 2947.216, "end": 2957.913, "text": " That gets at this notion of precision, one might think, in terms of there's an issue about, okay, there's the literal content of the proposition that's on offer.", "speaker": "Fraser Paterson"}, {"start": 2958.734, "end": 2961.178, "text": "We might imagine an active inference agent doing inference about this.", "speaker": "Fraser Paterson"}, {"start": 2961.759, "end": 2967.688, "text": "But above that is a question, how reliable is this thing that I'm hearing at all?", "speaker": "Fraser Paterson"}, {"start": 2967.668, "end": 2992.039, "text": " and we kind of see we saw the first glimpses of that with respect to this notion of precision in um in predictive coding so that is another problem which is you know there's all kinds of explanations out there there's all kinds of um things you could pay attention to propositional theories or whatever that you can do inference on but there's the further issue of okay well which one of them is relevant which one should i wait as more or less relevant", "speaker": "Fraser Paterson"}, {"start": 2992.019, "end": 3018.583, "text": " uh that's a that's a thorny problem and it gets at this issue of attention and precision um i think at least so very very thorny problem however um in general so yeah what we're not going to see because that is such a difficult problem we're not going to see a lot of talk and discussion about multi-agent active inference really um the book is meant to be the fundamentals of active inference so we're going to get a really solid appreciation", "speaker": "Fraser Paterson"}, {"start": 3018.563, "end": 3022.371, "text": " for what it means for one entity to be an active inference agent.", "speaker": "Fraser Paterson"}, {"start": 3022.391, "end": 3031.651, "text": "Having then understood that, you can then take that and push that entire picture inside the active inference agent or look at relationships between active inference agents.", "speaker": "Fraser Paterson"}, {"start": 3031.671, "end": 3033.595, "text": "But that's not going to be the focus of the book.", "speaker": "Fraser Paterson"}, {"start": 3034.396, "end": 3038.445, "text": "A lot of that is at the cutting edge of active inference research now.", "speaker": "Fraser Paterson"}, {"start": 3039.961, "end": 3044.874, "text": " Thanks for keeping it closer to the textbook by bringing up precision.", "speaker": "Andrew Pashea"}, {"start": 3044.934, "end": 3046.117, "text": "I very much agree with that.", "speaker": "Andrew Pashea"}, {"start": 3046.157, "end": 3052.915, "text": "And it just highlights the, I just want to briefly correct something since this is the not something you said pressure, but.", "speaker": "Andrew Pashea"}, {"start": 3054.459, "end": 3055.422, "text": "This.", "speaker": "Andrew Pashea"}, {"start": 3055.655, "end": 3065.723, "text": " Taking active inference and turning it into a verb and calling it active inferencing and then saying that people are doing it differently is not quite the right way to think about active inference.", "speaker": "Andrew Pashea"}, {"start": 3065.743, "end": 3070.095, "text": "So active inferences is generally setting up these principles.", "speaker": "Andrew Pashea"}, {"start": 3070.115, "end": 3070.997, "text": "It.", "speaker": "Andrew Pashea"}, {"start": 3070.977, "end": 3092.985, "text": " definitely agrees with a lot of modern science you know it's drawing from neuroscience in the fields um and so it would say active inference would say and the free energy principle would say that we're all doing this there's no like i'm doing this and but but you're doing something that isn't um it's rather that our models are different", "speaker": "Andrew Pashea"}, {"start": 3092.965, "end": 3093.386, "text": " Right?", "speaker": "Andrew Pashea"}, {"start": 3093.806, "end": 3097.411, "text": "Like, our variables that we're including in our respective models are different.", "speaker": "Andrew Pashea"}, {"start": 3097.772, "end": 3099.694, "text": "The precisions can definitely differ.", "speaker": "Andrew Pashea"}, {"start": 3099.714, "end": 3106.303, "text": "If I get information from someone in my group, I might have a higher precision and trust what they say more.", "speaker": "Andrew Pashea"}, {"start": 3106.764, "end": 3114.654, "text": "I might hear from another group, you know, other people who I know I already disagree with, so I'm going to just view them as whatever they say is a bunch of noise, right?", "speaker": "Andrew Pashea"}, {"start": 3115.155, "end": 3118.72, "text": "So a lot of the, you know, a lot of the", "speaker": "Andrew Pashea"}, {"start": 3119.054, "end": 3131.018, "text": " political bickering or, you know, the way that people argue with one another where it turns into this kind of messy thing where they just treat, you can think of it as they're treating each other and what they're saying is as noise, right?", "speaker": "Andrew Pashea"}, {"start": 3131.159, "end": 3134.245, "text": "It's like, oh, I just lump them all in with political group", "speaker": "Andrew Pashea"}, {"start": 3134.225, "end": 3156.351, "text": " be and uh all they ever say is noise so and i i already have my own prior beliefs about what that noise is all about and i disagree with it i think it's wrong and i think the premises are wrong and they can think the premises about what you say are wrong too right so it's it's i mean if you learn to do argumentation where you have the word premises available", "speaker": "Andrew Pashea"}, {"start": 3156.331, "end": 3172.112, "text": " then cool but i just i'm just trying to make the point if you want to talk about hunter gatherer societies or otherwise you also have to recognize the very language we use something that gets learned and the way that we get taught to think about different things like oh well you need to evaluate the premises not everyone", "speaker": "Andrew Pashea"}, {"start": 3172.092, "end": 3191.299, "text": " ask that so it's it's very yeah it's uh there's a lot there's a lot to be taken consider of i think the social sciences are a very interesting field right now though to be applying active inference so kudos on kind of thinking through those kinds of questions thank you thank you appreciate it", "speaker": "Andrew Pashea"}, {"start": 3192.308, "end": 3195.736, "text": " Do make sure, I've been absent for a little bit.", "speaker": "Fraser Paterson"}, {"start": 3196.237, "end": 3200.687, "text": "I'm gonna be much more active on the questions side of things on the CODA.", "speaker": "Fraser Paterson"}, {"start": 3200.748, "end": 3205.418, "text": "So do make sure if you've got questions, please do put them down in the questions tab here.", "speaker": "Fraser Paterson"}, {"start": 3205.458, "end": 3209.608, "text": "I've gone through and I've already answered quite a few of them for chapter five.", "speaker": "Fraser Paterson"}, {"start": 3209.588, "end": 3224.188, "text": " uh so coming down here chapter three i think i think they're all answered for chapter five now i'm gonna go back and answer everything that hasn't been answered so if you do have a burning question the best place to put it is um is on the coda i think i'm sharing now you guys can see", "speaker": "Fraser Paterson"}, {"start": 3224.573, "end": 3226.135, "text": " Are there any more live questions?", "speaker": "Fraser Paterson"}, {"start": 3227.777, "end": 3228.718, "text": "That would be nice.", "speaker": "Fraser Paterson"}, {"start": 3228.738, "end": 3233.584, "text": "We can stay around for maybe five or so minutes after the deadline if people so choose.", "speaker": "Fraser Paterson"}, {"start": 3234.645, "end": 3236.267, "text": "But we are coming up to the top of the hour.", "speaker": "Fraser Paterson"}, {"start": 3236.287, "end": 3241.994, "text": "So yeah, all of the questions in chapter five now are answered, or I've given my attempt.", "speaker": "Fraser Paterson"}, {"start": 3242.655, "end": 3246.379, "text": "As I say, I'll go through and attempt to answer all the previous ones as well.", "speaker": "Fraser Paterson"}, {"start": 3249.937, "end": 3276.737, "text": " think i'm gonna have to not do this for part one but going forward chapter six and such like i'm gonna try and create more of these sort of animations uh as to you know what's going on uh hopefully for a bit more um intuitive effect i have a little bit more time to do that now so i did that for chapter four but um yes okay coming back sun sun sun's in i'm not sure how to say your name sorry", "speaker": "Fraser Paterson"}, {"start": 3277.595, "end": 3301.122, "text": " okay uh can you hear me clearly yes yes okay i want to check if my understanding of the active inference process is correct my understanding is the generative model starts with prior beliefs receive observations and then reduces prediction errors through using bayesian", "speaker": "Sun Xin"}, {"start": 3301.102, "end": 3308.391, "text": " inference to engage in parameter learning so that it can update beliefs about the current state.", "speaker": "Sun Xin"}, {"start": 3309.172, "end": 3319.545, "text": "If prediction errors persist, the generative model engages in model selection, that is, to updating the generative model itself.", "speaker": "Sun Xin"}, {"start": 3319.985, "end": 3323.97, "text": "And I think this is what we call plasticity.", "speaker": "Sun Xin"}, {"start": 3323.95, "end": 3331.683, "text": " The ultimate goal of all this is to minimize expected free energy.", "speaker": "Sun Xin"}, {"start": 3332.404, "end": 3333.085, "text": "Is that right?", "speaker": "Sun Xin"}, {"start": 3334.548, "end": 3335.088, "text": "Wow, okay.", "speaker": "Fraser Paterson"}, {"start": 3335.129, "end": 3338.174, "text": "I mean, basically, yeah, that's substantively correct.", "speaker": "Fraser Paterson"}, {"start": 3339.115, "end": 3342.06, "text": "I will note, however, we haven't yet talked about expected free energy.", "speaker": "Fraser Paterson"}, {"start": 3342.395, "end": 3344.237, "text": " and planning and action selection.", "speaker": "Fraser Paterson"}, {"start": 3344.257, "end": 3345.358, "text": "So that is going to come later.", "speaker": "Fraser Paterson"}, {"start": 3346.379, "end": 3346.959, "text": "But yeah, you're right.", "speaker": "Fraser Paterson"}, {"start": 3346.979, "end": 3352.825, "text": "In terms of the general flavor of things, we have beliefs about hidden states in the world.", "speaker": "Fraser Paterson"}, {"start": 3353.686, "end": 3365.076, "text": "And we're going to update those beliefs by means of inference, specifically Bayesian inference, and more specifically, approximate Bayesian inference, where we're doing variational free energy minimization.", "speaker": "Fraser Paterson"}, {"start": 3365.997, "end": 3369.921, "text": "Or we've seen that we can also recast this in terms of predictive processing.", "speaker": "Fraser Paterson"}, {"start": 3370.66, "end": 3372.943, "text": " minimizing the precision way to prediction errors.", "speaker": "Fraser Paterson"}, {"start": 3372.963, "end": 3373.924, "text": "It's the same kind of story.", "speaker": "Fraser Paterson"}, {"start": 3374.926, "end": 3386.221, "text": "And what we need is we need a generative model, probability distribution of the hidden states and observations, which is to say a likelihood about observations and a prior belief about states.", "speaker": "Fraser Paterson"}, {"start": 3387.063, "end": 3387.283, "text": "Yes.", "speaker": "Fraser Paterson"}, {"start": 3388.725, "end": 3391.328, "text": "The other thing, you mentioned parameter learning.", "speaker": "Fraser Paterson"}, {"start": 3391.388, "end": 3392.77, "text": "So yes, exactly.", "speaker": "Fraser Paterson"}, {"start": 3392.79, "end": 3393.932, "text": "There's these two problems.", "speaker": "Fraser Paterson"}, {"start": 3394.032, "end": 3397.317, "text": "We have the problem of inferring the hidden states.", "speaker": "Fraser Paterson"}, {"start": 3397.357, "end": 3398.338, "text": "What should the hidden states be?", "speaker": "Fraser Paterson"}, {"start": 3398.977, "end": 3407.788, "text": " But in order to do that, we have a model which has knobs and dials called parameters, and we need to set those parameters before we do inference.", "speaker": "Fraser Paterson"}, {"start": 3408.529, "end": 3415.017, "text": "So we have kind of two problems, you know, approximate Bayesian inference, variation of free energy, precision weight prediction error, whatever you like.", "speaker": "Fraser Paterson"}, {"start": 3415.738, "end": 3419.183, "text": "But then above that, we need to deal with the problem of what should the setting of the parameters be?", "speaker": "Fraser Paterson"}, {"start": 3420.204, "end": 3426.853, "text": "And we've only really just begun to look at this in terms of prediction errors and hierarchical models.", "speaker": "Fraser Paterson"}, {"start": 3428.034, "end": 3428.815, "text": "But we saw", "speaker": "Fraser Paterson"}, {"start": 3429.149, "end": 3432.913, "text": " how to deal with that in terms of expectation maximization and earlier chapters.", "speaker": "Fraser Paterson"}, {"start": 3433.634, "end": 3443.644, "text": "We're going to be continuing to think about that problem of model learning, sorry, parameter learning and hidden state inference in terms of a hierarchical picture.", "speaker": "Fraser Paterson"}, {"start": 3443.964, "end": 3450.251, "text": "Because we generally can't do expectation maximization for the reason that we don't know the exact posterior.", "speaker": "Fraser Paterson"}, {"start": 3450.291, "end": 3458.299, "text": "So hopefully that didn't make you more confused, but yes, your understanding is substantively correct for sure, yeah.", "speaker": "Fraser Paterson"}, {"start": 3458.339, "end": 3458.599, "text": "Thank you.", "speaker": "Sun Xin"}, {"start": 3459.49, "end": 3461.011, "text": " No problem.", "speaker": "Fraser Paterson"}, {"start": 3461.032, "end": 3462.473, "text": "Do we have more questions?", "speaker": "Fraser Paterson"}, {"start": 3462.633, "end": 3463.254, "text": "More questions?", "speaker": "Fraser Paterson"}, {"start": 3464.655, "end": 3466.857, "text": "Maybe one more question, if there is one more.", "speaker": "Fraser Paterson"}, {"start": 3467.117, "end": 3469.52, "text": "And then I'll stop the recording.", "speaker": "Fraser Paterson"}, {"start": 3475.566, "end": 3475.866, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3475.886, "end": 3476.887, "text": "I might stop the recording here.", "speaker": "Fraser Paterson"}, {"start": 3478.068, "end": 3478.549, "text": "Well, Mark.", "speaker": "Fraser Paterson"}, {"start": 3478.569, "end": 3479.71, "text": "Hello, everyone, and welcome.", "speaker": "Fraser Paterson"}, {"start": 3479.91, "end": 3481.732, "text": "I've got my esteemed friend.", "speaker": "Fraser Paterson"}, {"start": 3482.753, "end": 3484.975, "text": "Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 3486.423, "end": 3489.528, "text": " That was interesting audio feedback.", "speaker": "Marc Broberg"}, {"start": 3489.548, "end": 3492.053, "text": "I thought I would jump in since nobody else asked us.", "speaker": "Marc Broberg"}, {"start": 3493.135, "end": 3499.105, "text": "But earlier you mentioned using gradient descent to learn parameters, right?", "speaker": "Marc Broberg"}, {"start": 3499.245, "end": 3501.289, "text": "Is that also used in perception?", "speaker": "Marc Broberg"}, {"start": 3501.69, "end": 3503.893, "text": "Is it the same approach?", "speaker": "Marc Broberg"}, {"start": 3503.994, "end": 3508.581, "text": "Yeah, I mean, gradient descent is used absolutely everywhere for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3508.602, "end": 3510.625, "text": "So it's a very, very general", "speaker": "Fraser Paterson"}, {"start": 3511.06, "end": 3512.522, "text": " optimization techniques.", "speaker": "Fraser Paterson"}, {"start": 3512.602, "end": 3520.19, "text": "So yes, it certainly can be used for learning, well, for perception.", "speaker": "Fraser Paterson"}, {"start": 3520.711, "end": 3522.513, "text": "I'm trying to find a nice picture here.", "speaker": "Fraser Paterson"}, {"start": 3523.173, "end": 3531.983, "text": "Typically, I guess going forward into like chapter six and so on, we're not going to spend a lot of time immediately on gradient descent.", "speaker": "Fraser Paterson"}, {"start": 3532.524, "end": 3533.445, "text": "Let me just make sure of that.", "speaker": "Fraser Paterson"}, {"start": 3533.565, "end": 3535.107, "text": "Yes, in other words, is the answer.", "speaker": "Fraser Paterson"}, {"start": 3535.547, "end": 3537.049, "text": "It's used for a lot of things.", "speaker": "Fraser Paterson"}, {"start": 3538.39, "end": 3539.852, "text": "Very, very general technique.", "speaker": "Fraser Paterson"}, {"start": 3540.237, "end": 3549.91, "text": " Um, although in, in, in a lot of the problems that we're gonna see later on, um, the kinds of, uh, yeah, no, we are gonna see in chapter six.", "speaker": "Fraser Paterson"}, {"start": 3549.93, "end": 3550.39, "text": "Absolutely.", "speaker": "Fraser Paterson"}, {"start": 3550.51, "end": 3550.891, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3550.911, "end": 3552.593, "text": "We're gonna talk about phase planes and so on.", "speaker": "Fraser Paterson"}, {"start": 3553.394, "end": 3553.635, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3553.955, "end": 3556.258, "text": "There are problems with it, but yes, we're gonna see it going forward.", "speaker": "Fraser Paterson"}, {"start": 3556.278, "end": 3559.242, "text": "It's a very, it's a very foundational technique.", "speaker": "Fraser Paterson"}, {"start": 3559.382, "end": 3562.807, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3562.827, "end": 3563.027, "text": "All right.", "speaker": "Fraser Paterson"}, {"start": 3563.047, "end": 3566.211, "text": "Any, any last minute questions for the YouTube recording before we, uh,", "speaker": "Fraser Paterson"}, {"start": 3569.093, "end": 3570.975, "text": " If not, I think I'm sorry.", "speaker": "Fraser Paterson"}, {"start": 3571.236, "end": 3572.377, "text": "Can I ask a quick question?", "speaker": "Dorsa"}, {"start": 3572.437, "end": 3574.2, "text": "Yeah, sure.", "speaker": "Fraser Paterson"}, {"start": 3574.74, "end": 3575.461, "text": "I do apologize.", "speaker": "Dorsa"}, {"start": 3575.702, "end": 3577.424, "text": "I have joined the group very late.", "speaker": "Dorsa"}, {"start": 3577.484, "end": 3579.767, "text": "So maybe this was addressed in like the past weeks.", "speaker": "Dorsa"}, {"start": 3579.827, "end": 3593.245, "text": "But are there any known like general equivalence or say convergence results between active inference and say just reinforcement learning?", "speaker": "Dorsa"}, {"start": 3593.478, "end": 3619.553, "text": " well uh there are known relationships if you go to appendix a um with the very first session was appendix a there's lots of very helpful uh discussion there about the relationship between active inference and reinforcement learning so yes there is there's quite a bit in terms of what you can do is you can look at the variational free energy or the expected free energy and active inference and you can see okay with all these terms", "speaker": "Fraser Paterson"}, {"start": 3619.955, "end": 3629.547, "text": " If I get rid of one of these terms or another one of these terms, I end up with KL control, or I end up with reinforcement learning, or I end up with risk-sensitive control.", "speaker": "Fraser Paterson"}, {"start": 3630.147, "end": 3643.784, "text": "There's kind of the idea that is forming that active inference is a very general way of talking about all of these things, and that things like reinforcement learning, things like KL control, like risk-sensitive control, these are kind of special cases of active inference.", "speaker": "Fraser Paterson"}, {"start": 3643.804, "end": 3646.988, "text": "So the answer is yes, there is a profound relationship.", "speaker": "Fraser Paterson"}, {"start": 3647.008, "end": 3649.371, "text": "We probably don't have time to get into it here.", "speaker": "Fraser Paterson"}, {"start": 3649.84, "end": 3659.489, "text": " One of the probably most pertinent differences between reinforcement learning and active inference is this idea of information gain.", "speaker": "Fraser Paterson"}, {"start": 3660.41, "end": 3665.735, "text": "Because we've seen with, well, we haven't yet seen with expected free energy how that works.", "speaker": "Fraser Paterson"}, {"start": 3665.755, "end": 3672.461, "text": "But very broadly, the idea with active inference is that we're modeling uncertainties from the beginning.", "speaker": "Fraser Paterson"}, {"start": 3673.062, "end": 3679.628, "text": "And we don't just have this scale of reward, this signal in reinforcement learning.", "speaker": "Fraser Paterson"}, {"start": 3680.114, "end": 3708.67, "text": " we're explicitly reasoning about uncertainties all the time and where we're able to account for information gain so it's probably not helped you too much but yeah there is a lot a lot there maybe as part of uh chapter five or in some cemetery summative section i'll put some stuff because there's a lot of existing stuff on the relationship between active infants and reinforcement learning yeah yeah thank you so much thank you no problem yeah just sort of like a like a um", "speaker": "Fraser Paterson"}, {"start": 3708.97, "end": 3717.107, "text": " a tale to what Frasier had said, just because I've also given some talks on the relationship with reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3717.147, "end": 3727.688, "text": "Yeah, the information gain part is very important, because for those who are used to thinking about reinforcement learning,", "speaker": "Andrew Pashea"}, {"start": 3727.668, "end": 3741.341, "text": " Of course, there are many ways to do reinforcement learning, but one common one is that whenever agents do actions, which again, like Fraser said, we're going to look more at action-based agents, agents who can act in part two.", "speaker": "Andrew Pashea"}, {"start": 3742.282, "end": 3752.331, "text": "But the big thing is that reinforcement learning, whenever the agents like infer, whenever they do those sorts of things, they're typically like reward driven.", "speaker": "Andrew Pashea"}, {"start": 3753.192, "end": 3756.595, "text": "So there can be like a KL control agent or there can be a variety.", "speaker": "Andrew Pashea"}, {"start": 3756.575, "end": 3779.157, "text": " other kinds of agents um so usually they they tend to be a little bit more um i don't want to use the word greedy but something like that like more reward focused um they look a lot more like the kind of like proverbial agent we would find in like economics or something um meanwhile in active inference um it would say like oh the way that", "speaker": "Andrew Pashea"}, {"start": 3779.137, "end": 3782.622, "text": " that things like curiosity exist.", "speaker": "Andrew Pashea"}, {"start": 3782.742, "end": 3787.429, "text": "Why does curiosity exist if we're actually always just driven towards a reward?", "speaker": "Andrew Pashea"}, {"start": 3787.509, "end": 3790.173, "text": "Once you know what the reward is, you should just go for that, right?", "speaker": "Andrew Pashea"}, {"start": 3790.213, "end": 3796.262, "text": "And you would have no reason to find some other strategy for going for it.", "speaker": "Andrew Pashea"}, {"start": 3796.463, "end": 3798.466, "text": "You would have no reason to go out of your way.", "speaker": "Andrew Pashea"}, {"start": 3798.606, "end": 3803.032, "text": "From what you already know, you just keep going for the reward as much as possible.", "speaker": "Andrew Pashea"}, {"start": 3803.393, "end": 3808.941, "text": "Perhaps you learn other things through happenstance along the way.", "speaker": "Andrew Pashea"}, {"start": 3809.258, "end": 3811.041, "text": " That's a big difference.", "speaker": "Fraser Paterson"}, {"start": 3811.722, "end": 3814.948, "text": "In reinforcement, you can just sort of start maximizing a reward signal.", "speaker": "Fraser Paterson"}, {"start": 3815.569, "end": 3822.601, "text": "But in active inference and in the, dare I say, in the real world, oftentimes you don't know how to just start maximizing rewards.", "speaker": "Fraser Paterson"}, {"start": 3822.621, "end": 3828.972, "text": "You need to resolve your uncertainty about how to start maximizing reward and then maximize rewards.", "speaker": "Fraser Paterson"}, {"start": 3828.952, "end": 3856.731, "text": " yeah exactly exactly and so what one common thing in reinforcement learning to try and like allow the agent to do something different rather than be so purely reward driven and of course there have been a lot of advances past this but usually it boils down to some kind of ad hoc rule like epsilon greedy framework which says okay go for the reward 90 of the time but 10 of the time do something random um not quite", "speaker": "Andrew Pashea"}, {"start": 3857.234, "end": 3868.402, "text": " how we work as far as I'm aware, what's found in the empirical literature, but it's been one way of resolving that kind of fixity on reward issue and reinforcement learning.", "speaker": "Andrew Pashea"}, {"start": 3868.743, "end": 3872.492, "text": "Meanwhile, active inference has a much more principled way that relates to this notion of", "speaker": "Andrew Pashea"}, {"start": 3872.472, "end": 3892.793, "text": " free energy and specifically expected free energy such that the agent will take it will find value in learning new things and exploring new things it will still maintain reference to what is rewarding to itself so it's not a random uh information gain it's like oh you know I usually", "speaker": "Andrew Pashea"}, {"start": 3893.06, "end": 3896.992, "text": " Um, when I play a sport, I throw the ball this way.", "speaker": "Andrew Pashea"}, {"start": 3897.012, "end": 3906.681, "text": "Uh, but what happens if I still try to throw the ball as I would so that I can still like accomplish the goal of like, uh, you know, whatever, throwing it to the other person.", "speaker": "Andrew Pashea"}, {"start": 3906.897, "end": 3909.963, "text": " but maybe I will kind of curve it or change it, right?", "speaker": "Andrew Pashea"}, {"start": 3910.343, "end": 3914.03, "text": "You're not randomly throwing it in the sky or in the opposite direction.", "speaker": "Andrew Pashea"}, {"start": 3914.431, "end": 3919.481, "text": "You're still trying to throw it to where you want to, but maybe you'll change your technique a bit, right?", "speaker": "Andrew Pashea"}, {"start": 3919.541, "end": 3927.235, "text": "There's a more kind of, you know, there's a sort of knowingness with respect to one's own model and how to make that model better.", "speaker": "Andrew Pashea"}, {"start": 3927.552, "end": 3953.037, "text": " based on things that you haven't explored yet or things that you can so so it's just much more involved it's much more principled and it's the kind of thing that if you produce this model in a concrete fashion you could even look at the time series of like how things change over time and figure out where it learned such and such and why and what was going on it's in its beliefs at that time as opposed to just being a black box model you know it's not all about", "speaker": "Andrew Pashea"}, {"start": 3953.523, "end": 3956.968, "text": " How do I perform the best whenever it comes to active inference?", "speaker": "Andrew Pashea"}, {"start": 3957.129, "end": 3965.422, "text": "It's not just about like, otherwise we could just make another deep neural network and, you know, 8 billion parameters and not know how to interpret any of them.", "speaker": "Andrew Pashea"}, {"start": 3966.223, "end": 3975.918, "text": "But active inference very much has to do with actually being able to interpret and understand the components of the model as beliefs that the agent has, and then they need to be exposed and clear.", "speaker": "Andrew Pashea"}, {"start": 3976.038, "end": 3977.32, "text": "Yeah.", "speaker": "Fraser Paterson"}, {"start": 3978.194, "end": 3983.299, "text": " I think we have time for Giancuomo and then Mark, and then we're probably going to have to call it there, guys.", "speaker": "Fraser Paterson"}, {"start": 3983.319, "end": 3986.502, "text": "So Giancuomo, fire away.", "speaker": "Fraser Paterson"}, {"start": 3986.522, "end": 3991.827, "text": "Very quickly, I think it ties in with what was just spoken and talked about now.", "speaker": "Giacomo Bruzzo"}, {"start": 3992.648, "end": 4007.982, "text": "Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is?", "speaker": "Giacomo Bruzzo"}, {"start": 4008.265, "end": 4033.468, "text": " Okay, can the agent quantify, and they seem to get a sense that they can through, there's an entropy term that maybe gives me an idea that somehow he could have like a confidence interval and say, I will say this, because if I look at the reward, the reinforcement learning machine that has been calibrated for learning, they tend to give you absolute certainty.", "speaker": "Giacomo Bruzzo"}, {"start": 4033.508, "end": 4035.029, "text": "They say, this is the answer.", "speaker": "Giacomo Bruzzo"}, {"start": 4035.109, "end": 4037.091, "text": "And you're like, no, no, no, it's not.", "speaker": "Giacomo Bruzzo"}, {"start": 4037.594, "end": 4057.919, "text": " And is there a sense in which it is a bit more nuanced, that it takes care that in a way that the path that he chooses through this very complex, high dimensional space of possibilities is actually more economical in the end, maybe slower, but more self-aware.", "speaker": "Giacomo Bruzzo"}, {"start": 4057.959, "end": 4064.007, "text": "Sorry if I'm using all the wrong words, but is there some kind of intuition like that going behind?", "speaker": "Giacomo Bruzzo"}, {"start": 4065.455, "end": 4093.251, "text": " well if i understand your question correctly uh it's is the agent itself able to offer a quantity to quantify its uncertainty in its predictions yes i mean the mere fact that we're uh so if i come here the mere fact that what we're doing from the beginning from the very beginning is we're reasoning about probability distributions", "speaker": "Fraser Paterson"}, {"start": 4093.687, "end": 4094.989, "text": " This is built in from the start.", "speaker": "Fraser Paterson"}, {"start": 4095.029, "end": 4103.639, "text": "So from the very beginning, we are reasoning about uncertainty because everything that we do is guided by the north star of Bayes' rule, right?", "speaker": "Fraser Paterson"}, {"start": 4104.22, "end": 4105.982, "text": "And okay, we can't do Bayes' rule exactly.", "speaker": "Fraser Paterson"}, {"start": 4106.022, "end": 4107.023, "text": "We have to do it approximately.", "speaker": "Fraser Paterson"}, {"start": 4107.083, "end": 4108.365, "text": "We do variational inference.", "speaker": "Fraser Paterson"}, {"start": 4108.405, "end": 4110.407, "text": "We do prediction error and minimization, that kind of thing.", "speaker": "Fraser Paterson"}, {"start": 4110.828, "end": 4114.052, "text": "But we're always, always, always reasoning about probability distribution.", "speaker": "Fraser Paterson"}, {"start": 4114.072, "end": 4118.597, "text": "So yes, every single active implementation ever is always", "speaker": "Fraser Paterson"}, {"start": 4120.535, "end": 4147.757, "text": " predictions because that's literally what it is to do active inference so that's the easy answer the more tricky answer is that when it comes to doing planning and action selection there's issues about okay in the future i need to plan stuff i need to think about what i'm going to do you know 10 time steps from now that have not happened um i need some way to reason about my uncertainty about things that haven't even happened yet", "speaker": "Fraser Paterson"}, {"start": 4147.737, "end": 4149.84, "text": " And there's issues around that and how to do that.", "speaker": "Fraser Paterson"}, {"start": 4150.821, "end": 4151.722, "text": "But we haven't seen that just yet.", "speaker": "Fraser Paterson"}, {"start": 4151.742, "end": 4153.024, "text": "So, yes, absolutely.", "speaker": "Fraser Paterson"}, {"start": 4153.084, "end": 4155.808, "text": "It's part and parcel of what it is to be an active infatuation.", "speaker": "Fraser Paterson"}, {"start": 4155.828, "end": 4157.15, "text": "It's the reason about uncertainty now.", "speaker": "Fraser Paterson"}, {"start": 4158.411, "end": 4158.932, "text": "Okay, thank you.", "speaker": "Fraser Paterson"}, {"start": 4159.613, "end": 4159.913, "text": "No problem.", "speaker": "Fraser Paterson"}, {"start": 4159.933, "end": 4162.036, "text": "And then, Mark, and then I think we can do... Oh, hang on.", "speaker": "Fraser Paterson"}, {"start": 4162.337, "end": 4167.163, "text": "Maybe just really quickly, did you have a comment, Andrew?", "speaker": "Fraser Paterson"}, {"start": 4167.413, "end": 4171.819, "text": " I also really have to go, but yeah, just briefly, you basically answered it.", "speaker": "Andrew Pashea"}, {"start": 4171.839, "end": 4173.622, "text": "Sorry, I was looking at the chat.", "speaker": "Andrew Pashea"}, {"start": 4173.642, "end": 4177.928, "text": "It's just, yeah, it depends on how the question is being asked.", "speaker": "Andrew Pashea"}, {"start": 4177.948, "end": 4185.339, "text": "Like if you want to go the full nine yards of like, oh, I'm imagining a person and can the person say like how uncertain they are?", "speaker": "Andrew Pashea"}, {"start": 4185.459, "end": 4191.348, "text": "Like that's going to take a little bit more than just like only looking at the simplified models.", "speaker": "Andrew Pashea"}, {"start": 4191.368, "end": 4193.05, "text": "We're looking at the textbook here.", "speaker": "Andrew Pashea"}, {"start": 4193.03, "end": 4216.928, "text": " um you know but as far as like um yeah as far as the models we've been looking at it's like they necessarily whenever we're looking at a probabilistic framework you can say oh it's you know in a categorical distribution uh for a for a fair coin it's like well it's 50 50 you know heads versus tails that it will land on right so that necessarily is a kind of uncertainty about what the real realization", "speaker": "Andrew Pashea"}, {"start": 4216.908, "end": 4230.176, "text": " of that kind of hidden state of the world is like, will it end up being heads or tails and, you know, 50 50 and so you can look at sort of like an entropy term on that categorical distribution is like, well, that's maximum entropy of 2 slots.", "speaker": "Andrew Pashea"}, {"start": 4230.256, "end": 4232.18, "text": "There are 2 possible things.", "speaker": "Andrew Pashea"}, {"start": 4232.22, "end": 4235.507, "text": "It could be is completely 50 50 is fully uncertain.", "speaker": "Andrew Pashea"}, {"start": 4235.487, "end": 4246.867, "text": " Uh, right so we necessarily have that and then the notions of uncertainty are also sort of baked into, uh, the, the, the, like, updating of prediction errors and precision.", "speaker": "Andrew Pashea"}, {"start": 4247.388, "end": 4247.609, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4247.709, "end": 4255.423, "text": "Because whenever you have precision, that's kind of like a gain, you know, kind of a, a, a, almost like a volume knob.", "speaker": "Andrew Pashea"}, {"start": 4255.803, "end": 4256.044, "text": "Right?", "speaker": "Andrew Pashea"}, {"start": 4256.104, "end": 4258.508, "text": "The more you turn up precision, the more you're.", "speaker": "Andrew Pashea"}, {"start": 4258.488, "end": 4263.134, "text": " you're going to take into account the prediction errors you're receiving and vice versa.", "speaker": "Andrew Pashea"}, {"start": 4263.214, "end": 4275.609, "text": "So that has to do with sort of the degree of trust or uncertainty around trusting, you know, some particular belief that you have or some particular sensory observation that you're receiving.", "speaker": "Andrew Pashea"}, {"start": 4275.669, "end": 4275.89, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4275.91, "end": 4279.294, "text": "So so uncertainty is very much like throughout.", "speaker": "Andrew Pashea"}, {"start": 4279.474, "end": 4284.18, "text": "I mean, it's a very ubiquitous term through many parts of active inference.", "speaker": "Andrew Pashea"}, {"start": 4284.2, "end": 4284.34, "text": "Yeah.", "speaker": "Andrew Pashea"}, {"start": 4284.48, "end": 4285.742, "text": "So it's essential.", "speaker": "Andrew Pashea"}, {"start": 4287.224, "end": 4288.225, "text": "So all right.", "speaker": "Jim DeLong"}, {"start": 4288.677, "end": 4311.58, "text": " so with precision with precision then could i uh turn it way up and watch the model uh uh believe it's watch the agent believe its model more than it believes its size yeah absolutely yeah and then vice versa i know my model's messed up just believe what you see", "speaker": "Jim DeLong"}, {"start": 4312.673, "end": 4337.179, "text": " right yeah it's a great thing and so that part of the it highlights the role of like why do we care about priors or something right like why you know there are many other models uh both you know theoretical models in neuroscience as well as machine learning and elsewhere where it's like we could just have a likelihood right we could just take in sensory information and update a belief and uh", "speaker": "Andrew Pashea"}, {"start": 4337.547, "end": 4338.789, "text": " who cares about priors.", "speaker": "Andrew Pashea"}, {"start": 4338.829, "end": 4342.634, "text": "And that's what we see with maximum likelihood estimation in chapter two there.", "speaker": "Andrew Pashea"}, {"start": 4343.655, "end": 4354.31, "text": "But the thing is like, you know, someone who fully believes their eyes and doesn't believe they're up, doesn't have any confidence in their own beliefs about priors or something.", "speaker": "Andrew Pashea"}, {"start": 4354.35, "end": 4366.086, "text": "It's like, if, you know, if I, you know, wake up in the middle of the night and it's dark and I swear, I saw a person in my room or something when in fact, maybe I was just waking up from a dream and just kind of like,", "speaker": "Andrew Pashea"}, {"start": 4366.403, "end": 4392.141, "text": " thought i saw something right like you'd be hyper reactive to the sensory information you receive you would not there would be no kind of stability from a prior belief that keeps you a little bit more grounded where that prior can be updated right it's not that you are born with a prior and it stays with you the whole life it's like that's why we have chapter three on learning where the prior itself can be learned just as the likelihood meaning how uh observations and hidden states um um", "speaker": "Andrew Pashea"}, {"start": 4392.661, "end": 4394.263, "text": " you know, how those connect up with each other.", "speaker": "Andrew Pashea"}, {"start": 4394.704, "end": 4403.396, "text": "So, so that's the significance of like, well, if I, on the other hand, if I overly believe my prior, then, then I'm just kind of stuck there.", "speaker": "Andrew Pashea"}, {"start": 4403.476, "end": 4413.409, "text": "And any information I receive, uh, will always be, you know, either it's contradictory to what I believe and therefore I don't trust it or it fully confirms what I already believe.", "speaker": "Andrew Pashea"}, {"start": 4413.55, "end": 4416.193, "text": "And so I fully trust it without question.", "speaker": "Andrew Pashea"}, {"start": 4416.654, "end": 4416.854, "text": "Right.", "speaker": "Andrew Pashea"}, {"start": 4416.874, "end": 4419.057, "text": "So I say all these things because, uh,", "speaker": "Andrew Pashea"}, {"start": 4419.037, "end": 4426.085, "text": " Part of what brought me to active inference was this stuff about, you know, how people communicate with one another and how they believe what they believe.", "speaker": "Andrew Pashea"}, {"start": 4426.666, "end": 4442.303, "text": "And then furthermore, how does this actually show in like psychiatry and psychology whenever it comes to people who believe a hallucination that they're having or people who like are kind of biased towards others in a particular way and all those sorts of things.", "speaker": "Andrew Pashea"}, {"start": 4442.343, "end": 4444.065, "text": "Yeah, it's very interesting to think about.", "speaker": "Andrew Pashea"}, {"start": 4445.294, "end": 4449.623, "text": " Yeah, do take a look at, I think I'm sharing, but figure 2.11.", "speaker": "Fraser Paterson"}, {"start": 4450.525, "end": 4459.965, "text": "This literally shows the effects of updating, well, the effect that prior precision has on the updating of belief.", "speaker": "Fraser Paterson"}, {"start": 4459.985, "end": 4462.29, "text": "So if the precision is really tight around the prior,", "speaker": "Fraser Paterson"}, {"start": 4462.692, "end": 4472.808, "text": " the belief hasn't really changed much, but if I have a kind of lax prior, not very precise, you can see that the update is dominated by the evidence coming in from my likelihood model.", "speaker": "Fraser Paterson"}, {"start": 4472.828, "end": 4474.511, "text": "So yeah, do take a look at that.", "speaker": "Fraser Paterson"}, {"start": 4474.531, "end": 4482.584, "text": "That's probably a very helpful motivating thing when it comes to the effect that the prior precision can have.", "speaker": "Fraser Paterson"}, {"start": 4482.684, "end": 4483.786, "text": "Yeah.", "speaker": "Jim DeLong"}, {"start": 4483.806, "end": 4484.146, "text": "Excellent.", "speaker": "Jim DeLong"}, {"start": 4484.206, "end": 4484.667, "text": "Thank you.", "speaker": "Jim DeLong"}, {"start": 4485.372, "end": 4485.512, "text": " Cool.", "speaker": "Fraser Paterson"}, {"start": 4485.532, "end": 4487.255, "text": "All right, guys, we're going to have to call it there.", "speaker": "Fraser Paterson"}, {"start": 4487.756, "end": 4491.963, "text": "Do put questions in the chats on the page, the code page.", "speaker": "Fraser Paterson"}, {"start": 4492.003, "end": 4493.606, "text": "I'll be a bit more attentive going forward.", "speaker": "Fraser Paterson"}, {"start": 4494.267, "end": 4495.489, "text": "I look forward to next week.", "speaker": "Fraser Paterson"}, {"start": 4495.509, "end": 4499.796, "text": "I'll be doing another session of this kind on Friday for the people who usually attend that session.", "speaker": "Fraser Paterson"}, {"start": 4499.816, "end": 4500.557, "text": "So thank you very much.", "speaker": "Fraser Paterson"}, {"start": 4500.918, "end": 4502.22, "text": "Stop sharing and stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4503.522, "end": 4504.704, "text": "Have we got a recording here?", "speaker": "Fraser Paterson"}, {"start": 4506.247, "end": 4507.269, "text": "Okay, stop the recording.", "speaker": "Fraser Paterson"}, {"start": 4507.749, "end": 4508.551, "text": "Goodbye, YouTube people.", "speaker": "Fraser Paterson"}, {"start": 4508.751, "end": 4509.372, "text": "Until next time.", "speaker": "Fraser Paterson"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt index 9d0693e97..5e56a2086 100644 --- a/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt +++ b/data/video/activeinferenceinstitute/TextbookGroup/Namjoshi2026/Cohort_1/Session_023/transcript.txt @@ -1320,7 +1320,7 @@ I think we have time for Giancuomo and then Mark, and then we're probably going So Giancuomo, fire away. -SPEAKER_00: +Giacomo Bruzzo: Very quickly, I think it ties in with what was just spoken and talked about now. Is there any sense in which in active interest an agent has a sense of how short-sighted it is or they are or whatever the pronoun of the agent is? From 0f8f597caf8e2dff874b3627d2cf12d4e78b378b Mon Sep 17 00:00:00 2001 From: Holly Grimm Date: Fri, 17 Jul 2026 16:44:08 -0600 Subject: [PATCH 4/4] =?UTF-8?q?data:=20GuestStream=5F128=20SPEAKER=5F00=20?= =?UTF-8?q?=3D=20Daniel=20Friedman=20=E2=80=94=20all=20speakers=20identifi?= =?UTF-8?q?ed?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_015NjFiMWkgrVy4foGF1Zbax --- .../GuestStream/GuestStream_128/metadata.json | 3 ++- .../GuestStream_128/transcript.json | 2 +- .../GuestStream_128/transcript.txt | 20 +++++++++---------- 3 files changed, 13 insertions(+), 12 deletions(-) diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json index 0fbbf114b..bc7dc9526 100644 --- a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/metadata.json @@ -29,7 +29,8 @@ "speakers": { "SPEAKER_03": "Daniel Friedman", "SPEAKER_05": "Omar Hashash", - "SPEAKER_04": "Christo Kurisummoottil Thomas" + "SPEAKER_04": "Christo Kurisummoottil Thomas", + "SPEAKER_00": "Daniel Friedman" } } ] diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json index 43c36e2eb..9b78a7631 100644 --- a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.json @@ -1 +1 @@ -[{"video_id": "vjtYYbO9jCY", "segments": [{"start": 13.193, "end": 14.215, "text": " Hello, welcome.", "speaker": "Daniel Friedman"}, {"start": 15.117, "end": 16.68, "text": "It's July 14th, 2026.", "speaker": "SPEAKER_00"}, {"start": 17.06, "end": 24.194, "text": "We're in active guest stream 128.1 on active inference as the test time scaling law for physical AI agents.", "speaker": "SPEAKER_00"}, {"start": 25.076, "end": 29.785, "text": "Thank you, Omar and Christo, for joining, and please take it away for the presentation.", "speaker": "SPEAKER_00"}, {"start": 32.819, "end": 33.34, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 34.882, "end": 38.145, "text": "Thank you, Daniel, for having us today in this presentation.", "speaker": "Omar Hashash"}, {"start": 40.268, "end": 44.693, "text": "Today, as you said, we'll be presenting one of our new works.", "speaker": "Omar Hashash"}, {"start": 45.334, "end": 49.579, "text": "So first of all, for the people that don't know me, I'm Omar Hashash.", "speaker": "Omar Hashash"}, {"start": 49.599, "end": 51.982, "text": "I'm a postdoc at Virginia Tech.", "speaker": "Omar Hashash"}, {"start": 52.683, "end": 60.733, "text": "And today I'll be also joined with my colleague, Christa Thomas, which is an assistant professor at WPI.", "speaker": "Omar Hashash"}, {"start": 61.372, "end": 68.162, "text": " And together we'll be presenting one of our recent works, which is active inference as a test time scaling law for physical AI.", "speaker": "Omar Hashash"}, {"start": 68.802, "end": 70.485, "text": "So it's a bit of an interesting work.", "speaker": "Omar Hashash"}, {"start": 70.685, "end": 84.645, "text": "It gives a bit of new concepts to the role of active inference and how it plays it in physical AI and why is it really necessary to be posed as a scaling law for physical AI agents.", "speaker": "Omar Hashash"}, {"start": 85.766, "end": 89.151, "text": "So just as a brief", "speaker": "Omar Hashash"}, {"start": 89.772, "end": 116.742, "text": " introduction uh so i recently finished my phd at the brand the department of electrical computer engineering at virginia tech back in december 2025 and where my focus was on wireless communications and ai including topics like world models and digital twins and edge intelligence and i'll pass it over here to crystal just a little bit to introduce himself", "speaker": "Omar Hashash"}, {"start": 118.105, "end": 140.907, "text": " uh yeah thanks for more uh i'm christopher thomas i am currently an assistant professor at uh wpi and i lead a research group uh named trustworthy resilient ai and networks um and i did my prayer to this i did my post postdoc associate at virginia tech ec department", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 140.887, "end": 151.417, "text": " And my research mainly lies at the intersection of AI, native wireless networks, semantic communication, physical AI, and mathematical foundations of AI.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 153.379, "end": 163.288, "text": "Yeah, so both of us have this kind of hybrid background on AI and wireless communications at the same time.", "speaker": "Omar Hashash"}, {"start": 163.808, "end": 170.775, "text": "And we're going to see how that plays a role in extending us more into neuroscience concepts like active inference.", "speaker": "Omar Hashash"}, {"start": 171.042, "end": 180.511, "text": " So without any further notice, I think we can start first with the fact that AI has become the most transformational technology that is changing all our worlds.", "speaker": "Omar Hashash"}, {"start": 180.932, "end": 191.102, "text": "And of course, one of those worlds is actually the wireless communications world, and especially the 6G, the next generation of wireless communications that we're expecting to see roughly in a couple of years.", "speaker": "Omar Hashash"}, {"start": 191.983, "end": 198.429, "text": "So we already started to touch on improvements with the current generations of AI that we have.", "speaker": "Omar Hashash"}, {"start": 198.489, "end": 201.032, "text": "We call these networks as AI-native networks.", "speaker": "Omar Hashash"}, {"start": 201.012, "end": 205.487, "text": " We started to see improvements in terms of efficiency and latency.", "speaker": "Omar Hashash"}, {"start": 205.747, "end": 207.272, "text": "The improvements are massive, actually.", "speaker": "Omar Hashash"}, {"start": 207.834, "end": 209.821, "text": "But that's only the good part of the story.", "speaker": "Omar Hashash"}, {"start": 210.463, "end": 214.195, "text": "But there's also, as any other story, there's also the bad and the ugly.", "speaker": "Omar Hashash"}, {"start": 214.968, "end": 216.791, "text": " So how do the bad and the ugly look like?", "speaker": "Omar Hashash"}, {"start": 217.171, "end": 218.193, "text": "Well, they look something like this.", "speaker": "Omar Hashash"}, {"start": 219.034, "end": 230.25, "text": "So the fact that we're training what we call as an AI native interface, for example, we train it to send beams to users in the network.", "speaker": "Omar Hashash"}, {"start": 230.27, "end": 233.555, "text": "So these users can be humans, they can be physical agents.", "speaker": "Omar Hashash"}, {"start": 234.196, "end": 237.28, "text": "So a part of these", "speaker": "Omar Hashash"}, {"start": 237.26, "end": 266.355, "text": " uh beams actually work very well but the fact is that the world is always non-stationary and dynamic and it's always changing so these so these kind of uh tasks hold some sort of problems and even fail at test time um so even if we look at the level of the agents themselves that are connected to the network we can see that we have plenty of examples on failures as well so for example we have autonomous vehicles that are trained on millions and millions of driving trials", "speaker": "Omar Hashash"}, {"start": 266.977, "end": 271.406, "text": " They still continue to fail in ways that we just can't comprehend in the world.", "speaker": "Omar Hashash"}, {"start": 272.428, "end": 286.054, "text": "At the same time, we have these kind of robots that are trained to win Olympics, for example, like this robot over here, but still crash into a man on an athletics track, even though it's trained to win an Olympics.", "speaker": "Omar Hashash"}, {"start": 286.996, "end": 289.761, "text": "So there's a lot of things that don't make a lot of sense over here.", "speaker": "Omar Hashash"}, {"start": 290.939, "end": 299.869, "text": " And ironically, with the current generation of AI, we do have these big, large language models that can actually win in math and math.", "speaker": "Omar Hashash"}, {"start": 299.967, "end": 317.171, "text": " So I think here now we can pose a question, which is, why is it that today's AI is good enough to win math Olympiad, but it fails at this type of Olympiad, or more generally, in very simple green world scenarios that are very easy for us, especially with this massive amount of training?", "speaker": "Omar Hashash"}, {"start": 317.912, "end": 325.502, "text": "I'm not going to answer this now, but I'm going to use it as a motivation for the rest of the presentation and to say that we're actually missing something very big over here.", "speaker": "Omar Hashash"}, {"start": 326.562, "end": 353.076, "text": " so so if i want to look now at what do we want from these ai ages so the real the physical agents the robots the the vehicle the network so what we want from these ai agents is to become adaptive to the worlds that are happening we can start from the first one the fact that the world is always changing uh we need to train it on something but there's always going to be more data that we didn't train on so we need to have a um", "speaker": "Omar Hashash"}, {"start": 353.545, "end": 357.969, "text": " a principled way that we act in the states that we didn't actually see before.", "speaker": "Omar Hashash"}, {"start": 359.231, "end": 362.894, "text": "At the same time, we want those agents to be generalizable.", "speaker": "Omar Hashash"}, {"start": 363.155, "end": 367.979, "text": "So we want them to train them on something, but at the same time, to be able to use them on something else.", "speaker": "Omar Hashash"}, {"start": 368.94, "end": 370.622, "text": "At the same time, we want to be autonomous.", "speaker": "Omar Hashash"}, {"start": 370.742, "end": 376.768, "text": "We don't want to stay in the loop or probably minimize our presence in the loop with the AI system as much as possible.", "speaker": "Omar Hashash"}, {"start": 377.709, "end": 381.493, "text": "At the same time, we want them to have this continual learning capability.", "speaker": "Omar Hashash"}, {"start": 381.625, "end": 385.329, "text": " So we wanted to learn on top and top and top of other things.", "speaker": "Omar Hashash"}, {"start": 385.349, "end": 396.36, "text": "So we're starting from a very far away part over here with the current generation of AI, which is mostly based on large language models like the GPT models that we have, which are a massive success.", "speaker": "Omar Hashash"}, {"start": 397.301, "end": 408.713, "text": "So part of the success that started with GPT, at least the new part, was introducing these reasoning capabilities.", "speaker": "Omar Hashash"}, {"start": 408.76, "end": 423.918, "text": " So I think back with the old models, the old three models, the old one model, we started to see how this reasoning can pop up into these models and actually give you a better answer for the agents.", "speaker": "Omar Hashash"}, {"start": 424.92, "end": 436.033, "text": "So for the first time, we had this new kind of test time scaling law, which started from the fact that if you give it more computed test time, you give the agent more computed test time, you can have a better answer.", "speaker": "Omar Hashash"}, {"start": 436.756, "end": 440.842, "text": " So this was followed by another kind of capability, which is called planning.", "speaker": "Omar Hashash"}, {"start": 442.024, "end": 447.532, "text": "And this is what ushered in the agentic era or AI agents that we see today.", "speaker": "Omar Hashash"}, {"start": 448.133, "end": 452.059, "text": "And this is where GPT-5 and all the clouds that we have today.", "speaker": "Omar Hashash"}, {"start": 452.639, "end": 459.169, "text": "So basically, planning is the capability of breaking down a complex task into a sequence of actions in order to achieve a goal.", "speaker": "Omar Hashash"}, {"start": 459.723, "end": 464.077, "text": " So this is where we currently roughly are right now, probably more or less.", "speaker": "Omar Hashash"}, {"start": 464.88, "end": 468.231, "text": "But we can see here that we're moving on a cognitive path.", "speaker": "Omar Hashash"}, {"start": 468.295, "end": 492.379, "text": " so what we can say is that the current forms of reasoning and planning are good for language models are probably good enough for language models but are not enough for real world agents because the world works in a different way than language so the world is driven by laws that works apart from language how language works in terms of syntax and all those stuff so but that that route seems to be right", "speaker": "Omar Hashash"}, {"start": 493.186, "end": 515.565, "text": " so if we were to build on these cognitive capabilities and the fact that they were starting to work so this means that there should be other cognitive capabilities that we're searching for in order to reach the characteristics of the AI systems that we're searching for so if you want to dig deep here and try to search for this missing cognitive capability that's needed to move forward with the world", "speaker": "Omar Hashash"}, {"start": 515.545, "end": 523.48, "text": " we can see that it needs to be a cognitive capability that is in direct touch with the world and it's focused on the world.", "speaker": "Omar Hashash"}, {"start": 523.797, "end": 529.185, "text": " And that capability is nothing but something called common sense.", "speaker": "Omar Hashash"}, {"start": 529.947, "end": 530.988, "text": "So what is common sense?", "speaker": "Omar Hashash"}, {"start": 531.189, "end": 540.644, "text": "Common sense is, as the name implies, it's the common amount of knowledge of how the world works that we humans use in order to make sense of our world.", "speaker": "Omar Hashash"}, {"start": 541.104, "end": 548.556, "text": "It's our ability to understand the world and make use of it, which is something that is found in each human of us.", "speaker": "Omar Hashash"}, {"start": 549.548, "end": 551.252, "text": " So that's great.", "speaker": "Omar Hashash"}, {"start": 551.432, "end": 553.036, "text": "How can we plug in this common sense?", "speaker": "Omar Hashash"}, {"start": 553.437, "end": 564.982, "text": "So common sense is a broadly defined term, but we kind of know if you want to introduce this common sense into AI age, we have to search for the component of the cognitive system that drives it.", "speaker": "Omar Hashash"}, {"start": 565.523, "end": 569.211, "text": "So that component is actually what we call as a work model.", "speaker": "Omar Hashash"}, {"start": 569.832, "end": 570.793, "text": " So what is a world model?", "speaker": "Omar Hashash"}, {"start": 571.054, "end": 583.772, "text": "A world model of the physical world allows humans to understand the world in terms of its real-time state, in terms of the causal structures that exist between the elements of the world.", "speaker": "Omar Hashash"}, {"start": 584.733, "end": 590.001, "text": "And at the same time, it allows us to understand the dynamical evolution of how this world will evolve with time.", "speaker": "Omar Hashash"}, {"start": 590.403, "end": 616.495, "text": " so but the introduction of the world model itself so the fact that you're taking a state of the world and trying to understand it and trying to see how it's going to evolve in the future this requires other cognitive capabilities such as perception for example so perception is the capability to actually understand the world and try to grasp it and to make meaning of it but if we try to zoom out now at this specific moment", "speaker": "Omar Hashash"}, {"start": 616.93, "end": 620.255, "text": " What are we trying to say with all of these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 620.295, "end": 628.046, "text": "So we said that we have that ladder that started with reasoning, we added planning, we added common sense for the world model, and now we're saying that we need perception.", "speaker": "Omar Hashash"}, {"start": 628.687, "end": 631.011, "text": "What are we trying to say with all these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 631.912, "end": 642.027, "text": "What we're indirectly acknowledging over here is the fact that we need a cognitive solution that includes all these cognitive capabilities on top of today's AI.", "speaker": "Omar Hashash"}, {"start": 643.255, "end": 645.239, "text": " So, and this is not something new.", "speaker": "Omar Hashash"}, {"start": 645.279, "end": 654.796, "text": "We already know that because we already know that the AI that we have today, that's roughly based on neural networks, they're called system one architectures from psychology.", "speaker": "Omar Hashash"}, {"start": 654.836, "end": 657.501, "text": "They're called system one architectures because they're reactive.", "speaker": "Omar Hashash"}, {"start": 657.937, "end": 677.177, "text": " uh but we know that we're missing these system two type of architectures so we know already that the system one it composes around roughly 95 of how we take actions in the world and this is what we've based ai systems on since the beginning and now", "speaker": "Omar Hashash"}, {"start": 677.343, "end": 689.04, "text": " What we're trying to do is to look at these other system two rational type of thinking ways and trying to see how can these come up with a new architecture in order to solve the problems that we have.", "speaker": "Omar Hashash"}, {"start": 690.083, "end": 714.37, "text": " so how does this architecture look like it looks something roughly like this it looks something like these cognitive modules that we have that i talked about the perception module and planning the world model all connected together uh of course over a network and the network acts here as a bridge uh or additional computing resource for the ages that exist in the world", "speaker": "Omar Hashash"}, {"start": 714.856, "end": 727.734, "text": " so that it benefits itself for the decisions, intelligent decisions that it can take, just like the beamforming example I was giving at the beginning, and all the other intelligent actions that it can help the agents do in the world.", "speaker": "Omar Hashash"}, {"start": 728.455, "end": 732.221, "text": "So this was based on one of our previous works.", "speaker": "Omar Hashash"}, {"start": 733.042, "end": 736.527, "text": "And here, I'll shout out to one of our colleagues", "speaker": "Omar Hashash"}, {"start": 736.507, "end": 745.302, "text": " which mentioned our presentation in one of the theoretical neurobiology groups before and", "speaker": "Omar Hashash"}, {"start": 745.805, "end": 747.327, "text": " gave us credit for this nice work.", "speaker": "Omar Hashash"}, {"start": 747.988, "end": 751.852, "text": "So this is actually us, yours truly over here with Crystal.", "speaker": "Omar Hashash"}, {"start": 752.853, "end": 771.256, "text": "And what he was presenting is that our work is kind of an intersection with other leading works in the field, like, for example, Yann LeCun's vision of modular cognitive brain architecture and his famous model of the world.", "speaker": "Omar Hashash"}, {"start": 772.197, "end": 774.9, "text": "But what we did over here is that we actually", "speaker": "Omar Hashash"}, {"start": 775.403, "end": 777.348, "text": " We push things a little bit forward.", "speaker": "Omar Hashash"}, {"start": 777.949, "end": 785.346, "text": "So what we kind of did is that we saw the other intersection with what Carfis is actually also presenting with active inference and the free energy principle.", "speaker": "Omar Hashash"}, {"start": 786.088, "end": 793.986, "text": "And we tried to combine all of these things together into one kind of coherent story around world models that we're going to see here today.", "speaker": "Omar Hashash"}, {"start": 794.557, "end": 802.567, "text": " So going back to this architecture that I was talking about, the first step to actually use this architecture is to sense the world.", "speaker": "Omar Hashash"}, {"start": 802.888, "end": 806.553, "text": "So we already, in terms of wireless networks, already studied this before.", "speaker": "Omar Hashash"}, {"start": 807.734, "end": 816.405, "text": "We have the abilities now in next generation networks to develop sensing capabilities that allow us to divide this world over wireless networks.", "speaker": "Omar Hashash"}, {"start": 816.986, "end": 818.688, "text": "We said this before in previous works.", "speaker": "Omar Hashash"}, {"start": 819.369, "end": 822.994, "text": "And I'll try to highlight here a little bit on the fact that we have these little", "speaker": "Omar Hashash"}, {"start": 822.974, "end": 843.784, "text": " red dots that i have over here in red which are called digital twins so digital twins are the copies or the replicas or the different models of the ai agents of these physical ai agents so they're the models of the agents the physical agents that exist in the world and we have the world models", "speaker": "Omar Hashash"}, {"start": 844.068, "end": 847.213, "text": " that are basically the other things around it.", "speaker": "Omar Hashash"}, {"start": 847.394, "end": 852.923, "text": "And digital twins are a part of the world model itself, along with the different elements that exist in the world.", "speaker": "Omar Hashash"}, {"start": 854.085, "end": 859.414, "text": "So basically, this is just the high-level concept how digital twins work.", "speaker": "Omar Hashash"}, {"start": 859.854, "end": 864.061, "text": "So they're based on the fact that they're a part of the world model.", "speaker": "Omar Hashash"}, {"start": 865.103, "end": 866.405, "text": "They intersect with world models.", "speaker": "Omar Hashash"}, {"start": 867.126, "end": 869.35, "text": "Both of them, they allow", "speaker": "Omar Hashash"}, {"start": 870.19, "end": 875.258, "text": " Basically, with a digital twin, you can actually move that work model forward.", "speaker": "Omar Hashash"}, {"start": 875.619, "end": 880.307, "text": "And this intersects with the notion of AI, which is called action-conditioned work models.", "speaker": "Omar Hashash"}, {"start": 881.088, "end": 884.173, "text": "So digital twins allow you to move forward in time.", "speaker": "Omar Hashash"}, {"start": 884.814, "end": 891.405, "text": "At the same time, they allow you to send some configuration back to what we call the physical twin in the physical world.", "speaker": "Omar Hashash"}, {"start": 891.723, "end": 895.347, "text": " So these configurations, we still don't know what they are, basically, in 6G.", "speaker": "Omar Hashash"}, {"start": 896.028, "end": 898.171, "text": "And this is actually what we'll be uncovering in this work.", "speaker": "Omar Hashash"}, {"start": 898.732, "end": 910.646, "text": "So basically, as I was saying, digital twins, they take real-time updates from the world, and they return some real-time optimization feedback back to the agents in the world.", "speaker": "Omar Hashash"}, {"start": 911.147, "end": 917.815, "text": "And their role is roughly somewhere around what-if analysis and doing predictions and monitoring for the AI agent.", "speaker": "Omar Hashash"}, {"start": 918.402, "end": 926.399, "text": " And we've also shown in one of our previous work before that these agents can have continual learning capabilities, but we're going to show it more in depth over here today.", "speaker": "Omar Hashash"}, {"start": 927.321, "end": 935.618, "text": "So roughly going back to that architecture that I was talking about, now we can look at that part where we talk about the agent, the digital twin.", "speaker": "Omar Hashash"}, {"start": 936.526, "end": 955.041, "text": " of the agent as going forward in time as a process of reasoning, because we're seeing it from a different lens, not from a 6G wireless communications lens, but probably from a neuroscience or AI point of view, where reasoning is actually trying to optimize some certain action for the agent in the world.", "speaker": "Omar Hashash"}, {"start": 955.443, "end": 960.032, "text": " But at the same time, what we want to do over here is similar to what LLMs do.", "speaker": "Omar Hashash"}, {"start": 960.193, "end": 967.568, "text": "So we want to reason in order to take, just like LLMs reason to give you a better answer for some harder prompt, for example.", "speaker": "Omar Hashash"}, {"start": 968.169, "end": 975.223, "text": "What we want to do is that we actually want to use this reasoning capability that's now on the network in order", "speaker": "Omar Hashash"}, {"start": 975.44, "end": 979.008, "text": " to allow the agent to reason to take a better action in the world.", "speaker": "Omar Hashash"}, {"start": 979.549, "end": 991.677, "text": "But that will require something like test time scaling, the one that was shown empirically in large language models, but is basically missing today in these physical AI systems.", "speaker": "Omar Hashash"}, {"start": 991.657, "end": 1020.162, "text": " so this is basically the goal of our work today we're going to present how these how this reasoning over the network can benefit the ai agents so that it reasons to make to take a better action and generalize in new unforeseen scenarios that it hasn't been trained on and to see how this actually results once built from first principles like active inference it results in a test time scaling law similar to the test time scanning law that we saw with large language models", "speaker": "Omar Hashash"}, {"start": 1022.2, "end": 1025.307, "text": " So this is our work that I'll be presenting today.", "speaker": "Omar Hashash"}, {"start": 1025.368, "end": 1034.269, "text": "And it has been recently released on archive for people that want to read more in depth about it.", "speaker": "Omar Hashash"}, {"start": 1035.008, "end": 1038.514, "text": " So let's start with test time scaling in the real world.", "speaker": "Omar Hashash"}, {"start": 1039.215, "end": 1043.022, "text": "So before we start, we have to look at the story from the beginning.", "speaker": "Omar Hashash"}, {"start": 1043.502, "end": 1047.269, "text": "So what we were doing before with our L agents, for example, is that we used to train them.", "speaker": "Omar Hashash"}, {"start": 1048.11, "end": 1056.865, "text": "And then after a certain time, we would reach a, for example, we're talking, of course, you were talking, for example, about some example of autonomous driving.", "speaker": "Omar Hashash"}, {"start": 1057.604, "end": 1078.228, "text": " what we used to do is that we used to train these autonomous agents and then we used to plug them in the world after some convergence of policy we just to plug them in the world and just we just hope that things work out perfectly fine but that's not always the case and ai fails once the world changes and of course we have a lot of examples about that", "speaker": "Omar Hashash"}, {"start": 1078.563, "end": 1095.458, "text": " so before before we go to solve the problem what we have to do is to actually go back to the origin so what are we actually doing over there how does the brain actually does this type of learning and to see what we are actually missing from the story", "speaker": "Omar Hashash"}, {"start": 1095.843, "end": 1099.968, "text": " So to start first, we have to acknowledge that we have two regions of the brain that are working together.", "speaker": "Omar Hashash"}, {"start": 1101.47, "end": 1109.179, "text": "Basically, we have the prefrontal cortex where all the learning, reasoning, planning, the world model, everything happens over there.", "speaker": "Omar Hashash"}, {"start": 1110.1, "end": 1118.61, "text": "And then after that, once we reach that convergence that I was talking about over here, what we do is that we take the policy and plug it back in the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1119.957, "end": 1129.11, "text": " And after that, roughly what we do is that we really don't need all the effort from the prefrontal cortex anymore because that's the smart part of our brain.", "speaker": "Omar Hashash"}, {"start": 1129.751, "end": 1134.558, "text": "We don't need to devote any big effort for that task.", "speaker": "Omar Hashash"}, {"start": 1134.818, "end": 1137.723, "text": "And it becomes somehow a subconscious task at that specific moment.", "speaker": "Omar Hashash"}, {"start": 1138.384, "end": 1141.348, "text": "And we start to rely more on our policy while working.", "speaker": "Omar Hashash"}, {"start": 1141.368, "end": 1142.91, "text": "And this is actually what we did before.", "speaker": "Omar Hashash"}, {"start": 1143.92, "end": 1151.331, "text": " But the question is now, if I were to ask you, how much do you think about driving while driving?", "speaker": "Omar Hashash"}, {"start": 1152.307, "end": 1155.872, "text": " the answer would probably be around zero.", "speaker": "Omar Hashash"}, {"start": 1156.273, "end": 1157.875, "text": "You don't think about driving while driving.", "speaker": "Omar Hashash"}, {"start": 1158.496, "end": 1170.433, "text": "And this is something we know, and this is something agrees with what the famous psychologist, Daniel Kahneman, was a Nobel Prize winner, actually identified before.", "speaker": "Omar Hashash"}, {"start": 1170.854, "end": 1178.365, "text": "So he roughly said that we use our intuition system one around 95% of the time in order to take actions.", "speaker": "Omar Hashash"}, {"start": 1178.345, "end": 1194.295, "text": " but at the same time there's this five percent that we also take actions with so when does this five percent pop up actually and why is it just five percent so one example of how we can use that is through surprise", "speaker": "Omar Hashash"}, {"start": 1194.562, "end": 1207.816, "text": " So for example, if you're here driving a vehicle, and let's say you're on a highway, and then you get a jaywalking pedestrian, for example, that suddenly wants to cross the road, you weren't expecting it.", "speaker": "Omar Hashash"}, {"start": 1208.377, "end": 1212.481, "text": "And actually, after that, you got a surprise that you didn't expect, right?", "speaker": "Omar Hashash"}, {"start": 1212.581, "end": 1216.806, "text": "Because if you didn't expect something, you wouldn't get a surprise at the beginning.", "speaker": "Omar Hashash"}, {"start": 1218.627, "end": 1222.932, "text": "And once that happens, what you do in that case is that you start to think.", "speaker": "Omar Hashash"}, {"start": 1223.468, "end": 1228.56, "text": " So if I move forward, I probably hit the person, but I don't want that to happen.", "speaker": "Omar Hashash"}, {"start": 1229.643, "end": 1240.65, "text": "So this is what you'd likely be saying, or probably saying that if I possibly go slower, the person behind me will probably bump into me and I'll head into a crash, another crash.", "speaker": "Omar Hashash"}, {"start": 1240.917, "end": 1250.809, "text": " So probably what I need to do is probably just lower my speed just a little bit to allow the person to move in front of me, cross whatever they want to go.", "speaker": "Omar Hashash"}, {"start": 1251.59, "end": 1254.314, "text": "And then after that, I can just continue my path.", "speaker": "Omar Hashash"}, {"start": 1254.914, "end": 1263.565, "text": "So this type of thinking over here is exactly necessary in order to pass that certain situation that I wasn't really trained on.", "speaker": "Omar Hashash"}, {"start": 1263.605, "end": 1265.187, "text": "I didn't really expect in front of me.", "speaker": "Omar Hashash"}, {"start": 1266.01, "end": 1272.919, "text": " So roughly, if I want to combine it now, so this is kind of how we do this type of decision making.", "speaker": "Omar Hashash"}, {"start": 1273.5, "end": 1279.929, "text": "So basically, when you're driving, if there's no prediction error in front of you, you did not expect anything surprising.", "speaker": "Omar Hashash"}, {"start": 1279.989, "end": 1282.812, "text": "Everything in the world is working precisely as you would want it to do.", "speaker": "Omar Hashash"}, {"start": 1284.815, "end": 1289.521, "text": "You basically use your Bayes of Ganglia and basically the policy that exists in your Bayes of Ganglia.", "speaker": "Omar Hashash"}, {"start": 1289.902, "end": 1292.946, "text": "But at the same time, if you do have a prediction error,", "speaker": "Omar Hashash"}, {"start": 1292.926, "end": 1310.513, "text": " this means that your policy is somehow sub-optimal and you need to go through your frontal cortex again you need to kick it in again until it handle this new situation for you so what you would do in that case is that you would reason before you take the right action at that specific moment", "speaker": "Omar Hashash"}, {"start": 1311.067, "end": 1314.841, "text": " So now I probably ask you, where did we see the system before?", "speaker": "Omar Hashash"}, {"start": 1314.901, "end": 1318.735, "text": "This kind of system that toggles between reasoning and not reasoning.", "speaker": "Omar Hashash"}, {"start": 1318.795, "end": 1323.051, "text": "This is exactly what we've seen in", "speaker": "Omar Hashash"}, {"start": 1323.993, "end": 1326.696, "text": " models and large language models like GPT-5.", "speaker": "Omar Hashash"}, {"start": 1327.117, "end": 1328.318, "text": "So that's smart routing.", "speaker": "Omar Hashash"}, {"start": 1329.2, "end": 1332.444, "text": "So this is, you could say probably this is from where it came from.", "speaker": "Omar Hashash"}, {"start": 1332.984, "end": 1346.742, "text": "I'm going to see how that connects afterwards, but basically what the agents, the soft kind of agents, digital agents, language agents that we see today, when they reason or don't reason, they follow that specific scheme that is in the brain.", "speaker": "Omar Hashash"}, {"start": 1347.683, "end": 1351.688, "text": "So basically the prediction error is the, um,", "speaker": "Omar Hashash"}, {"start": 1352.697, "end": 1356.365, "text": " is the natural trigger, rather than having a holistic router.", "speaker": "Omar Hashash"}, {"start": 1356.405, "end": 1361.957, "text": "The router is trying to approximate that trigger, natural trigger.", "speaker": "Omar Hashash"}, {"start": 1362.077, "end": 1367.85, "text": "But the question here becomes, how do we actually deal with the scaling at that specific moment?", "speaker": "Omar Hashash"}, {"start": 1368.01, "end": 1371.077, "text": "I said that we thought and took an action.", "speaker": "Omar Hashash"}, {"start": 1371.513, "end": 1373.659, "text": " how is the action actually resulting over here?", "speaker": "Omar Hashash"}, {"start": 1373.679, "end": 1375.303, "text": "How is it actually coming into place?", "speaker": "Omar Hashash"}, {"start": 1375.403, "end": 1376.245, "text": "We still don't know.", "speaker": "Omar Hashash"}, {"start": 1376.306, "end": 1382.221, "text": "And that's something still mysterious in the AI space.", "speaker": "Omar Hashash"}, {"start": 1382.562, "end": 1385.269, "text": "But you see here what I presented just now in terms of that", "speaker": "Omar Hashash"}, {"start": 1385.772, "end": 1389.076, "text": " prediction error and trying to reason or to move that prediction error.", "speaker": "Omar Hashash"}, {"start": 1389.597, "end": 1391.539, "text": "This is exactly what active inference is.", "speaker": "Omar Hashash"}, {"start": 1391.9, "end": 1398.708, "text": "So active inference is a first principle that describes how all living systems survive in the world.", "speaker": "Omar Hashash"}, {"start": 1399.349, "end": 1406.358, "text": "And returning back to the example that I presented in the beginning, that humanoid robot and that people that is crashing.", "speaker": "Omar Hashash"}, {"start": 1406.698, "end": 1410.463, "text": "So this is how that in other ways that", "speaker": "Omar Hashash"}, {"start": 1410.848, "end": 1421.334, "text": " the absence of this minimum amount of freezing on top of the AIs that we have today can actually explain why those systems are dying systems in the world.", "speaker": "Omar Hashash"}, {"start": 1421.976, "end": 1428.452, "text": "If this is how all systems survive in the world, all living systems like us, humans or animals,", "speaker": "Omar Hashash"}, {"start": 1428.432, "end": 1441.952, "text": " survival in the world, and the fact that this is absent in those autonomous agents, autonomous physical AI agents that exist in the world, you can possibly find an explanation to the fact why these systems are dying.", "speaker": "Omar Hashash"}, {"start": 1443.89, "end": 1458.986, "text": " So of course, uh, in contrast to what, uh, to what we have, uh, we can see here that what we need to do actually is to allow those systems to survive at this time, because as I said, it's the first principle of that explains how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1459.806, "end": 1467.715, "text": "So survival at that specific moment over here, when we reached the conversions, so we were reaching something in non equilibrium steady state in terms of physics.", "speaker": "Omar Hashash"}, {"start": 1468.495, "end": 1471.078, "text": "So from a genetic point of view, you need to maintain", "speaker": "Omar Hashash"}, {"start": 1471.514, "end": 1478.004, "text": " non-equilibrium steady state by preserving the policy and the world model so that you can survive in the world.", "speaker": "Omar Hashash"}, {"start": 1479.366, "end": 1482.01, "text": "So preserving that actually has two routes.", "speaker": "Omar Hashash"}, {"start": 1482.31, "end": 1488.159, "text": "The first route is what existing works do, which is that they assume stationary conditions about the environment.", "speaker": "Omar Hashash"}, {"start": 1489.261, "end": 1493.627, "text": "So in that case, the system works just in the states that you actually saw on the training set.", "speaker": "Omar Hashash"}, {"start": 1494.549, "end": 1497.753, "text": "And survival is limited to just those conditions.", "speaker": "Omar Hashash"}, {"start": 1499.015, "end": 1499.496, "text": "But", "speaker": "Omar Hashash"}, {"start": 1499.577, "end": 1506.792, "text": " There's an alternative route for the system to survive, which is by integrating the first principles of survival from the beginning.", "speaker": "Omar Hashash"}, {"start": 1507.293, "end": 1510.379, "text": "So you don't need to actually make your world frozen at this time.", "speaker": "Omar Hashash"}, {"start": 1510.64, "end": 1512.844, "text": "You can keep it dynamic and non-stationary.", "speaker": "Omar Hashash"}, {"start": 1513.195, "end": 1519.923, "text": " but give the agent the first principle of survival so that it survives in dynamic, non-stationary worlds.", "speaker": "Omar Hashash"}, {"start": 1520.124, "end": 1526.031, "text": "So in that case, survival is no longer limited just to what the agent saw in the training set.", "speaker": "Omar Hashash"}, {"start": 1526.412, "end": 1531.959, "text": "Now it actually survives even in new situations, and it allows it to generalize, which was our main goal from the beginning.", "speaker": "Omar Hashash"}, {"start": 1532.319, "end": 1536.364, "text": "What we want is an agent that generalizes in unforeseen scenarios that we haven't seen before.", "speaker": "Omar Hashash"}, {"start": 1536.965, "end": 1539.388, "text": "And in that case, by definition,", "speaker": "Omar Hashash"}, {"start": 1539.655, "end": 1547.154, "text": " Surviving in the world is the main facet of a system, of an AI agent, basically, that generalizes.", "speaker": "Omar Hashash"}, {"start": 1548.478, "end": 1552.568, "text": "So by guaranteeing survival, you're actually guaranteeing that the agent generalizes.", "speaker": "Omar Hashash"}, {"start": 1554.303, "end": 1563.952, "text": " So, roughly speaking, this is what active inferences and what Carl Friston presented it as a way to remove prediction errors from the brain.", "speaker": "Omar Hashash"}, {"start": 1563.992, "end": 1566.234, "text": "Of course, we always have these sensory information.", "speaker": "Omar Hashash"}, {"start": 1567.635, "end": 1574.722, "text": "We're seeing if we don't have a prediction error, it means that we're in the region that minimizes free energy or basically surprised and we don't need to take any action.", "speaker": "Omar Hashash"}, {"start": 1575.122, "end": 1581.188, "text": "However, if we do have a prediction error, we will try to act on the world so that we minimize that prediction error.", "speaker": "Omar Hashash"}, {"start": 1583.007, "end": 1612.525, "text": " uh so of course in this way we can see that reasoning to reduce surprise is the minimum amount of reasoning in any in any living system to survive in the world that's the minimum amount of intelligence that you need to have in order to exist in the world over a long time so what we kind of did over here is that we tried to integrate this into what we know in terms of agents that we just don't need them to just survive we want to make them usable so that they can actually perform tasks", "speaker": "Omar Hashash"}, {"start": 1612.91, "end": 1618.963, "text": " So we went back to integrate this with what we had in terms of reinforcement learning and the policies that we had.", "speaker": "Omar Hashash"}, {"start": 1619.785, "end": 1631.29, "text": "So what happens in this case is that what we saw is that if we take in states and if those states are somehow unforeseen, it means that we can go reason", "speaker": "Omar Hashash"}, {"start": 1633.16, "end": 1637.007, "text": " in order to remove the prediction error that results from this unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 1637.688, "end": 1651.995, "text": "And at the same time, if we can use that test time scaling, we can use it to alter the actions to fit the specific scenario in order to make it optimal in that specific scenario so that the agent can actually generalize in that specific moment.", "speaker": "Omar Hashash"}, {"start": 1652.38, "end": 1666.356, "text": " So roughly speaking, if you want to put it down graphically over here in the figure, what we were doing with reinforcement learning is that we were looking at those states that are highly probable and trying to focus on those states.", "speaker": "Omar Hashash"}, {"start": 1667.602, "end": 1669.824, "text": " finding a policy that fits those states.", "speaker": "Omar Hashash"}, {"start": 1670.105, "end": 1672.187, "text": "This is basically what reinforcement learning is doing.", "speaker": "Omar Hashash"}, {"start": 1673.148, "end": 1676.792, "text": "And after that, what we did is that we froze the world.", "speaker": "Omar Hashash"}, {"start": 1676.812, "end": 1678.374, "text": "We said that the world is not going to change.", "speaker": "Omar Hashash"}, {"start": 1678.954, "end": 1682.278, "text": "So we're always going to be stuck in those states over there and those policies.", "speaker": "Omar Hashash"}, {"start": 1682.338, "end": 1684.581, "text": "But we know that that fades in the real world.", "speaker": "Omar Hashash"}, {"start": 1685.061, "end": 1691.348, "text": "What's actually happening in the real world is that because it's non-stationary, it will probably push you to a state", "speaker": "Omar Hashash"}, {"start": 1691.328, "end": 1694.413, "text": " That is not a problem for you.", "speaker": "Omar Hashash"}, {"start": 1694.573, "end": 1700.343, "text": "It's not a part of the set that you've seen in your training distribution.", "speaker": "Omar Hashash"}, {"start": 1701.285, "end": 1714.006, "text": "What you need to do with that type of reasoning to minimize prediction errors is actually what pushes you back or scales you back to your characteristic set, which is the set of states where you can use your policy again.", "speaker": "Omar Hashash"}, {"start": 1714.467, "end": 1717.352, "text": "So this is basically what we mean by test time scaling.", "speaker": "Omar Hashash"}, {"start": 1718.665, "end": 1734.702, "text": " In this case, we can see, and this is a very nice viewpoint on the story, which is that every living system, in this case, shows a general objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1736.324, "end": 1742.59, "text": "And specifically, I'm saying here something general, because this is something that all living systems perform in the world.", "speaker": "Omar Hashash"}, {"start": 1743.191, "end": 1747.255, "text": "So the fact that you are a living system, it means that you have an objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1748.416, "end": 1768.348, "text": " or else you would die so in that case we are governed by general objectives like survival over here and at the same time we have narrow objectives which allow us to achieve our specific tasks and goals with subsumed within that specific region of survival that we have", "speaker": "Omar Hashash"}, {"start": 1768.953, "end": 1775.687, "text": " So I'll just give here an example, just like the jaywalking example of here, just to say how this system kind of works.", "speaker": "Omar Hashash"}, {"start": 1776.509, "end": 1782.06, "text": "So basically, what we have is the fact that I have an autonomous driving vehicle, for example.", "speaker": "Omar Hashash"}, {"start": 1782.141, "end": 1783.283, "text": "It's driving.", "speaker": "Omar Hashash"}, {"start": 1783.482, "end": 1785.004, "text": " And everything's fine.", "speaker": "Omar Hashash"}, {"start": 1785.104, "end": 1786.365, "text": "I have no prediction error.", "speaker": "Omar Hashash"}, {"start": 1786.746, "end": 1788.988, "text": "This is something that we call passive influence in that case.", "speaker": "Omar Hashash"}, {"start": 1789.108, "end": 1791.051, "text": "So we're just taking in sensory observations.", "speaker": "Omar Hashash"}, {"start": 1791.091, "end": 1792.953, "text": "Everything is working as expected.", "speaker": "Omar Hashash"}, {"start": 1793.433, "end": 1794.194, "text": "I have no problem.", "speaker": "Omar Hashash"}, {"start": 1794.935, "end": 1799.801, "text": "And then suddenly what I have is that an agent that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1800.241, "end": 1801.743, "text": "Sorry, a human that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1803.144, "end": 1812.595, "text": "In that case, the network, of course, the network is acting here as an additional computing resource in addition to the agent that we have.", "speaker": "Omar Hashash"}, {"start": 1813.3, "end": 1815.845, "text": " you don't need to focus a lot on what the network is doing.", "speaker": "Omar Hashash"}, {"start": 1815.865, "end": 1818.269, "text": "The network is similar to the agent in that case.", "speaker": "Omar Hashash"}, {"start": 1818.75, "end": 1820.894, "text": "They're kind of one system together.", "speaker": "Omar Hashash"}, {"start": 1822.396, "end": 1830.972, "text": "So in this case, the network that we have will detect that there's a prediction error, and it will switch from passive to active inference about the world.", "speaker": "Omar Hashash"}, {"start": 1831.508, "end": 1836.993, "text": " So what that means is that the network will engage in reasoning in order to minimize the surprise.", "speaker": "Omar Hashash"}, {"start": 1837.554, "end": 1846.143, "text": "And it's going to use those digital twins that I talked about at the beginning, which capture the model of the agent and the work model itself in order to reason.", "speaker": "Omar Hashash"}, {"start": 1846.543, "end": 1854.511, "text": "And what we want to do is that we want to reason, just like I was talking about, I was giving an example here, should I move forward faster?", "speaker": "Omar Hashash"}, {"start": 1854.531, "end": 1855.492, "text": "Should I move slower?", "speaker": "Omar Hashash"}, {"start": 1855.552, "end": 1861.498, "text": "So all that thinking together, we're going to use it in order to put down a value function.", "speaker": "Omar Hashash"}, {"start": 1862.255, "end": 1873.032, "text": " So the value function here, we call it as a delta v. So basically, I'm trying to measure which action allows me to resolve my prediction error in that specific state that I am in.", "speaker": "Omar Hashash"}, {"start": 1873.052, "end": 1881.165, "text": "And then what I want to do is that I'm going to take in that feedback from the network and send it back to the physical AI agent.", "speaker": "Omar Hashash"}, {"start": 1881.566, "end": 1882.828, "text": "The physical AI agent", "speaker": "Omar Hashash"}, {"start": 1883.804, "end": 1886.367, "text": " What it's going to do is that it's going to use this feedback.", "speaker": "Omar Hashash"}, {"start": 1886.888, "end": 1891.374, "text": "It's like the prefrontal cortex sending the signals down to the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1891.875, "end": 1895.84, "text": "So the policy is over here at the level of the agent of the vehicle here.", "speaker": "Omar Hashash"}, {"start": 1896.301, "end": 1899.986, "text": "And it's going to use it in order to reason and do inference.", "speaker": "Omar Hashash"}, {"start": 1900.006, "end": 1904.412, "text": "So it's going to scale its policy from pi node to pi node prime, for example, here.", "speaker": "Omar Hashash"}, {"start": 1904.712, "end": 1909.0, "text": " So in that specific case, the action that it's going to take is going to be different than its normal.", "speaker": "Omar Hashash"}, {"start": 1909.44, "end": 1916.193, "text": "It's going to be possibly to decrease its velocity in order to avoid an accident, for instance.", "speaker": "Omar Hashash"}, {"start": 1916.213, "end": 1927.192, "text": "And then what happens is that after this jaywalking pedestrian just passes, I go back to my regular cases where I have my prediction error resolved and the pedestrian passes.", "speaker": "Omar Hashash"}, {"start": 1927.232, "end": 1929.336, "text": "So I can go back just to my passive inventory.", "speaker": "Omar Hashash"}, {"start": 1930.227, "end": 1943.289, "text": " So what you can see here is that our world is basically a transition always between foreseen scenarios, unforeseen scenarios, resolving these unforeseen scenarios, and then going back to foreseen scenarios.", "speaker": "Omar Hashash"}, {"start": 1943.57, "end": 1945.152, "text": "This is how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1945.753, "end": 1948.438, "text": "We're always transitioning between those different states.", "speaker": "Omar Hashash"}, {"start": 1949.937, "end": 1953.542, "text": " So now I'm going to go into the system model.", "speaker": "Omar Hashash"}, {"start": 1954.222, "end": 1957.887, "text": "So as I said before, we have an agent that is in the world.", "speaker": "Omar Hashash"}, {"start": 1958.768, "end": 1959.97, "text": "It has a policy panel.", "speaker": "Omar Hashash"}, {"start": 1960.891, "end": 1964.275, "text": "And we have a world model that exists over the network in this architecture.", "speaker": "Omar Hashash"}, {"start": 1965.136, "end": 1971.945, "text": "One part of the world model is composed from these digital tools, these different alternative policies that the agent can have.", "speaker": "Omar Hashash"}, {"start": 1971.925, "end": 1973.948, "text": " or models of the agent.", "speaker": "Omar Hashash"}, {"start": 1974.749, "end": 1984.383, "text": "At the same time, we have other things that are called assets which capture the different things that exist in the world, just like the human, the houses, everything else besides the agents themselves.", "speaker": "Omar Hashash"}, {"start": 1985.344, "end": 1988.269, "text": "And what's going to happen is that we're going to take sensing observations.", "speaker": "Omar Hashash"}, {"start": 1988.849, "end": 1993.877, "text": "So it's not the agent itself that's doing the sensing, it's actually the network that's doing the sensing on behalf of the agent.", "speaker": "Omar Hashash"}, {"start": 1994.978, "end": 1997.282, "text": "And what we're going to do is that we're going to do perception.", "speaker": "Omar Hashash"}, {"start": 1997.422, "end": 2000.987, "text": "So we're basically going to map our observations to", "speaker": "Omar Hashash"}, {"start": 2001.49, "end": 2009.058, "text": " the set of states that we have trying to make sense of the states, trying to infer what's going to happen over here in terms of the states.", "speaker": "Omar Hashash"}, {"start": 2009.538, "end": 2018.588, "text": "And then once we have a surprise, so we have an unforeseen scenario, what's going to happen is that the network is going to engage in counterfactual reasoning to minimize surprise.", "speaker": "Omar Hashash"}, {"start": 2019.289, "end": 2025.575, "text": "So it's going to think about all these different alternative routes with the different policies that they may have over here, different alternative policies.", "speaker": "Omar Hashash"}, {"start": 2026.496, "end": 2031.261, "text": "And then, as I said, it's going to send back these digital twin configurations.", "speaker": "Omar Hashash"}, {"start": 2032.405, "end": 2038.804, "text": " back to the agent in order to scale the policy and generalize in the unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2039.577, "end": 2043.742, "text": " So mathematically, we can look at perception, of course, as an inference process.", "speaker": "Omar Hashash"}, {"start": 2043.802, "end": 2051.032, "text": "But to do that, we have to first define our generative world model in terms of the states and observations and the policies that we have.", "speaker": "Omar Hashash"}, {"start": 2052.494, "end": 2060.764, "text": "Of course, before we do perception, we have to predict what the expected observation is going to be.", "speaker": "Omar Hashash"}, {"start": 2061.465, "end": 2064.048, "text": "And then we're going to get our sensing observations from the world.", "speaker": "Omar Hashash"}, {"start": 2064.509, "end": 2066.952, "text": "And then we're going to measure if there's a surprise or not.", "speaker": "Omar Hashash"}, {"start": 2066.932, "end": 2070.28, "text": " If we do have a surprise, of course, this is how we measure the surprise.", "speaker": "Omar Hashash"}, {"start": 2070.32, "end": 2071.644, "text": "It's just the easiest surprise.", "speaker": "Omar Hashash"}, {"start": 2073.127, "end": 2082.19, "text": "And after that, if we do have a surprise, it means that we're going to cross some threshold epsilon in order to signal that there's an unforeseen scenario in front of us.", "speaker": "Omar Hashash"}, {"start": 2082.17, "end": 2105.005, "text": " so the first thing that should happen in that case is that we should actually calculate the posterior because basically what we predicted about the world is not right we have an error so we have the first thing we have to do is to make sense of our world again by calculating this posterior after we saw our observations which is basically an inference process of it just we can see before we engage in plan and reason", "speaker": "Omar Hashash"}, {"start": 2104.985, "end": 2111.657, "text": " So, the planning part, it constitutes these digital twins and work models that engage in counterfactual reasoning.", "speaker": "Omar Hashash"}, {"start": 2111.697, "end": 2117.447, "text": "So, basically, the what-if analysis that I was giving an example about, in order to plan the plausible future world states.", "speaker": "Omar Hashash"}, {"start": 2117.647, "end": 2122.656, "text": "And the goal, of course, of this reasoning is to actually minimize future surprise.", "speaker": "Omar Hashash"}, {"start": 2122.957, "end": 2128.742, "text": " So our goal is to see how much surprise is going to result from each of these policies over here.", "speaker": "Omar Hashash"}, {"start": 2128.762, "end": 2135.889, "text": "And you can see that it's going to be nothing but a measure of the entropy at each of those states in the future.", "speaker": "Omar Hashash"}, {"start": 2135.969, "end": 2140.933, "text": "So from t plus 1 until the end of the time horizon that we have.", "speaker": "Omar Hashash"}, {"start": 2140.953, "end": 2149.2, "text": "And then after that, what we're going to do is that we're going to use this reasoning that we did in order to build a value function.", "speaker": "Omar Hashash"}, {"start": 2149.721, "end": 2150.922, "text": "We call it delta v.", "speaker": "Omar Hashash"}, {"start": 2152.252, "end": 2154.935, "text": " And where the reward is actually minimizing surprise.", "speaker": "Omar Hashash"}, {"start": 2154.955, "end": 2160.563, "text": "So as I said at the beginning, we have a general objective in the brain, which is to survive.", "speaker": "Omar Hashash"}, {"start": 2161.364, "end": 2164.348, "text": "So basically it has a value function over here.", "speaker": "Omar Hashash"}, {"start": 2164.368, "end": 2174.901, "text": "And then what we're going to do is that we're going to take this value and we're going to send it back to the agents in the world in order to scale the policy.", "speaker": "Omar Hashash"}, {"start": 2175.37, "end": 2196.602, "text": " how that looks like it's a little bit similar to what perception did so we have the perception as inference of course planning is also as inference over here because planning is an extension of minimizing surprises in the future so it's also planning as inference and now we have also we have to model action as inference but here it's a different type of conference", "speaker": "Omar Hashash"}, {"start": 2196.582, "end": 2212.228, "text": " the fact that when we did perception as influence the evidence that happened in front of us was 100 true it happened in front of us in the world so we're confident about it but when you want to do action as influence it's a little bit different", "speaker": "Omar Hashash"}, {"start": 2212.208, "end": 2217.479, "text": " So it's based on imaginary virtual data in our brain, right?", "speaker": "Omar Hashash"}, {"start": 2217.499, "end": 2220.204, "text": "When we were thinking, we were actually generating data.", "speaker": "Omar Hashash"}, {"start": 2220.445, "end": 2221.888, "text": "That thing didn't happen in the world.", "speaker": "Omar Hashash"}, {"start": 2222.108, "end": 2223.471, "text": "It happened inside our brain.", "speaker": "Omar Hashash"}, {"start": 2224.152, "end": 2230.926, "text": "But luckily, we have Judea Pearl, who came up with this method of virtual evidence.", "speaker": "Omar Hashash"}, {"start": 2231.446, "end": 2236.833, "text": " So he has a nice technique in order to update the beliefs in that case.", "speaker": "Omar Hashash"}, {"start": 2237.474, "end": 2242.48, "text": "So we can see here, if we write it down, this is going to be something called a soft Bayesian update.", "speaker": "Omar Hashash"}, {"start": 2243.161, "end": 2249.629, "text": "So it's soft because you don't really need to do a complete inference update over here.", "speaker": "Omar Hashash"}, {"start": 2250.27, "end": 2259.462, "text": "So if you write it down, you can see here at the last line, you can write it in terms of a feed forward and reasoning or inference term.", "speaker": "Omar Hashash"}, {"start": 2259.83, "end": 2260.891, "text": " which is something very nice.", "speaker": "Omar Hashash"}, {"start": 2261.211, "end": 2275.066, "text": "Because if you really look at it, if you put that theta, which is basically your normalized surprise, if the world is mostly predictable, so your theta or surprise is basically equal to 0.", "speaker": "Omar Hashash"}, {"start": 2275.647, "end": 2284.116, "text": "And if you put that exponential part over here equal to 0, you basically result in the feedforward policy that we had in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2285.142, "end": 2289.269, "text": " The fact is that action is always an inference process.", "speaker": "Omar Hashash"}, {"start": 2289.569, "end": 2293.195, "text": "It's always modulated by that exponential term that we have over here.", "speaker": "Omar Hashash"}, {"start": 2293.676, "end": 2299.746, "text": "And this is what scales the policy once we have prediction errors and once we have an unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2300.427, "end": 2311.065, "text": "So the policies in today's solution, they remain limited to the case where surprise is null due to the stationary assumptions about the world, just like we do in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2311.045, "end": 2314.932, "text": " So reinforcement learning is not stationary by itself at this time.", "speaker": "Omar Hashash"}, {"start": 2315.032, "end": 2323.968, "text": "We made it stationary because we did not compensate for this factor over here in equation 5, which is clearly a special case of the equation.", "speaker": "Omar Hashash"}, {"start": 2325.03, "end": 2329.678, "text": "However, the inference at this part can be challenging if you want to really solve the problem.", "speaker": "Omar Hashash"}, {"start": 2330.03, "end": 2335.64, "text": " It can be challenging because marginalizing over all the possible states is computationally intractable in practice.", "speaker": "Omar Hashash"}, {"start": 2336.341, "end": 2341.29, "text": "So to find a better solution, we need to resort to a variational Bayesian inference solution.", "speaker": "Omar Hashash"}, {"start": 2341.53, "end": 2347.06, "text": "And I'll keep it here for Christo to present the solution.", "speaker": "Omar Hashash"}, {"start": 2349.625, "end": 2352.51, "text": "I think I'll stop sharing for Christo to share.", "speaker": "Daniel Friedman"}, {"start": 2373.854, "end": 2374.595, "text": " Thanks, Omar.", "speaker": "Daniel Friedman"}, {"start": 2376.958, "end": 2380.243, "text": "So I will go over the solution roadmap first.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2381.305, "end": 2387.193, "text": "As Omar mentioned, our solution is grounded in variational variation inference principle.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2388.815, "end": 2401.954, "text": "So this evolves as a four-step cognitive loop, starting with perception, which is about perceiving the current state by minimizing a variational free energy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2402.474, "end": 2420.297, "text": " and then uh you know we go to the future time steps where we you know estimate the expected free energy static from the inferred state and then uh you know infer policy for actions that minimize the future surprise so that's the planning phase", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2420.277, "end": 2437.392, "text": " and then in the action phase uh so there is we introduced the concept of marco blanket which acts as a statistical boundary between agent and the environment and then policy scaling will be more modeled as a gradient descent on the free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2437.372, "end": 2453.727, "text": " um and finally okay we incorporate this uh updated policy as well as uh the land posterior distributions about the states into you know updating the world model that so that again happens via vfe minimization so", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2453.707, "end": 2464.773, "text": " These four stages is grounded on the same or minimizing the same variation free energy concept.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2465.174, "end": 2467.199, "text": "That's about the solution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2467.86, "end": 2469.825, "text": "Then first, let's look at the perception.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2469.805, "end": 2479.079, "text": " So perception is modeled as inference, inferring the posterior distribution of the states S given the observations.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2479.8, "end": 2494.362, "text": "But if you expand it in terms of the Bayes rule, you can see that this is interactable, computationally interactable because of the marginal distribution of the observations which appear in the denominator here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2495.017, "end": 2509.87, "text": " so here we move to a to find a tractable posterior distribution we actually uh formulate the kl divergence between the approximate posterior q of s comma pi zero", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2509.917, "end": 2536.633, "text": " and the KL divergence to the actual or the true posterior so that's that's the concept of variational inference so where if you actually expand this KL divergence we can see that it can be written as a variation free energy times variation of free energy as well as an additional yeah plus or so basically variation free energy can be written as a KL divergence plus a surprise time which is ln p of time here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2536.985, "end": 2566.772, "text": " uh and so minimizing the scale divergence become equal to uh yeah minimizing variation and then okay once we actually once we actually find a posterior that minimizes the kl divergence along with that we all we are also computing or we are also estimating the surprise quantity which is here uh so that's the advantage of this variational variation inference formulation here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2567.528, "end": 2591.494, "text": " um yeah so yeah so once we find an optimal or optimal posterior distribution that actually means that uh the resetting variation free energy is equal to the surprise quantity uh so which surprise the estimation of the surprise is actually needed for the policy scaling that uh homer mentioned in the previous slide", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2593.162, "end": 2597.248, "text": " So now the question is how we actually go about solving this.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2597.608, "end": 2606.241, "text": "So this can be modeled as a sequential inference process, which can be formulated as a partially observed Markov decision process as shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2606.281, "end": 2610.587, "text": "And this leads to the following factor graph as is shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2611.048, "end": 2617.617, "text": "So where we can see that, okay, so we actually modeled the transition from the state's", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2617.597, "end": 2620.328, "text": " to the observations using the matrix A here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2620.389, "end": 2626.695, "text": "So we actually model all these transition probabilities as categorical distributions.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2626.827, "end": 2654.252, "text": " and we use the matrix b to actually model the transition from the previous state to the current state st uh yeah so now optimizing the or minimizing the variation of free energy this leads to an expression as is shown here uh so okay this basically leads to variational message passing uh wherein we can see that okay so there is the message one which is basically uh accounting for the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2654.232, "end": 2675.38, "text": " uh yeah basically a forward transition which basically depends upon the forward transition probability and message 2 which depends upon the backward transition probability which is the message 2 component here and then message 3 actually uh incorporates the uh observation likelihood distribution", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2676.136, "end": 2689.173, "text": " And further, in order to compute the posterior distribution, which is basically defined as the s variable here, we basically take a softmax over the logarithm tense here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2689.694, "end": 2695.461, "text": "So that's how we actually obtain the ln q of s or the posterior distribution here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2696.42, "end": 2721.668, "text": " and uh so this posterior distribution can be shown to be okay so as we mentioned we started with minimizing the variation free energy component so at converge at the equilibrium or yeah at the convergence this um optimal posterior distribution should um converge to that of the true posterior so it basically is minimizing the um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2722.475, "end": 2745.208, "text": " variational free energy component so basically it's a gradient descent on the um free energy um component so that's what basically this uh yeah perception component basically shows and we can actually show that this uh variational free energy can be written to be uh okay the posterior times an error on the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2745.188, "end": 2757.477, "text": " uh posterior posterior distribution so basically minimizing the variation of free energy is equivalent to minimizing the error on the estimated posterior or inferred posterior distributions", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2759.262, "end": 2779.414, "text": " and yeah so now uh when we go to planning so what we actually do is that uh so we look at future time steps and we we want to find the policy that minimizes the average free energy uh across time so that's basically so so for that we need expected free energy component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2779.394, "end": 2797.639, "text": " and we show that this expected free energy component can be nicely decomposed into two components so one is the risk component which actually uh measures the kl divergence between uh the you know approximate uh oh basically the inferred um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2798.615, "end": 2820.285, "text": " distribution of observations given the current policy how that so the inferred uh distribute inferred observational distribution how that is actually comparing to the uh expected or preferred observations or the preferences that we have uh for the agent so that's basically the risk component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2820.265, "end": 2845.429, "text": " plus there is an ambiguity component that uh so basically basically it's a entropy of the probability distributions which is basically the likelihood probability distribution here which so basically if you want to minimize the expected free energy you should make sure that the inferred world model is actually close to that or inferred observations are basically close to that of the preferred observations as well as", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2845.409, "end": 2855.234, "text": " we should try to reduce the ambiguity in terms of the observations that the agent observe in future time steps.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2855.314, "end": 2858.001, "text": "So that's basically the expected free energy component.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2859.297, "end": 2877.12, "text": " uh yeah so now uh the the prior over the posterior the the prior over the policies can be written as s of max of uh minus gamma g so we should yeah the policy should ensure that it should always go in the direction of minimizing the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2877.438, "end": 2897.472, "text": " and uh yeah inferred policy uh inferred action policy can be written as um yeah summation over all the policies pi uh where yeah corresponding to action a times the the prior distribution of the particular policy pi here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2897.452, "end": 2917.3, "text": " uh and we can yeah so that's that's the inferred policy and then finally as omar mentioned in the previous slides what we actually do is using this inferred policy so this inferred policy basically uh needs an estimate of the surprise so so basically inferred policy depends upon the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2917.28, "end": 2921.485, "text": " which basically is equivalent to the surprise quantity, as I mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2921.525, "end": 2930.417, "text": "So basically, you cannot compute an exact value for the inference policy, but you can only estimate using the surprise quantity.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2930.897, "end": 2940.79, "text": "So to estimate the surprise, we need to perform reasoning.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2940.81, "end": 2945.616, "text": "So how much reasoning that you need to do depends upon this parameter called theta here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2945.596, "end": 2952.41, "text": " So that is similar to that of the test time computing LLMs like Omar mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2952.43, "end": 2959.504, "text": "So that's how we perform the planning and do the test time scaling.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2961.273, "end": 2968.691, "text": " Yeah, so now the question is whether all these computations are really stable or not.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2969.292, "end": 2972.821, "text": "For that, we need to look at the concept of Markov blanket.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2973.021, "end": 2978.033, "text": "So an AI agent that generalizes is a random dynamical system", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2978.013, "end": 2997.574, "text": " that exist over time so for the a for this ai agent to be able to generalize across unforeseen scenarios so there should exist a marco blanket which basically means that uh so marco blanket is consisting of those sensors okay so the agent interacts with the environment through the sensory state", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2997.723, "end": 3019.933, "text": " and so so basically environment influences the agent through the sensory state and the agent influences the environment through the active state so this active so this marco blanket compose which is composed of the active states as well as the sensory state uh that is something which you know separates the internal agent state from the external states", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3020.369, "end": 3046.14, "text": " yeah so that's how this the closed loop of planning um perception planning action and sensing happens and the key equation here is that so we can write the states x the flow for the state x just composed of this marker blanket states as shown in as is shown in here which is composed of two temps so one is the one is a great dissipative gradient time and another is the solenoid", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3046.12, "end": 3053.953, "text": " uh time which basically both of this time depends upon the log of the model evidence which is ln p of x here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3054.018, "end": 3078.979, "text": " uh so yeah for the agent to exist over time so there should be a mod uh marco blanket which also means that it should actually maximize the this log p of x quantity or the model evidence quantity which is basically equal to minimizing a free and i mean minimizing uh my uh l minus ln p of x which is basically corresponding to uh free energy estimate", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3078.959, "end": 3101.592, "text": " uh yeah so that's what basically this um concept of marco blanket tells us and also this also gives us a principled criterion for principle stopping criterion which basically means that uh so when do when do the agent actually stop reasoning so that can be written like uh yeah that can be written as depending upon this quantity called theta that i mentioned", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3101.572, "end": 3107.6, "text": " So yeah, theta basically decides how much of reasoning that the agent should solve.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3107.62, "end": 3118.954, "text": "So the stopping criteria is written as like one minus theta times the KL divergence between the feedforward policy as well as the inferred or the tested time scaled policy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3120.176, "end": 3123.78, "text": "So yeah, so that's basically about the stopping criteria.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3124.762, "end": 3129.688, "text": "Yeah, so now we talked about perception planning and action policies.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3130.293, "end": 3138.884, "text": " So now we have to incorporate this land policies or the land posterior distributions into the world model.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3139.365, "end": 3150.84, "text": "For that, we can actually write the generative world model as is shown here, which is composed of the joint distribution of observation state policy as well as the A and B parameters here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3150.86, "end": 3153.203, "text": "We can actually write this joint distribution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3153.964, "end": 3158.089, "text": "And then, yeah, so how do we actually compute this?", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3158.423, "end": 3186.237, "text": " distributions a and b so if you do variation inference we can actually write the posterior distributions for these components a as is shown here so basically what we do is yeah in the variation free energy expression we can actually filter out all those stems that are independent of a then if you actually consider this prior distribution of a to be a Dirichlet distribution so we choose Dirichlet because", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3186.217, "end": 3214.654, "text": " so as i mentioned i mentioned that this your a a is actually a categorical distribution so in order for the posterior to be tractable uh we need to find a conjugate distribution so that's the reason why we basically choose this dirichlet distribution so traditionally distribution can actually be parameterized by this quantity called alpha ij here and we actually can find a closed form expression for this alpha ij", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3214.634, "end": 3230.787, "text": " um yeah a close to home expression for this posterior distribution as is shown here so as you can see here um so what basically this expression does is that it basically okay basically counts the number of joint occurrences of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3230.767, "end": 3254.074, "text": " uh the state um j as well as the the particular observation auto equal to i so basically but yeah so basically as we get more evidence about the world observations so we actually basically update the observation model based on that so that's what basically this observation model posterior inference does similarly", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3254.054, "end": 3282.508, "text": " uh yeah i won't go into the details of these things yeah similarly we actually compute the posterior distribution for the transition model b again we use the same procedure of minimizing the variation inference and we also assume that the b also follows the dirichlet prior distribution which basically leads to a tractable posterior distribution as is shown here here again we can see that basically what we do is basically counting the number of co-occurrence of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3282.488, "end": 3309.711, "text": " uh s store minus one and s store so that's what basically this distribution does uh so this basically leads to uh okay given that um yeah so and given that we also assume that the states um corresponding to different um physical assets are independent in the posterior distribution we can actually write the posterior distribution as a scaled factor as is shown here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3309.961, "end": 3336.711, "text": " uh and yeah so so basically this b i j times basically compute the ij value of this b matrix uh and finally we also compute the posterior distribution for the policy again considering that the prior distribution for the policy follows a dirichlet distribution leading to a tractable posterior distribution as i mentioned uh previously so that's that's how we basically", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3336.691, "end": 3364.504, "text": " um sold the transition model observation and the policy uh for the world model updates so uh to conclude uh the entire these four stages uh perception planning action policy and finally world model update occurred by minimizing a single quantity called variational free energy uh yeah so that's that's that's about the solution so i will um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3365.699, "end": 3367.428, "text": " give it back to Omar now.", "speaker": "Daniel Friedman"}, {"start": 3372.895, "end": 3373.498, "text": "Okay.", "speaker": "Daniel Friedman"}, {"start": 3381.393, "end": 3384.397, "text": " OK, thank you, Christo, for the solution.", "speaker": "Daniel Friedman"}, {"start": 3384.917, "end": 3395.71, "text": "So as Christo was saying, just to recap both parts of the story, the first part shows how we can actually model everything.", "speaker": "Omar Hashash"}, {"start": 3395.95, "end": 3409.206, "text": "So all the perception, planning, action, everything can be modeled as an inference process in this case, which is mathematically also equivalent to a gradient descent process, basically machine learning.", "speaker": "Omar Hashash"}, {"start": 3409.186, "end": 3425.344, "text": " At the same time, we can find tractable solutions through approximate Bayesian inference in order to do the perception, the planning, the action, and the learning at the end to enable this continual type of learning.", "speaker": "Omar Hashash"}, {"start": 3425.945, "end": 3433.033, "text": "So now what we're going to do is that we're going to see some simulation results around the jaywalking.", "speaker": "Omar Hashash"}, {"start": 3433.013, "end": 3453.52, "text": " example that we have been covering since the beginning of the presentation so we're going to simulate now at test time basically our solution which is a test time solution in comparison to other solutions like Q-learning for example and see how they're going to react in that specific scenario", "speaker": "Omar Hashash"}, {"start": 3453.5, "end": 3468.91, "text": " so what's going to happen is that we're going to train uh new solutions we're going to consider the base solution our solution to be a q learning on top of it we have this minimum reasoning active influence capability that we're talking about the reasoning to minimize free energy", "speaker": "Omar Hashash"}, {"start": 3468.89, "end": 3474.46, "text": " At the same time, we have a Q-learning solution that does not have that capability.", "speaker": "Omar Hashash"}, {"start": 3474.941, "end": 3482.855, "text": "So first of all, of course, when the world changes and now we're going to have a jaywalking pedestrian that the agent has not seen before.", "speaker": "Omar Hashash"}, {"start": 3483.115, "end": 3485.7, "text": "So for example, when you have a green traffic light,", "speaker": "Omar Hashash"}, {"start": 3486.575, "end": 3489.198, "text": " it always sees that there's no pedestrian over there.", "speaker": "Omar Hashash"}, {"start": 3489.278, "end": 3494.265, "text": "So it's kind of confident that it needs to move forward and just go towards its goal.", "speaker": "Omar Hashash"}, {"start": 3494.285, "end": 3495.346, "text": "This is what it was trained for.", "speaker": "Omar Hashash"}, {"start": 3495.786, "end": 3502.074, "text": "But at that specific moment, when the jaywalking pedestrian passes, it should have a prediction error.", "speaker": "Omar Hashash"}, {"start": 3502.094, "end": 3506.66, "text": "But the fact that it has no world model on top of its policy, it cannot do it.", "speaker": "Omar Hashash"}, {"start": 3506.68, "end": 3508.983, "text": "So this is exactly what happens here on the left-hand side.", "speaker": "Omar Hashash"}, {"start": 3510.064, "end": 3513.128, "text": "We can see here that the agent starts at distance 3.", "speaker": "Omar Hashash"}, {"start": 3513.148, "end": 3515.451, "text": "It starts to move towards distance 1.", "speaker": "Omar Hashash"}, {"start": 3516.359, "end": 3525.545, "text": " But then what's going to happen is that at step two, the agent, the jaywalking pedestrian appears.", "speaker": "Omar Hashash"}, {"start": 3526.147, "end": 3527.23, "text": "In that case,", "speaker": "Omar Hashash"}, {"start": 3528.34, "end": 3534.049, "text": " What's going to happen, as I said, because it relies just on its policy, it's going to move forward.", "speaker": "Omar Hashash"}, {"start": 3534.249, "end": 3538.896, "text": "And it's going to still accelerate, because this is what it was used to in the beginning.", "speaker": "Omar Hashash"}, {"start": 3539.397, "end": 3542.222, "text": "And it did not realize that its world changed in that case.", "speaker": "Omar Hashash"}, {"start": 3542.803, "end": 3546.368, "text": "So it starts to accelerate, and it crashes on the pedestrian, of course.", "speaker": "Omar Hashash"}, {"start": 3546.388, "end": 3551.496, "text": "This is typical with methods like Q-learning.", "speaker": "Omar Hashash"}, {"start": 3551.476, "end": 3558.445, "text": " But however, if we look at our solution, the scale policy one, we can see here that what happened is that the agent stopped at distance two.", "speaker": "Omar Hashash"}, {"start": 3558.465, "end": 3559.766, "text": "It did not move after that.", "speaker": "Omar Hashash"}, {"start": 3560.367, "end": 3568.718, "text": "So you can see here in the background, there's a change towards a yellow color, which means that the agent is now in a reasoning type of mode.", "speaker": "Omar Hashash"}, {"start": 3569.298, "end": 3573.103, "text": "So what happened is that unlike the Q-learning agent, it actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3573.724, "end": 3575.827, "text": "So the Q-learning agent actually sped up.", "speaker": "Omar Hashash"}, {"start": 3576.487, "end": 3578.53, "text": "This one actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3578.51, "end": 3590.577, "text": " and kept itself in its place until the jaywalking pedestrian actually left, after which the prediction error is resolved and it starts to accelerate again.", "speaker": "Omar Hashash"}, {"start": 3590.618, "end": 3593.464, "text": "Now, that's a scenario we haven't trained it on before.", "speaker": "Omar Hashash"}, {"start": 3593.798, "end": 3602.738, "text": " But the fact that it has a world model and can reason in those specific cases, it can actually figure out a way in order to keep itself surviving.", "speaker": "Omar Hashash"}, {"start": 3603.179, "end": 3611.858, "text": "And of course, after that, it continues after that specific moment from distance two until it reaches its destination at the other end.", "speaker": "Omar Hashash"}, {"start": 3612.715, "end": 3613.877, "text": " behind that intersection.", "speaker": "Omar Hashash"}, {"start": 3614.798, "end": 3625.855, "text": "So looking here at the rewards, at least, if we look at the Q-Learning rewards, of course, what happened over here is that when the agent crashed, it had a very big negative penalty over here, of course.", "speaker": "Omar Hashash"}, {"start": 3627.037, "end": 3630.522, "text": "However, in our case, what happened is that the agent was able", "speaker": "Omar Hashash"}, {"start": 3630.502, "end": 3654.214, "text": " uh to detect that there's something wrong okay there's a surprise so it detected the surprise and you can see it here from the purple line over here so the surprise went above the threshold which is this 0.2 this epsilon threshold and after that it started to reason so you can see here what happened is that the agent is still getting some kind of reward in the background for surviving", "speaker": "Omar Hashash"}, {"start": 3655.544, "end": 3664.114, "text": " And after that, what was happening in this region between the two lines over here, the agent was actually reasoning.", "speaker": "Omar Hashash"}, {"start": 3664.815, "end": 3675.288, "text": "So it was reasoning in order to adapt and scale its policy in order to avoid the crashing scenario that it may face over here, although it lost a bit of its rewards.", "speaker": "Omar Hashash"}, {"start": 3675.929, "end": 3678.011, "text": "So we're going to see this now in a different state.", "speaker": "Omar Hashash"}, {"start": 3678.031, "end": 3682.476, "text": "We're going to compare our solution with a Q-learning solution and a Bayesian RL solution.", "speaker": "Omar Hashash"}, {"start": 3682.516, "end": 3685.5, "text": "Of course, Q-learning is a", "speaker": "Omar Hashash"}, {"start": 3685.48, "end": 3691.39, "text": " is a model-free scenario, model-free solution, and Bayesian RL is a model-based solution.", "speaker": "Omar Hashash"}, {"start": 3691.971, "end": 3694.095, "text": "So in order just to draw contrast with them.", "speaker": "Omar Hashash"}, {"start": 3694.615, "end": 3697.881, "text": "So what happens here is that we're going to train the three scenarios.", "speaker": "Omar Hashash"}, {"start": 3698.222, "end": 3703.27, "text": "We can see here that roughly they have the same training reward.", "speaker": "Omar Hashash"}, {"start": 3703.25, "end": 3708.221, "text": " And then at test time, we're going to let them see the jaywalking pedestrian.", "speaker": "Omar Hashash"}, {"start": 3709.002, "end": 3717.561, "text": "What's going to happen is that the reward will drop massively for Q-learning and Bayesian RL, and also for our solution.", "speaker": "Omar Hashash"}, {"start": 3717.581, "end": 3723.874, "text": "It's going to drop, but not that severe as the Q-learning and Bayesian RL scenarios.", "speaker": "Omar Hashash"}, {"start": 3723.854, "end": 3726.958, "text": " Although it does, it really drops.", "speaker": "Omar Hashash"}, {"start": 3727.059, "end": 3729.562, "text": "However, this comes at the expense of success.", "speaker": "Omar Hashash"}, {"start": 3730.183, "end": 3741.72, "text": "So our agent, in order to scale, it had to sacrifice its narrow rewards in order to ensure the success of its long-term objective, right?", "speaker": "Omar Hashash"}, {"start": 3741.74, "end": 3743.342, "text": "Because this is what RL is about.", "speaker": "Omar Hashash"}, {"start": 3743.362, "end": 3746.166, "text": "It's about maximizing the long-term reward.", "speaker": "Omar Hashash"}, {"start": 3746.534, "end": 3768.39, "text": " uh in order to ensure success while efficiently utilizing inference so what it did over here is that it showed us that you don't actually need to do reasoning and inference all the time you just need to do it when needed so it can be adaptive not like what a lot of people think now that if you add always add reasoning on top of the models it's always better", "speaker": "Omar Hashash"}, {"start": 3768.758, "end": 3775.127, "text": " So the efficient human-like way is that reasoning is only included once you know this is human behavior.", "speaker": "Omar Hashash"}, {"start": 3775.167, "end": 3789.366, "text": "And so it showed around 36% improvement more than above the Bayesian RL solution in terms of utilizing the computing resources in that case.", "speaker": "Omar Hashash"}, {"start": 3789.386, "end": 3795.154, "text": "So it shows a trade-off between model-based and model-free solutions.", "speaker": "Omar Hashash"}, {"start": 3795.775, "end": 3809.017, "text": " And after that, what we did is that we trained the agent that we have with our solution on a jaywalking pedestrian scenario.", "speaker": "Omar Hashash"}, {"start": 3809.037, "end": 3814.426, "text": "But we allowed it to see that scenario again and again in order to see what's going to happen.", "speaker": "Omar Hashash"}, {"start": 3815.007, "end": 3821.277, "text": "So basically what happened in that case is that the first time it saw the scenario, it had a very high surprise.", "speaker": "Omar Hashash"}, {"start": 3821.257, "end": 3845.307, "text": " after that the surprise started to decrease over time because it started to update and started to make sense that i'm always going to see a jaywalking pedestrian at that specific scenario it got used to it basically and you can see it here from the fact that it starts to predict from the other graph you can start to see in the blue graph that it starts to predict that there's a pedestrian whenever i have a green light", "speaker": "Omar Hashash"}, {"start": 3845.759, "end": 3851.13, "text": " So it starts to update the probability of the agent having a pedestrian in that case.", "speaker": "Omar Hashash"}, {"start": 3852.132, "end": 3856.861, "text": "So basically, it becomes less surprising for the agent.", "speaker": "Omar Hashash"}, {"start": 3856.881, "end": 3865.438, "text": "And at the same time, at the level of the policy itself, it also updates the probability of taking that action in that specific state.", "speaker": "Omar Hashash"}, {"start": 3865.621, "end": 3877.632, "text": " So basically what's happening is that every time it's seeing the scenario, it's reinforcing that I should see a jaywalking pedestrian at that green light in the jaywalking scenario that I'm going to see.", "speaker": "Omar Hashash"}, {"start": 3878.052, "end": 3887.841, "text": "And I know very well what my action, that specific action that I was able to get through my reasoning, I know very well what that action should be and it reinforces it in the policy.", "speaker": "Omar Hashash"}, {"start": 3888.542, "end": 3894.447, "text": "So after 50 episodes, what happens is that the decision making moves from being", "speaker": "Omar Hashash"}, {"start": 3894.427, "end": 3915.321, "text": " controlled through inference through reasoning to become feed forward because basically it becomes a part of its work model and the policy after that in order to get it basically it gets reinforced over time at test time so you can see here that even with we have reinforcement learning now at test time and this is basically how the agent can become a continual learning agent", "speaker": "Omar Hashash"}, {"start": 3916.567, "end": 3924.176, "text": " So just to conclude this now, this is just a comprehensive roadmap of wireless communication and AI working together.", "speaker": "Omar Hashash"}, {"start": 3924.216, "end": 3930.504, "text": "Basically, our networks are moving forward very fast, just like AI.", "speaker": "Omar Hashash"}, {"start": 3930.544, "end": 3939.696, "text": "And what we want is that we have this ultimate intersection between the two so that we can reach what we call as AGI native networks in the future.", "speaker": "Omar Hashash"}, {"start": 3939.716, "end": 3943.3, "text": "So we're moving from AI native networks to AGI native networks.", "speaker": "Omar Hashash"}, {"start": 3943.28, "end": 3957.675, "text": " where these things build on the different technologies and systems that we talked about today, the world model, the digital twins, our presented test time scaling, and more broadly, a big cognitive system to architecture of the brain.", "speaker": "Omar Hashash"}, {"start": 3958.376, "end": 3963.721, "text": "And with that, I'll conclude the presentation, and I'll be happy to answer any questions you may have.", "speaker": "Omar Hashash"}, {"start": 3964.642, "end": 3965.503, "text": "Crystal and I, of course.", "speaker": "Daniel Friedman"}, {"start": 3968.426, "end": 3968.787, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 3969.427, "end": 3969.928, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 3972.743, "end": 3975.365, "text": " Yeah, just one short personal comment.", "speaker": "SPEAKER_00"}, {"start": 3975.505, "end": 3983.172, "text": "I think this really builds nicely from Christo's previous presentation, which was about the semantic information theory.", "speaker": "SPEAKER_00"}, {"start": 3983.192, "end": 4002.75, "text": "And this takes it and really connects it quite well to the passive, active planning, online learning setting and how those features are being implemented in a functional sense by the different recent generations of LLM.", "speaker": "SPEAKER_00"}, {"start": 4002.73, "end": 4010.245, "text": " and sort of associated systems like the skills, tools, sub-agent dispatch, chain of thought, all those kinds of things.", "speaker": "SPEAKER_00"}, {"start": 4011.808, "end": 4012.069, "text": "All right.", "speaker": "SPEAKER_00"}, {"start": 4013.732, "end": 4014.854, "text": "Krista, do you want to add anything?", "speaker": "Daniel Friedman"}, {"start": 4015.255, "end": 4016.858, "text": "And then I'll read some questions.", "speaker": "Daniel Friedman"}, {"start": 4023.301, "end": 4023.502, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 4023.522, "end": 4024.925, "text": "All right.", "speaker": "Daniel Friedman"}, {"start": 4025.306, "end": 4033.547, "text": "If anyone in live chat asks a question, I'll read it, but I'll start with a few questions that were submitted by email from Joel Robinson.", "speaker": "SPEAKER_00"}, {"start": 4035.89, "end": 4046.562, "text": " Okay, Jill wrote, I've been reading active inference as the test time scaling law for physical AI agents and had several questions that arise from my own work on regulatory architectures.", "speaker": "SPEAKER_00"}, {"start": 4047.323, "end": 4057.934, "text": "I'm asking these because the empirical literature suggests certain variables must exist in any physical or biological agent, and I'm trying to understand how they map onto your formulation.", "speaker": "SPEAKER_00"}, {"start": 4059.155, "end": 4059.636, "text": "So here we go.", "speaker": "SPEAKER_00"}, {"start": 4061.135, "end": 4064.099, "text": " Question one, inference capacity as a state variable.", "speaker": "SPEAKER_00"}, {"start": 4064.84, "end": 4078.016, "text": "You write that the scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time, and that that policy update is modeled as a soft Bayesian inference process.", "speaker": "SPEAKER_00"}, {"start": 4079.838, "end": 4084.324, "text": "In cognitive neuroscience, inference capacity is state dependent and collapses under load.", "speaker": "SPEAKER_00"}, {"start": 4084.878, "end": 4090.288, "text": " So where in your formulation is the agent's capacity to perform inference represented?", "speaker": "SPEAKER_00"}, {"start": 4091.069, "end": 4099.624, "text": "Is inference assumed to be always available, or is there a state variable that modulates whether the agent can actually execute the Bayesian update?", "speaker": "SPEAKER_00"}, {"start": 4100.526, "end": 4103.511, "text": "And I'm going to paste this into the Zoom chat as well.", "speaker": "SPEAKER_00"}, {"start": 4107.322, "end": 4115.953, "text": " Yeah, so I think it's the last part of the question that's really exciting.", "speaker": "Omar Hashash"}, {"start": 4116.414, "end": 4126.087, "text": "So I do agree with the question itself as being there are limitations for the architecture.", "speaker": "Omar Hashash"}, {"start": 4127.469, "end": 4133.717, "text": "But in our case, I think we're considering mostly kind of like optimal scenarios, optimal execution scenarios.", "speaker": "Omar Hashash"}, {"start": 4134.777, "end": 4140.367, "text": " So the thing is that I agree is that I think inference collapses.", "speaker": "Omar Hashash"}, {"start": 4140.847, "end": 4145.575, "text": "I think the word was collapses in certain scenarios.", "speaker": "Omar Hashash"}, {"start": 4145.596, "end": 4156.574, "text": "But in our case, we specifically highlighted the fact here that we are using our system two, sorry, our system one most of the time to take actions.", "speaker": "Omar Hashash"}, {"start": 4157.256, "end": 4160.481, "text": "So in that case, our resources for reasoning", "speaker": "Omar Hashash"}, {"start": 4160.731, "end": 4188.409, "text": " assuming that we have a lot of time possibly not instantaneous reflective reflexive like a reflex while driving that i have enough time to think about the different scenarios that i can take in that specific scenario so i'm not considering here a where like an example where an agent where a human for example needs to take a split second action i'm feeling that there needs to be enough time", "speaker": "Omar Hashash"}, {"start": 4188.389, "end": 4201.842, "text": " so that the agent thinks because assuming i'm on a highway and there's like roughly four five six seconds which i think is more than enough for a human to think in that case at least in the broad sense", "speaker": "Omar Hashash"}, {"start": 4202.21, "end": 4213.169, "text": " So these are the type of problems that we're considering over here, which in that case, I think the inference part doesn't crash specifically in that case or collapses in that case.", "speaker": "Omar Hashash"}, {"start": 4213.189, "end": 4226.091, "text": "But I do acknowledge the fact that there are situations where this kind of has to be all short circuited, I would say, and you just have to pass to just taking an instantaneous reflex.", "speaker": "Omar Hashash"}, {"start": 4228.045, "end": 4254.222, "text": " Yeah, one point I'll add there, and this is often brought up by the RxInfer and sort of message graph developers, is that with the message graph doing local free energy minimization, you can have different frequencies or different number of iterations of optimization between observations.", "speaker": "SPEAKER_00"}, {"start": 4254.202, "end": 4257.747, "text": " You could be doing an observation slower or at a faster frequency.", "speaker": "SPEAKER_00"}, {"start": 4258.188, "end": 4262.534, "text": "You could have intermittent or a fix or a variable compute budget.", "speaker": "SPEAKER_00"}, {"start": 4263.275, "end": 4277.796, "text": "And so that's like an advantage for online methods is that you're kind of just doing as good as you can in an online fashion given the amount of cognitive or computational resourcing available.", "speaker": "SPEAKER_00"}, {"start": 4279.244, "end": 4289.809, "text": " Modulo, anything like a catastrophic failure, that kind of breaks out of the box of the performance on the self-driving car itself breaking down mechanically.", "speaker": "SPEAKER_00"}, {"start": 4289.849, "end": 4291.994, "text": "It's sort of the mortal computing.", "speaker": "SPEAKER_00"}, {"start": 4292.054, "end": 4294.62, "text": "That's like the embodiment side.", "speaker": "SPEAKER_00"}, {"start": 4294.6, "end": 4319.376, "text": " and here is like really scoping out a lot of the formalisms and methods for the computational the cognitive that kind of presuppose the integrity of the functioning machine even though that becomes also a question when you're talking about these like network systems yeah i i i believe you said it in the best way possible and so uh it's uh", "speaker": "SPEAKER_00"}, {"start": 4319.71, "end": 4323.435, "text": " What you raised over here are the perfect times at that specific moment.", "speaker": "Omar Hashash"}, {"start": 4323.876, "end": 4327.781, "text": "So of course, we're working on a strict budget over here, but we're considering that.", "speaker": "Omar Hashash"}, {"start": 4329.303, "end": 4331.265, "text": "So I'm processing it in a different way.", "speaker": "Omar Hashash"}, {"start": 4331.666, "end": 4344.503, "text": "So I think our goal of this work was to lay down the foundation of how can an agent exist in the world in order to solve the problem that", "speaker": "Omar Hashash"}, {"start": 4345.023, "end": 4349.61, "text": " the problem that we were facing or possibly what we were missing before in that case.", "speaker": "Omar Hashash"}, {"start": 4350.692, "end": 4360.568, "text": "Of course, the more practical sense that you were talking about, which is different frequencies that you may be taking in samples in about the world, the different noise", "speaker": "Omar Hashash"}, {"start": 4360.835, "end": 4365.928, "text": " the noises that you may have, how much computationally feasible is that.", "speaker": "Omar Hashash"}, {"start": 4366.068, "end": 4368.515, "text": "I think these things are right on point.", "speaker": "Omar Hashash"}, {"start": 4369.056, "end": 4371.803, "text": "And they should follow exactly from this specific moment.", "speaker": "Omar Hashash"}, {"start": 4371.823, "end": 4375.232, "text": "So I think this is roughly a", "speaker": "Omar Hashash"}, {"start": 4375.887, "end": 4399.026, "text": " a i wouldn't say a sketch but i would say as a initial solution of a problem uh that we can start with and of course on top of that we can add a different uh feasibility type of thing i assume there's a lot of solution that needs to come up in this space more but the tricky part about it was uh as you were saying about the different um", "speaker": "Omar Hashash"}, {"start": 4399.006, "end": 4403.395, "text": " connecting it into the broader AI scenario, like what was really missing.", "speaker": "Omar Hashash"}, {"start": 4403.415, "end": 4407.644, "text": "Like we're seeing these different LLMs, chain of thought, reasoning, planning, agentic.", "speaker": "Omar Hashash"}, {"start": 4408.566, "end": 4412.034, "text": "But we already knew that, like we already had agents before.", "speaker": "Omar Hashash"}, {"start": 4412.054, "end": 4413.156, "text": "So what was actually missing?", "speaker": "Omar Hashash"}, {"start": 4413.176, "end": 4416.002, "text": "What did we really not see before?", "speaker": "Omar Hashash"}, {"start": 4416.269, "end": 4433.012, "text": " And it was actually that fact that I think that the new part in LLMs was mostly based on the test time part more than the regular vanilla kind of scaling the training part, the model size, the parameters.", "speaker": "Omar Hashash"}, {"start": 4432.992, "end": 4442.342, "text": " So we try to focus on that more, try to see why is it really that I have an empirical scaling law with language models, and what's the intersection over there?", "speaker": "Omar Hashash"}, {"start": 4443.183, "end": 4447.768, "text": "So yeah, it was putting down the theoretical part of the story, I think, at first.", "speaker": "Omar Hashash"}, {"start": 4448.229, "end": 4453.575, "text": "But definitely, the next steps should be based on making it more computationally tractable.", "speaker": "Omar Hashash"}, {"start": 4455.016, "end": 4455.317, "text": "Awesome.", "speaker": "SPEAKER_00"}, {"start": 4455.457, "end": 4457.219, "text": "All right, here's another question from Joel.", "speaker": "SPEAKER_00"}, {"start": 4457.86, "end": 4458.24, "text": "He wrote,", "speaker": "SPEAKER_00"}, {"start": 4460.06, "end": 4484.515, "text": " temporal depth and prediction window collapse you note that humans simulate counterfactual scenarios and plan future states of the world and this is central to resolving prediction error empirically temporal depth is not fixed it narrows under stress fatigue or overload is temporal depth treated as a dynamic state variable variable in your model if the prediction window collapse how does the scaling law behave", "speaker": "SPEAKER_00"}, {"start": 4488.19, "end": 4495.268, "text": " So if I got the answer correctly, you were asking about the temporal depth?", "speaker": "Omar Hashash"}, {"start": 4496.872, "end": 4499.078, "text": "Yeah, like the depth of planning.", "speaker": "SPEAKER_00"}, {"start": 4500.577, "end": 4502.921, "text": " OK, yeah.", "speaker": "SPEAKER_00"}, {"start": 4502.981, "end": 4511.875, "text": "So of course, the depth here is, of course, focused on we're having it as the least shallow layer, I would say.", "speaker": "Omar Hashash"}, {"start": 4511.935, "end": 4517.323, "text": "Of course, we're not considering that we have a hierarchical form of the world, of course, which we know about.", "speaker": "Omar Hashash"}, {"start": 4518.385, "end": 4523.513, "text": "And we're not considering the different temporal aspects of the world that can exist over there.", "speaker": "Omar Hashash"}, {"start": 4523.493, "end": 4528.862, "text": " Of course, the next steps should be that this world is composed in a hierarchical way.", "speaker": "Omar Hashash"}, {"start": 4529.262, "end": 4530.645, "text": "There's a lot of structure in there.", "speaker": "Omar Hashash"}, {"start": 4531.166, "end": 4540.982, "text": "At which level of planning should we actually limit ourselves in so that we don't fall into the kind of collapse that Joel's talking about over here?", "speaker": "Omar Hashash"}, {"start": 4541.002, "end": 4548.454, "text": "So I think even there are some interesting parts over here related to hierarchical active inference.", "speaker": "Omar Hashash"}, {"start": 4548.822, "end": 4557.631, "text": " So the hierarchical part of modeling the world, which I assume should be also our next objective over here.", "speaker": "Omar Hashash"}, {"start": 4557.932, "end": 4560.52, "text": "I'm not sure if Christo wants to add a certain part over here.", "speaker": "Daniel Friedman"}, {"start": 4564.972, "end": 4576.104, "text": " Yeah, so I think Christo, because he had some words on hierarchical abstractions and hierarchical representations.", "speaker": "Omar Hashash"}, {"start": 4576.785, "end": 4582.371, "text": "So this is a different part of the depth that we were talking about, about the world model.", "speaker": "Omar Hashash"}, {"start": 4582.391, "end": 4588.498, "text": "So we're talking about the depth of the world in terms of the aspects that it may have hierarchically.", "speaker": "Omar Hashash"}, {"start": 4588.478, "end": 4603.135, "text": " And at the same time, we're talking about the depth in terms of time going millisecond by millisecond, or going in terms of representations that can possibly last longer over time, that don't change in an instantaneous manner.", "speaker": "Omar Hashash"}, {"start": 4604.157, "end": 4610.284, "text": "Yeah, I'll just copy two more questions in from Joel.", "speaker": "SPEAKER_00"}, {"start": 4613.707, "end": 4630.551, "text": " on some related topics these are all great great comments so just so it's all um there the one one area was about generalization and again i think it speaks to this nexus of basically like", "speaker": "SPEAKER_00"}, {"start": 4630.531, "end": 4635.501, "text": " there's what is the computational, like structure learning?", "speaker": "SPEAKER_00"}, {"start": 4635.862, "end": 4639.65, "text": "There still could be novel situations that are outside your structure learning envelope.", "speaker": "SPEAKER_00"}, {"start": 4640.371, "end": 4647.105, "text": "There could be like unknown unknowns, or there could be things that were, or there could even be structural possibilities that aren't adjacencies.", "speaker": "SPEAKER_00"}, {"start": 4647.727, "end": 4650.85, "text": " Then there's the implementation of the computer resources.", "speaker": "SPEAKER_00"}, {"start": 4651.831, "end": 4655.194, "text": "The math is like the computer science is free.", "speaker": "SPEAKER_00"}, {"start": 4656.195, "end": 4666.664, "text": "And then there's these questions about the capacity and how the compute system's capacity re-enters into the test time.", "speaker": "SPEAKER_00"}, {"start": 4666.704, "end": 4675.472, "text": "And what if you develop the test time in one situation and then it has a poor drop-off", "speaker": "SPEAKER_00"}, {"start": 4675.452, "end": 4702.911, "text": " in a compute limited setting but i think that your that um point about this um training objective or constraint being an underappreciated or underutilized factor in large model training is very key and i hope that it comes through yeah so um", "speaker": "SPEAKER_00"}, {"start": 4703.498, "end": 4710.371, "text": " I'll try to hit on the part which you're talking about, the unknown unknowns, those parts.", "speaker": "Omar Hashash"}, {"start": 4710.711, "end": 4717.123, "text": "So of course, there are still room enough for a lot of things to pop up in the world.", "speaker": "Omar Hashash"}, {"start": 4717.143, "end": 4725.118, "text": "And the fact that what we were dealing with the world is we're trying to deal with the part of the world", "speaker": "Omar Hashash"}, {"start": 4726.094, "end": 4737.353, "text": " that we don't really know about so that's the whole premise of the whole story which is the fact that you want to escape from the training grounds and escape into the test time part", "speaker": "Omar Hashash"}, {"start": 4738.126, "end": 4740.21, "text": " Of course, at this time, there's always going to be capacity.", "speaker": "Omar Hashash"}, {"start": 4740.23, "end": 4753.715, "text": "I think we do have, we included some resources in our paper about like, for example, if you have a new scenario that that's actually not explained by none of our states, let's say if you're doing perception.", "speaker": "Omar Hashash"}, {"start": 4754.255, "end": 4757.061, "text": "So in that case, you would need to, let's say, for example, expand your model.", "speaker": "Omar Hashash"}, {"start": 4757.261, "end": 4762.21, "text": "So you need to go through from a Bayesian sense, you would need to look at Bayesian model expansion.", "speaker": "Omar Hashash"}, {"start": 4762.73, "end": 4767.641, "text": " But of course, these have other limitations, and you have to specify the criteria.", "speaker": "Omar Hashash"}, {"start": 4767.661, "end": 4777.363, "text": "Just like we said that we have a prediction error, and we have a certain epsilon, that if we cross that prediction error, this means that you're going to foresee the scenario.", "speaker": "Omar Hashash"}, {"start": 4777.597, "end": 4788.038, "text": " would probably be in a case where you're trying to fit in your certain state of the world into one of the states that you have, but probably none of those states are actually explaining your situation.", "speaker": "Omar Hashash"}, {"start": 4788.198, "end": 4791.925, "text": "So in that case, the most logical thing to say is that this doesn't fit anything.", "speaker": "Omar Hashash"}, {"start": 4791.966, "end": 4793.749, "text": "This is something very new.", "speaker": "Omar Hashash"}, {"start": 4793.729, "end": 4799.375, "text": " you might probably need to say that this is something like a composition of different states together.", "speaker": "Omar Hashash"}, {"start": 4799.395, "end": 4801.037, "text": "I haven't seen this before.", "speaker": "Omar Hashash"}, {"start": 4801.097, "end": 4810.668, "text": "But in our case, let's say we took the simplistic first step about this, which is the fact that just the probability distribution is changing.", "speaker": "Omar Hashash"}, {"start": 4811.509, "end": 4819.939, "text": "So in a very simple scenario that all of us, I think, see roughly in our lifetime, which is the fact that someone is jaywalking.", "speaker": "Omar Hashash"}, {"start": 4819.919, "end": 4821.822, "text": " So this is something very easy to grasp.", "speaker": "Omar Hashash"}, {"start": 4822.062, "end": 4829.594, "text": "And the fact that we already know what a human is, what a traffic light is, what a green light is, we already know the traffic rules.", "speaker": "Omar Hashash"}, {"start": 4830.115, "end": 4831.617, "text": "So this is something very well known.", "speaker": "Omar Hashash"}, {"start": 4832.258, "end": 4840.772, "text": "In that case, we just flip the rules a little bit so that rather than crossing at a red light, you're crossing at a green light to see what the behavior is.", "speaker": "Omar Hashash"}, {"start": 4840.852, "end": 4844.037, "text": "But of course, I'll keep your imagination", "speaker": "Omar Hashash"}, {"start": 4844.017, "end": 4865.805, "text": " to help you in that case to see how much still we have a lot more to deal with in the future but i think that should be the main focus of this whole thing is trying to figure out step by step what are those critical things how can they actually be posed at this time so that we found find a", "speaker": "Omar Hashash"}, {"start": 4866.477, "end": 4869.164, "text": " first principle kind of solution to these systems.", "speaker": "Omar Hashash"}, {"start": 4869.185, "end": 4874.259, "text": "So building on the fact that all of us solve that problem in that certain way.", "speaker": "Omar Hashash"}, {"start": 4875.823, "end": 4877.989, "text": "So I think that should hit directly on that point.", "speaker": "Daniel Friedman"}, {"start": 4880.585, "end": 4896.881, "text": " Yeah, one area that I see this being very relevant to is the distillation and the on-policy distillation, the teacher, the co-learner, all these kind of model transformation and training strategies.", "speaker": "SPEAKER_00"}, {"start": 4898.583, "end": 4910.355, "text": "All different policy training strategies and this as a comparable measure, even if one", "speaker": "SPEAKER_00"}, {"start": 4911.989, "end": 4930.825, "text": " um test time is well it's 30 hours for this long horizon coding agent it's eight hours for this camera image model even though they're totally different domains they're not going to have a domain benchmark that's comparable and then for the models that are multimodal", "speaker": "SPEAKER_00"}, {"start": 4930.805, "end": 4938.658, "text": " or they're used in a multimodal setting, there aren't benchmarks at the domain level that would apply.", "speaker": "SPEAKER_00"}, {"start": 4938.678, "end": 4947.952, "text": "So it ends up being a sort of weakest link question with these systems sometimes.", "speaker": "SPEAKER_00"}, {"start": 4949.956, "end": 4957.808, "text": "I think that connects to the model training strategies, which are upstream of how well these models do.", "speaker": "SPEAKER_00"}, {"start": 4958.851, "end": 4972.725, "text": " And the fact, this is actually the first point that I wanted to talk about, which is you don't really have a benchmark of what's good or bad, which is the fact that you need to figure things out.", "speaker": "Omar Hashash"}, {"start": 4973.165, "end": 4985.938, "text": "In order to do that, there needs to be a first principle kind of way to make sure at least that it's doing the minimum best possible thing that any human could do.", "speaker": "Omar Hashash"}, {"start": 4986.745, "end": 4994.597, "text": " which is the fact of why we're looking at this active inference part, which is putting it as a limiting factor, actually than putting it as an upper bound.", "speaker": "Omar Hashash"}, {"start": 4995.078, "end": 5002.069, "text": "So we're saying that this is the minimum amount of reason that you need to have in order to solve this type of problem.", "speaker": "Omar Hashash"}, {"start": 5002.45, "end": 5005.114, "text": "But this does not mean that it's the only reason.", "speaker": "Omar Hashash"}, {"start": 5005.154, "end": 5012.145, "text": "So you can add on top of that other things and you can just expand as much as possible, but at least,", "speaker": "Omar Hashash"}, {"start": 5012.463, "end": 5024.216, "text": " to exist in the world and to solve problems like the problems that we see with these autonomous vehicles, Teslas, Waymos, to think about those humanoid robots that just crash down and fall.", "speaker": "Omar Hashash"}, {"start": 5024.957, "end": 5029.081, "text": "So we're trying to at least address that type of problem.", "speaker": "Omar Hashash"}, {"start": 5029.982, "end": 5031.724, "text": "But of course, that's bare minimum.", "speaker": "Daniel Friedman"}, {"start": 5035.228, "end": 5035.588, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 5035.969, "end": 5037.07, "text": "Yeah, and the paper.", "speaker": "Daniel Friedman"}, {"start": 5038.248, "end": 5040.19, "text": " really lays out that direction.", "speaker": "SPEAKER_00"}, {"start": 5041.351, "end": 5042.592, "text": "Do you have any last comments?", "speaker": "SPEAKER_00"}, {"start": 5042.712, "end": 5045.774, "text": "Otherwise, I hope this has been a good entry point to the topic.", "speaker": "SPEAKER_00"}, {"start": 5047.436, "end": 5048.056, "text": "Yeah.", "speaker": "Omar Hashash"}, {"start": 5048.797, "end": 5050.498, "text": "It's actually been a pleasure.", "speaker": "Omar Hashash"}, {"start": 5050.518, "end": 5054.222, "text": "I think the questions were really great and exciting.", "speaker": "Omar Hashash"}, {"start": 5055.543, "end": 5057.424, "text": "And I really enjoyed presenting over here today.", "speaker": "Omar Hashash"}, {"start": 5058.345, "end": 5059.146, "text": "Thank you for having us.", "speaker": "Omar Hashash"}, {"start": 5060.367, "end": 5060.727, "text": "Thank you.", "speaker": "SPEAKER_00"}, {"start": 5061.007, "end": 5063.569, "text": "Thank you to authors and to Joel for the questions.", "speaker": "SPEAKER_00"}, {"start": 5064.21, "end": 5065.331, "text": "So, bye.", "speaker": "Daniel Friedman"}, {"start": 5066.932, "end": 5067.152, "text": "Bye-bye.", "speaker": "Daniel Friedman"}, {"start": 5067.172, "end": 5067.453, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 5067.533, "end": 5067.833, "text": "Thank you.", "speaker": "Daniel Friedman"}]}] \ No newline at end of file +[{"video_id": "vjtYYbO9jCY", "segments": [{"start": 13.193, "end": 14.215, "text": " Hello, welcome.", "speaker": "Daniel Friedman"}, {"start": 15.117, "end": 16.68, "text": "It's July 14th, 2026.", "speaker": "Daniel Friedman"}, {"start": 17.06, "end": 24.194, "text": "We're in active guest stream 128.1 on active inference as the test time scaling law for physical AI agents.", "speaker": "Daniel Friedman"}, {"start": 25.076, "end": 29.785, "text": "Thank you, Omar and Christo, for joining, and please take it away for the presentation.", "speaker": "Daniel Friedman"}, {"start": 32.819, "end": 33.34, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 34.882, "end": 38.145, "text": "Thank you, Daniel, for having us today in this presentation.", "speaker": "Omar Hashash"}, {"start": 40.268, "end": 44.693, "text": "Today, as you said, we'll be presenting one of our new works.", "speaker": "Omar Hashash"}, {"start": 45.334, "end": 49.579, "text": "So first of all, for the people that don't know me, I'm Omar Hashash.", "speaker": "Omar Hashash"}, {"start": 49.599, "end": 51.982, "text": "I'm a postdoc at Virginia Tech.", "speaker": "Omar Hashash"}, {"start": 52.683, "end": 60.733, "text": "And today I'll be also joined with my colleague, Christa Thomas, which is an assistant professor at WPI.", "speaker": "Omar Hashash"}, {"start": 61.372, "end": 68.162, "text": " And together we'll be presenting one of our recent works, which is active inference as a test time scaling law for physical AI.", "speaker": "Omar Hashash"}, {"start": 68.802, "end": 70.485, "text": "So it's a bit of an interesting work.", "speaker": "Omar Hashash"}, {"start": 70.685, "end": 84.645, "text": "It gives a bit of new concepts to the role of active inference and how it plays it in physical AI and why is it really necessary to be posed as a scaling law for physical AI agents.", "speaker": "Omar Hashash"}, {"start": 85.766, "end": 89.151, "text": "So just as a brief", "speaker": "Omar Hashash"}, {"start": 89.772, "end": 116.742, "text": " introduction uh so i recently finished my phd at the brand the department of electrical computer engineering at virginia tech back in december 2025 and where my focus was on wireless communications and ai including topics like world models and digital twins and edge intelligence and i'll pass it over here to crystal just a little bit to introduce himself", "speaker": "Omar Hashash"}, {"start": 118.105, "end": 140.907, "text": " uh yeah thanks for more uh i'm christopher thomas i am currently an assistant professor at uh wpi and i lead a research group uh named trustworthy resilient ai and networks um and i did my prayer to this i did my post postdoc associate at virginia tech ec department", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 140.887, "end": 151.417, "text": " And my research mainly lies at the intersection of AI, native wireless networks, semantic communication, physical AI, and mathematical foundations of AI.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 153.379, "end": 163.288, "text": "Yeah, so both of us have this kind of hybrid background on AI and wireless communications at the same time.", "speaker": "Omar Hashash"}, {"start": 163.808, "end": 170.775, "text": "And we're going to see how that plays a role in extending us more into neuroscience concepts like active inference.", "speaker": "Omar Hashash"}, {"start": 171.042, "end": 180.511, "text": " So without any further notice, I think we can start first with the fact that AI has become the most transformational technology that is changing all our worlds.", "speaker": "Omar Hashash"}, {"start": 180.932, "end": 191.102, "text": "And of course, one of those worlds is actually the wireless communications world, and especially the 6G, the next generation of wireless communications that we're expecting to see roughly in a couple of years.", "speaker": "Omar Hashash"}, {"start": 191.983, "end": 198.429, "text": "So we already started to touch on improvements with the current generations of AI that we have.", "speaker": "Omar Hashash"}, {"start": 198.489, "end": 201.032, "text": "We call these networks as AI-native networks.", "speaker": "Omar Hashash"}, {"start": 201.012, "end": 205.487, "text": " We started to see improvements in terms of efficiency and latency.", "speaker": "Omar Hashash"}, {"start": 205.747, "end": 207.272, "text": "The improvements are massive, actually.", "speaker": "Omar Hashash"}, {"start": 207.834, "end": 209.821, "text": "But that's only the good part of the story.", "speaker": "Omar Hashash"}, {"start": 210.463, "end": 214.195, "text": "But there's also, as any other story, there's also the bad and the ugly.", "speaker": "Omar Hashash"}, {"start": 214.968, "end": 216.791, "text": " So how do the bad and the ugly look like?", "speaker": "Omar Hashash"}, {"start": 217.171, "end": 218.193, "text": "Well, they look something like this.", "speaker": "Omar Hashash"}, {"start": 219.034, "end": 230.25, "text": "So the fact that we're training what we call as an AI native interface, for example, we train it to send beams to users in the network.", "speaker": "Omar Hashash"}, {"start": 230.27, "end": 233.555, "text": "So these users can be humans, they can be physical agents.", "speaker": "Omar Hashash"}, {"start": 234.196, "end": 237.28, "text": "So a part of these", "speaker": "Omar Hashash"}, {"start": 237.26, "end": 266.355, "text": " uh beams actually work very well but the fact is that the world is always non-stationary and dynamic and it's always changing so these so these kind of uh tasks hold some sort of problems and even fail at test time um so even if we look at the level of the agents themselves that are connected to the network we can see that we have plenty of examples on failures as well so for example we have autonomous vehicles that are trained on millions and millions of driving trials", "speaker": "Omar Hashash"}, {"start": 266.977, "end": 271.406, "text": " They still continue to fail in ways that we just can't comprehend in the world.", "speaker": "Omar Hashash"}, {"start": 272.428, "end": 286.054, "text": "At the same time, we have these kind of robots that are trained to win Olympics, for example, like this robot over here, but still crash into a man on an athletics track, even though it's trained to win an Olympics.", "speaker": "Omar Hashash"}, {"start": 286.996, "end": 289.761, "text": "So there's a lot of things that don't make a lot of sense over here.", "speaker": "Omar Hashash"}, {"start": 290.939, "end": 299.869, "text": " And ironically, with the current generation of AI, we do have these big, large language models that can actually win in math and math.", "speaker": "Omar Hashash"}, {"start": 299.967, "end": 317.171, "text": " So I think here now we can pose a question, which is, why is it that today's AI is good enough to win math Olympiad, but it fails at this type of Olympiad, or more generally, in very simple green world scenarios that are very easy for us, especially with this massive amount of training?", "speaker": "Omar Hashash"}, {"start": 317.912, "end": 325.502, "text": "I'm not going to answer this now, but I'm going to use it as a motivation for the rest of the presentation and to say that we're actually missing something very big over here.", "speaker": "Omar Hashash"}, {"start": 326.562, "end": 353.076, "text": " so so if i want to look now at what do we want from these ai ages so the real the physical agents the robots the the vehicle the network so what we want from these ai agents is to become adaptive to the worlds that are happening we can start from the first one the fact that the world is always changing uh we need to train it on something but there's always going to be more data that we didn't train on so we need to have a um", "speaker": "Omar Hashash"}, {"start": 353.545, "end": 357.969, "text": " a principled way that we act in the states that we didn't actually see before.", "speaker": "Omar Hashash"}, {"start": 359.231, "end": 362.894, "text": "At the same time, we want those agents to be generalizable.", "speaker": "Omar Hashash"}, {"start": 363.155, "end": 367.979, "text": "So we want them to train them on something, but at the same time, to be able to use them on something else.", "speaker": "Omar Hashash"}, {"start": 368.94, "end": 370.622, "text": "At the same time, we want to be autonomous.", "speaker": "Omar Hashash"}, {"start": 370.742, "end": 376.768, "text": "We don't want to stay in the loop or probably minimize our presence in the loop with the AI system as much as possible.", "speaker": "Omar Hashash"}, {"start": 377.709, "end": 381.493, "text": "At the same time, we want them to have this continual learning capability.", "speaker": "Omar Hashash"}, {"start": 381.625, "end": 385.329, "text": " So we wanted to learn on top and top and top of other things.", "speaker": "Omar Hashash"}, {"start": 385.349, "end": 396.36, "text": "So we're starting from a very far away part over here with the current generation of AI, which is mostly based on large language models like the GPT models that we have, which are a massive success.", "speaker": "Omar Hashash"}, {"start": 397.301, "end": 408.713, "text": "So part of the success that started with GPT, at least the new part, was introducing these reasoning capabilities.", "speaker": "Omar Hashash"}, {"start": 408.76, "end": 423.918, "text": " So I think back with the old models, the old three models, the old one model, we started to see how this reasoning can pop up into these models and actually give you a better answer for the agents.", "speaker": "Omar Hashash"}, {"start": 424.92, "end": 436.033, "text": "So for the first time, we had this new kind of test time scaling law, which started from the fact that if you give it more computed test time, you give the agent more computed test time, you can have a better answer.", "speaker": "Omar Hashash"}, {"start": 436.756, "end": 440.842, "text": " So this was followed by another kind of capability, which is called planning.", "speaker": "Omar Hashash"}, {"start": 442.024, "end": 447.532, "text": "And this is what ushered in the agentic era or AI agents that we see today.", "speaker": "Omar Hashash"}, {"start": 448.133, "end": 452.059, "text": "And this is where GPT-5 and all the clouds that we have today.", "speaker": "Omar Hashash"}, {"start": 452.639, "end": 459.169, "text": "So basically, planning is the capability of breaking down a complex task into a sequence of actions in order to achieve a goal.", "speaker": "Omar Hashash"}, {"start": 459.723, "end": 464.077, "text": " So this is where we currently roughly are right now, probably more or less.", "speaker": "Omar Hashash"}, {"start": 464.88, "end": 468.231, "text": "But we can see here that we're moving on a cognitive path.", "speaker": "Omar Hashash"}, {"start": 468.295, "end": 492.379, "text": " so what we can say is that the current forms of reasoning and planning are good for language models are probably good enough for language models but are not enough for real world agents because the world works in a different way than language so the world is driven by laws that works apart from language how language works in terms of syntax and all those stuff so but that that route seems to be right", "speaker": "Omar Hashash"}, {"start": 493.186, "end": 515.565, "text": " so if we were to build on these cognitive capabilities and the fact that they were starting to work so this means that there should be other cognitive capabilities that we're searching for in order to reach the characteristics of the AI systems that we're searching for so if you want to dig deep here and try to search for this missing cognitive capability that's needed to move forward with the world", "speaker": "Omar Hashash"}, {"start": 515.545, "end": 523.48, "text": " we can see that it needs to be a cognitive capability that is in direct touch with the world and it's focused on the world.", "speaker": "Omar Hashash"}, {"start": 523.797, "end": 529.185, "text": " And that capability is nothing but something called common sense.", "speaker": "Omar Hashash"}, {"start": 529.947, "end": 530.988, "text": "So what is common sense?", "speaker": "Omar Hashash"}, {"start": 531.189, "end": 540.644, "text": "Common sense is, as the name implies, it's the common amount of knowledge of how the world works that we humans use in order to make sense of our world.", "speaker": "Omar Hashash"}, {"start": 541.104, "end": 548.556, "text": "It's our ability to understand the world and make use of it, which is something that is found in each human of us.", "speaker": "Omar Hashash"}, {"start": 549.548, "end": 551.252, "text": " So that's great.", "speaker": "Omar Hashash"}, {"start": 551.432, "end": 553.036, "text": "How can we plug in this common sense?", "speaker": "Omar Hashash"}, {"start": 553.437, "end": 564.982, "text": "So common sense is a broadly defined term, but we kind of know if you want to introduce this common sense into AI age, we have to search for the component of the cognitive system that drives it.", "speaker": "Omar Hashash"}, {"start": 565.523, "end": 569.211, "text": "So that component is actually what we call as a work model.", "speaker": "Omar Hashash"}, {"start": 569.832, "end": 570.793, "text": " So what is a world model?", "speaker": "Omar Hashash"}, {"start": 571.054, "end": 583.772, "text": "A world model of the physical world allows humans to understand the world in terms of its real-time state, in terms of the causal structures that exist between the elements of the world.", "speaker": "Omar Hashash"}, {"start": 584.733, "end": 590.001, "text": "And at the same time, it allows us to understand the dynamical evolution of how this world will evolve with time.", "speaker": "Omar Hashash"}, {"start": 590.403, "end": 616.495, "text": " so but the introduction of the world model itself so the fact that you're taking a state of the world and trying to understand it and trying to see how it's going to evolve in the future this requires other cognitive capabilities such as perception for example so perception is the capability to actually understand the world and try to grasp it and to make meaning of it but if we try to zoom out now at this specific moment", "speaker": "Omar Hashash"}, {"start": 616.93, "end": 620.255, "text": " What are we trying to say with all of these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 620.295, "end": 628.046, "text": "So we said that we have that ladder that started with reasoning, we added planning, we added common sense for the world model, and now we're saying that we need perception.", "speaker": "Omar Hashash"}, {"start": 628.687, "end": 631.011, "text": "What are we trying to say with all these cognitive capabilities?", "speaker": "Omar Hashash"}, {"start": 631.912, "end": 642.027, "text": "What we're indirectly acknowledging over here is the fact that we need a cognitive solution that includes all these cognitive capabilities on top of today's AI.", "speaker": "Omar Hashash"}, {"start": 643.255, "end": 645.239, "text": " So, and this is not something new.", "speaker": "Omar Hashash"}, {"start": 645.279, "end": 654.796, "text": "We already know that because we already know that the AI that we have today, that's roughly based on neural networks, they're called system one architectures from psychology.", "speaker": "Omar Hashash"}, {"start": 654.836, "end": 657.501, "text": "They're called system one architectures because they're reactive.", "speaker": "Omar Hashash"}, {"start": 657.937, "end": 677.177, "text": " uh but we know that we're missing these system two type of architectures so we know already that the system one it composes around roughly 95 of how we take actions in the world and this is what we've based ai systems on since the beginning and now", "speaker": "Omar Hashash"}, {"start": 677.343, "end": 689.04, "text": " What we're trying to do is to look at these other system two rational type of thinking ways and trying to see how can these come up with a new architecture in order to solve the problems that we have.", "speaker": "Omar Hashash"}, {"start": 690.083, "end": 714.37, "text": " so how does this architecture look like it looks something roughly like this it looks something like these cognitive modules that we have that i talked about the perception module and planning the world model all connected together uh of course over a network and the network acts here as a bridge uh or additional computing resource for the ages that exist in the world", "speaker": "Omar Hashash"}, {"start": 714.856, "end": 727.734, "text": " so that it benefits itself for the decisions, intelligent decisions that it can take, just like the beamforming example I was giving at the beginning, and all the other intelligent actions that it can help the agents do in the world.", "speaker": "Omar Hashash"}, {"start": 728.455, "end": 732.221, "text": "So this was based on one of our previous works.", "speaker": "Omar Hashash"}, {"start": 733.042, "end": 736.527, "text": "And here, I'll shout out to one of our colleagues", "speaker": "Omar Hashash"}, {"start": 736.507, "end": 745.302, "text": " which mentioned our presentation in one of the theoretical neurobiology groups before and", "speaker": "Omar Hashash"}, {"start": 745.805, "end": 747.327, "text": " gave us credit for this nice work.", "speaker": "Omar Hashash"}, {"start": 747.988, "end": 751.852, "text": "So this is actually us, yours truly over here with Crystal.", "speaker": "Omar Hashash"}, {"start": 752.853, "end": 771.256, "text": "And what he was presenting is that our work is kind of an intersection with other leading works in the field, like, for example, Yann LeCun's vision of modular cognitive brain architecture and his famous model of the world.", "speaker": "Omar Hashash"}, {"start": 772.197, "end": 774.9, "text": "But what we did over here is that we actually", "speaker": "Omar Hashash"}, {"start": 775.403, "end": 777.348, "text": " We push things a little bit forward.", "speaker": "Omar Hashash"}, {"start": 777.949, "end": 785.346, "text": "So what we kind of did is that we saw the other intersection with what Carfis is actually also presenting with active inference and the free energy principle.", "speaker": "Omar Hashash"}, {"start": 786.088, "end": 793.986, "text": "And we tried to combine all of these things together into one kind of coherent story around world models that we're going to see here today.", "speaker": "Omar Hashash"}, {"start": 794.557, "end": 802.567, "text": " So going back to this architecture that I was talking about, the first step to actually use this architecture is to sense the world.", "speaker": "Omar Hashash"}, {"start": 802.888, "end": 806.553, "text": "So we already, in terms of wireless networks, already studied this before.", "speaker": "Omar Hashash"}, {"start": 807.734, "end": 816.405, "text": "We have the abilities now in next generation networks to develop sensing capabilities that allow us to divide this world over wireless networks.", "speaker": "Omar Hashash"}, {"start": 816.986, "end": 818.688, "text": "We said this before in previous works.", "speaker": "Omar Hashash"}, {"start": 819.369, "end": 822.994, "text": "And I'll try to highlight here a little bit on the fact that we have these little", "speaker": "Omar Hashash"}, {"start": 822.974, "end": 843.784, "text": " red dots that i have over here in red which are called digital twins so digital twins are the copies or the replicas or the different models of the ai agents of these physical ai agents so they're the models of the agents the physical agents that exist in the world and we have the world models", "speaker": "Omar Hashash"}, {"start": 844.068, "end": 847.213, "text": " that are basically the other things around it.", "speaker": "Omar Hashash"}, {"start": 847.394, "end": 852.923, "text": "And digital twins are a part of the world model itself, along with the different elements that exist in the world.", "speaker": "Omar Hashash"}, {"start": 854.085, "end": 859.414, "text": "So basically, this is just the high-level concept how digital twins work.", "speaker": "Omar Hashash"}, {"start": 859.854, "end": 864.061, "text": "So they're based on the fact that they're a part of the world model.", "speaker": "Omar Hashash"}, {"start": 865.103, "end": 866.405, "text": "They intersect with world models.", "speaker": "Omar Hashash"}, {"start": 867.126, "end": 869.35, "text": "Both of them, they allow", "speaker": "Omar Hashash"}, {"start": 870.19, "end": 875.258, "text": " Basically, with a digital twin, you can actually move that work model forward.", "speaker": "Omar Hashash"}, {"start": 875.619, "end": 880.307, "text": "And this intersects with the notion of AI, which is called action-conditioned work models.", "speaker": "Omar Hashash"}, {"start": 881.088, "end": 884.173, "text": "So digital twins allow you to move forward in time.", "speaker": "Omar Hashash"}, {"start": 884.814, "end": 891.405, "text": "At the same time, they allow you to send some configuration back to what we call the physical twin in the physical world.", "speaker": "Omar Hashash"}, {"start": 891.723, "end": 895.347, "text": " So these configurations, we still don't know what they are, basically, in 6G.", "speaker": "Omar Hashash"}, {"start": 896.028, "end": 898.171, "text": "And this is actually what we'll be uncovering in this work.", "speaker": "Omar Hashash"}, {"start": 898.732, "end": 910.646, "text": "So basically, as I was saying, digital twins, they take real-time updates from the world, and they return some real-time optimization feedback back to the agents in the world.", "speaker": "Omar Hashash"}, {"start": 911.147, "end": 917.815, "text": "And their role is roughly somewhere around what-if analysis and doing predictions and monitoring for the AI agent.", "speaker": "Omar Hashash"}, {"start": 918.402, "end": 926.399, "text": " And we've also shown in one of our previous work before that these agents can have continual learning capabilities, but we're going to show it more in depth over here today.", "speaker": "Omar Hashash"}, {"start": 927.321, "end": 935.618, "text": "So roughly going back to that architecture that I was talking about, now we can look at that part where we talk about the agent, the digital twin.", "speaker": "Omar Hashash"}, {"start": 936.526, "end": 955.041, "text": " of the agent as going forward in time as a process of reasoning, because we're seeing it from a different lens, not from a 6G wireless communications lens, but probably from a neuroscience or AI point of view, where reasoning is actually trying to optimize some certain action for the agent in the world.", "speaker": "Omar Hashash"}, {"start": 955.443, "end": 960.032, "text": " But at the same time, what we want to do over here is similar to what LLMs do.", "speaker": "Omar Hashash"}, {"start": 960.193, "end": 967.568, "text": "So we want to reason in order to take, just like LLMs reason to give you a better answer for some harder prompt, for example.", "speaker": "Omar Hashash"}, {"start": 968.169, "end": 975.223, "text": "What we want to do is that we actually want to use this reasoning capability that's now on the network in order", "speaker": "Omar Hashash"}, {"start": 975.44, "end": 979.008, "text": " to allow the agent to reason to take a better action in the world.", "speaker": "Omar Hashash"}, {"start": 979.549, "end": 991.677, "text": "But that will require something like test time scaling, the one that was shown empirically in large language models, but is basically missing today in these physical AI systems.", "speaker": "Omar Hashash"}, {"start": 991.657, "end": 1020.162, "text": " so this is basically the goal of our work today we're going to present how these how this reasoning over the network can benefit the ai agents so that it reasons to make to take a better action and generalize in new unforeseen scenarios that it hasn't been trained on and to see how this actually results once built from first principles like active inference it results in a test time scaling law similar to the test time scanning law that we saw with large language models", "speaker": "Omar Hashash"}, {"start": 1022.2, "end": 1025.307, "text": " So this is our work that I'll be presenting today.", "speaker": "Omar Hashash"}, {"start": 1025.368, "end": 1034.269, "text": "And it has been recently released on archive for people that want to read more in depth about it.", "speaker": "Omar Hashash"}, {"start": 1035.008, "end": 1038.514, "text": " So let's start with test time scaling in the real world.", "speaker": "Omar Hashash"}, {"start": 1039.215, "end": 1043.022, "text": "So before we start, we have to look at the story from the beginning.", "speaker": "Omar Hashash"}, {"start": 1043.502, "end": 1047.269, "text": "So what we were doing before with our L agents, for example, is that we used to train them.", "speaker": "Omar Hashash"}, {"start": 1048.11, "end": 1056.865, "text": "And then after a certain time, we would reach a, for example, we're talking, of course, you were talking, for example, about some example of autonomous driving.", "speaker": "Omar Hashash"}, {"start": 1057.604, "end": 1078.228, "text": " what we used to do is that we used to train these autonomous agents and then we used to plug them in the world after some convergence of policy we just to plug them in the world and just we just hope that things work out perfectly fine but that's not always the case and ai fails once the world changes and of course we have a lot of examples about that", "speaker": "Omar Hashash"}, {"start": 1078.563, "end": 1095.458, "text": " so before before we go to solve the problem what we have to do is to actually go back to the origin so what are we actually doing over there how does the brain actually does this type of learning and to see what we are actually missing from the story", "speaker": "Omar Hashash"}, {"start": 1095.843, "end": 1099.968, "text": " So to start first, we have to acknowledge that we have two regions of the brain that are working together.", "speaker": "Omar Hashash"}, {"start": 1101.47, "end": 1109.179, "text": "Basically, we have the prefrontal cortex where all the learning, reasoning, planning, the world model, everything happens over there.", "speaker": "Omar Hashash"}, {"start": 1110.1, "end": 1118.61, "text": "And then after that, once we reach that convergence that I was talking about over here, what we do is that we take the policy and plug it back in the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1119.957, "end": 1129.11, "text": " And after that, roughly what we do is that we really don't need all the effort from the prefrontal cortex anymore because that's the smart part of our brain.", "speaker": "Omar Hashash"}, {"start": 1129.751, "end": 1134.558, "text": "We don't need to devote any big effort for that task.", "speaker": "Omar Hashash"}, {"start": 1134.818, "end": 1137.723, "text": "And it becomes somehow a subconscious task at that specific moment.", "speaker": "Omar Hashash"}, {"start": 1138.384, "end": 1141.348, "text": "And we start to rely more on our policy while working.", "speaker": "Omar Hashash"}, {"start": 1141.368, "end": 1142.91, "text": "And this is actually what we did before.", "speaker": "Omar Hashash"}, {"start": 1143.92, "end": 1151.331, "text": " But the question is now, if I were to ask you, how much do you think about driving while driving?", "speaker": "Omar Hashash"}, {"start": 1152.307, "end": 1155.872, "text": " the answer would probably be around zero.", "speaker": "Omar Hashash"}, {"start": 1156.273, "end": 1157.875, "text": "You don't think about driving while driving.", "speaker": "Omar Hashash"}, {"start": 1158.496, "end": 1170.433, "text": "And this is something we know, and this is something agrees with what the famous psychologist, Daniel Kahneman, was a Nobel Prize winner, actually identified before.", "speaker": "Omar Hashash"}, {"start": 1170.854, "end": 1178.365, "text": "So he roughly said that we use our intuition system one around 95% of the time in order to take actions.", "speaker": "Omar Hashash"}, {"start": 1178.345, "end": 1194.295, "text": " but at the same time there's this five percent that we also take actions with so when does this five percent pop up actually and why is it just five percent so one example of how we can use that is through surprise", "speaker": "Omar Hashash"}, {"start": 1194.562, "end": 1207.816, "text": " So for example, if you're here driving a vehicle, and let's say you're on a highway, and then you get a jaywalking pedestrian, for example, that suddenly wants to cross the road, you weren't expecting it.", "speaker": "Omar Hashash"}, {"start": 1208.377, "end": 1212.481, "text": "And actually, after that, you got a surprise that you didn't expect, right?", "speaker": "Omar Hashash"}, {"start": 1212.581, "end": 1216.806, "text": "Because if you didn't expect something, you wouldn't get a surprise at the beginning.", "speaker": "Omar Hashash"}, {"start": 1218.627, "end": 1222.932, "text": "And once that happens, what you do in that case is that you start to think.", "speaker": "Omar Hashash"}, {"start": 1223.468, "end": 1228.56, "text": " So if I move forward, I probably hit the person, but I don't want that to happen.", "speaker": "Omar Hashash"}, {"start": 1229.643, "end": 1240.65, "text": "So this is what you'd likely be saying, or probably saying that if I possibly go slower, the person behind me will probably bump into me and I'll head into a crash, another crash.", "speaker": "Omar Hashash"}, {"start": 1240.917, "end": 1250.809, "text": " So probably what I need to do is probably just lower my speed just a little bit to allow the person to move in front of me, cross whatever they want to go.", "speaker": "Omar Hashash"}, {"start": 1251.59, "end": 1254.314, "text": "And then after that, I can just continue my path.", "speaker": "Omar Hashash"}, {"start": 1254.914, "end": 1263.565, "text": "So this type of thinking over here is exactly necessary in order to pass that certain situation that I wasn't really trained on.", "speaker": "Omar Hashash"}, {"start": 1263.605, "end": 1265.187, "text": "I didn't really expect in front of me.", "speaker": "Omar Hashash"}, {"start": 1266.01, "end": 1272.919, "text": " So roughly, if I want to combine it now, so this is kind of how we do this type of decision making.", "speaker": "Omar Hashash"}, {"start": 1273.5, "end": 1279.929, "text": "So basically, when you're driving, if there's no prediction error in front of you, you did not expect anything surprising.", "speaker": "Omar Hashash"}, {"start": 1279.989, "end": 1282.812, "text": "Everything in the world is working precisely as you would want it to do.", "speaker": "Omar Hashash"}, {"start": 1284.815, "end": 1289.521, "text": "You basically use your Bayes of Ganglia and basically the policy that exists in your Bayes of Ganglia.", "speaker": "Omar Hashash"}, {"start": 1289.902, "end": 1292.946, "text": "But at the same time, if you do have a prediction error,", "speaker": "Omar Hashash"}, {"start": 1292.926, "end": 1310.513, "text": " this means that your policy is somehow sub-optimal and you need to go through your frontal cortex again you need to kick it in again until it handle this new situation for you so what you would do in that case is that you would reason before you take the right action at that specific moment", "speaker": "Omar Hashash"}, {"start": 1311.067, "end": 1314.841, "text": " So now I probably ask you, where did we see the system before?", "speaker": "Omar Hashash"}, {"start": 1314.901, "end": 1318.735, "text": "This kind of system that toggles between reasoning and not reasoning.", "speaker": "Omar Hashash"}, {"start": 1318.795, "end": 1323.051, "text": "This is exactly what we've seen in", "speaker": "Omar Hashash"}, {"start": 1323.993, "end": 1326.696, "text": " models and large language models like GPT-5.", "speaker": "Omar Hashash"}, {"start": 1327.117, "end": 1328.318, "text": "So that's smart routing.", "speaker": "Omar Hashash"}, {"start": 1329.2, "end": 1332.444, "text": "So this is, you could say probably this is from where it came from.", "speaker": "Omar Hashash"}, {"start": 1332.984, "end": 1346.742, "text": "I'm going to see how that connects afterwards, but basically what the agents, the soft kind of agents, digital agents, language agents that we see today, when they reason or don't reason, they follow that specific scheme that is in the brain.", "speaker": "Omar Hashash"}, {"start": 1347.683, "end": 1351.688, "text": "So basically the prediction error is the, um,", "speaker": "Omar Hashash"}, {"start": 1352.697, "end": 1356.365, "text": " is the natural trigger, rather than having a holistic router.", "speaker": "Omar Hashash"}, {"start": 1356.405, "end": 1361.957, "text": "The router is trying to approximate that trigger, natural trigger.", "speaker": "Omar Hashash"}, {"start": 1362.077, "end": 1367.85, "text": "But the question here becomes, how do we actually deal with the scaling at that specific moment?", "speaker": "Omar Hashash"}, {"start": 1368.01, "end": 1371.077, "text": "I said that we thought and took an action.", "speaker": "Omar Hashash"}, {"start": 1371.513, "end": 1373.659, "text": " how is the action actually resulting over here?", "speaker": "Omar Hashash"}, {"start": 1373.679, "end": 1375.303, "text": "How is it actually coming into place?", "speaker": "Omar Hashash"}, {"start": 1375.403, "end": 1376.245, "text": "We still don't know.", "speaker": "Omar Hashash"}, {"start": 1376.306, "end": 1382.221, "text": "And that's something still mysterious in the AI space.", "speaker": "Omar Hashash"}, {"start": 1382.562, "end": 1385.269, "text": "But you see here what I presented just now in terms of that", "speaker": "Omar Hashash"}, {"start": 1385.772, "end": 1389.076, "text": " prediction error and trying to reason or to move that prediction error.", "speaker": "Omar Hashash"}, {"start": 1389.597, "end": 1391.539, "text": "This is exactly what active inference is.", "speaker": "Omar Hashash"}, {"start": 1391.9, "end": 1398.708, "text": "So active inference is a first principle that describes how all living systems survive in the world.", "speaker": "Omar Hashash"}, {"start": 1399.349, "end": 1406.358, "text": "And returning back to the example that I presented in the beginning, that humanoid robot and that people that is crashing.", "speaker": "Omar Hashash"}, {"start": 1406.698, "end": 1410.463, "text": "So this is how that in other ways that", "speaker": "Omar Hashash"}, {"start": 1410.848, "end": 1421.334, "text": " the absence of this minimum amount of freezing on top of the AIs that we have today can actually explain why those systems are dying systems in the world.", "speaker": "Omar Hashash"}, {"start": 1421.976, "end": 1428.452, "text": "If this is how all systems survive in the world, all living systems like us, humans or animals,", "speaker": "Omar Hashash"}, {"start": 1428.432, "end": 1441.952, "text": " survival in the world, and the fact that this is absent in those autonomous agents, autonomous physical AI agents that exist in the world, you can possibly find an explanation to the fact why these systems are dying.", "speaker": "Omar Hashash"}, {"start": 1443.89, "end": 1458.986, "text": " So of course, uh, in contrast to what, uh, to what we have, uh, we can see here that what we need to do actually is to allow those systems to survive at this time, because as I said, it's the first principle of that explains how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1459.806, "end": 1467.715, "text": "So survival at that specific moment over here, when we reached the conversions, so we were reaching something in non equilibrium steady state in terms of physics.", "speaker": "Omar Hashash"}, {"start": 1468.495, "end": 1471.078, "text": "So from a genetic point of view, you need to maintain", "speaker": "Omar Hashash"}, {"start": 1471.514, "end": 1478.004, "text": " non-equilibrium steady state by preserving the policy and the world model so that you can survive in the world.", "speaker": "Omar Hashash"}, {"start": 1479.366, "end": 1482.01, "text": "So preserving that actually has two routes.", "speaker": "Omar Hashash"}, {"start": 1482.31, "end": 1488.159, "text": "The first route is what existing works do, which is that they assume stationary conditions about the environment.", "speaker": "Omar Hashash"}, {"start": 1489.261, "end": 1493.627, "text": "So in that case, the system works just in the states that you actually saw on the training set.", "speaker": "Omar Hashash"}, {"start": 1494.549, "end": 1497.753, "text": "And survival is limited to just those conditions.", "speaker": "Omar Hashash"}, {"start": 1499.015, "end": 1499.496, "text": "But", "speaker": "Omar Hashash"}, {"start": 1499.577, "end": 1506.792, "text": " There's an alternative route for the system to survive, which is by integrating the first principles of survival from the beginning.", "speaker": "Omar Hashash"}, {"start": 1507.293, "end": 1510.379, "text": "So you don't need to actually make your world frozen at this time.", "speaker": "Omar Hashash"}, {"start": 1510.64, "end": 1512.844, "text": "You can keep it dynamic and non-stationary.", "speaker": "Omar Hashash"}, {"start": 1513.195, "end": 1519.923, "text": " but give the agent the first principle of survival so that it survives in dynamic, non-stationary worlds.", "speaker": "Omar Hashash"}, {"start": 1520.124, "end": 1526.031, "text": "So in that case, survival is no longer limited just to what the agent saw in the training set.", "speaker": "Omar Hashash"}, {"start": 1526.412, "end": 1531.959, "text": "Now it actually survives even in new situations, and it allows it to generalize, which was our main goal from the beginning.", "speaker": "Omar Hashash"}, {"start": 1532.319, "end": 1536.364, "text": "What we want is an agent that generalizes in unforeseen scenarios that we haven't seen before.", "speaker": "Omar Hashash"}, {"start": 1536.965, "end": 1539.388, "text": "And in that case, by definition,", "speaker": "Omar Hashash"}, {"start": 1539.655, "end": 1547.154, "text": " Surviving in the world is the main facet of a system, of an AI agent, basically, that generalizes.", "speaker": "Omar Hashash"}, {"start": 1548.478, "end": 1552.568, "text": "So by guaranteeing survival, you're actually guaranteeing that the agent generalizes.", "speaker": "Omar Hashash"}, {"start": 1554.303, "end": 1563.952, "text": " So, roughly speaking, this is what active inferences and what Carl Friston presented it as a way to remove prediction errors from the brain.", "speaker": "Omar Hashash"}, {"start": 1563.992, "end": 1566.234, "text": "Of course, we always have these sensory information.", "speaker": "Omar Hashash"}, {"start": 1567.635, "end": 1574.722, "text": "We're seeing if we don't have a prediction error, it means that we're in the region that minimizes free energy or basically surprised and we don't need to take any action.", "speaker": "Omar Hashash"}, {"start": 1575.122, "end": 1581.188, "text": "However, if we do have a prediction error, we will try to act on the world so that we minimize that prediction error.", "speaker": "Omar Hashash"}, {"start": 1583.007, "end": 1612.525, "text": " uh so of course in this way we can see that reasoning to reduce surprise is the minimum amount of reasoning in any in any living system to survive in the world that's the minimum amount of intelligence that you need to have in order to exist in the world over a long time so what we kind of did over here is that we tried to integrate this into what we know in terms of agents that we just don't need them to just survive we want to make them usable so that they can actually perform tasks", "speaker": "Omar Hashash"}, {"start": 1612.91, "end": 1618.963, "text": " So we went back to integrate this with what we had in terms of reinforcement learning and the policies that we had.", "speaker": "Omar Hashash"}, {"start": 1619.785, "end": 1631.29, "text": "So what happens in this case is that what we saw is that if we take in states and if those states are somehow unforeseen, it means that we can go reason", "speaker": "Omar Hashash"}, {"start": 1633.16, "end": 1637.007, "text": " in order to remove the prediction error that results from this unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 1637.688, "end": 1651.995, "text": "And at the same time, if we can use that test time scaling, we can use it to alter the actions to fit the specific scenario in order to make it optimal in that specific scenario so that the agent can actually generalize in that specific moment.", "speaker": "Omar Hashash"}, {"start": 1652.38, "end": 1666.356, "text": " So roughly speaking, if you want to put it down graphically over here in the figure, what we were doing with reinforcement learning is that we were looking at those states that are highly probable and trying to focus on those states.", "speaker": "Omar Hashash"}, {"start": 1667.602, "end": 1669.824, "text": " finding a policy that fits those states.", "speaker": "Omar Hashash"}, {"start": 1670.105, "end": 1672.187, "text": "This is basically what reinforcement learning is doing.", "speaker": "Omar Hashash"}, {"start": 1673.148, "end": 1676.792, "text": "And after that, what we did is that we froze the world.", "speaker": "Omar Hashash"}, {"start": 1676.812, "end": 1678.374, "text": "We said that the world is not going to change.", "speaker": "Omar Hashash"}, {"start": 1678.954, "end": 1682.278, "text": "So we're always going to be stuck in those states over there and those policies.", "speaker": "Omar Hashash"}, {"start": 1682.338, "end": 1684.581, "text": "But we know that that fades in the real world.", "speaker": "Omar Hashash"}, {"start": 1685.061, "end": 1691.348, "text": "What's actually happening in the real world is that because it's non-stationary, it will probably push you to a state", "speaker": "Omar Hashash"}, {"start": 1691.328, "end": 1694.413, "text": " That is not a problem for you.", "speaker": "Omar Hashash"}, {"start": 1694.573, "end": 1700.343, "text": "It's not a part of the set that you've seen in your training distribution.", "speaker": "Omar Hashash"}, {"start": 1701.285, "end": 1714.006, "text": "What you need to do with that type of reasoning to minimize prediction errors is actually what pushes you back or scales you back to your characteristic set, which is the set of states where you can use your policy again.", "speaker": "Omar Hashash"}, {"start": 1714.467, "end": 1717.352, "text": "So this is basically what we mean by test time scaling.", "speaker": "Omar Hashash"}, {"start": 1718.665, "end": 1734.702, "text": " In this case, we can see, and this is a very nice viewpoint on the story, which is that every living system, in this case, shows a general objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1736.324, "end": 1742.59, "text": "And specifically, I'm saying here something general, because this is something that all living systems perform in the world.", "speaker": "Omar Hashash"}, {"start": 1743.191, "end": 1747.255, "text": "So the fact that you are a living system, it means that you have an objective to survive in the world.", "speaker": "Omar Hashash"}, {"start": 1748.416, "end": 1768.348, "text": " or else you would die so in that case we are governed by general objectives like survival over here and at the same time we have narrow objectives which allow us to achieve our specific tasks and goals with subsumed within that specific region of survival that we have", "speaker": "Omar Hashash"}, {"start": 1768.953, "end": 1775.687, "text": " So I'll just give here an example, just like the jaywalking example of here, just to say how this system kind of works.", "speaker": "Omar Hashash"}, {"start": 1776.509, "end": 1782.06, "text": "So basically, what we have is the fact that I have an autonomous driving vehicle, for example.", "speaker": "Omar Hashash"}, {"start": 1782.141, "end": 1783.283, "text": "It's driving.", "speaker": "Omar Hashash"}, {"start": 1783.482, "end": 1785.004, "text": " And everything's fine.", "speaker": "Omar Hashash"}, {"start": 1785.104, "end": 1786.365, "text": "I have no prediction error.", "speaker": "Omar Hashash"}, {"start": 1786.746, "end": 1788.988, "text": "This is something that we call passive influence in that case.", "speaker": "Omar Hashash"}, {"start": 1789.108, "end": 1791.051, "text": "So we're just taking in sensory observations.", "speaker": "Omar Hashash"}, {"start": 1791.091, "end": 1792.953, "text": "Everything is working as expected.", "speaker": "Omar Hashash"}, {"start": 1793.433, "end": 1794.194, "text": "I have no problem.", "speaker": "Omar Hashash"}, {"start": 1794.935, "end": 1799.801, "text": "And then suddenly what I have is that an agent that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1800.241, "end": 1801.743, "text": "Sorry, a human that decides to jaywalk.", "speaker": "Omar Hashash"}, {"start": 1803.144, "end": 1812.595, "text": "In that case, the network, of course, the network is acting here as an additional computing resource in addition to the agent that we have.", "speaker": "Omar Hashash"}, {"start": 1813.3, "end": 1815.845, "text": " you don't need to focus a lot on what the network is doing.", "speaker": "Omar Hashash"}, {"start": 1815.865, "end": 1818.269, "text": "The network is similar to the agent in that case.", "speaker": "Omar Hashash"}, {"start": 1818.75, "end": 1820.894, "text": "They're kind of one system together.", "speaker": "Omar Hashash"}, {"start": 1822.396, "end": 1830.972, "text": "So in this case, the network that we have will detect that there's a prediction error, and it will switch from passive to active inference about the world.", "speaker": "Omar Hashash"}, {"start": 1831.508, "end": 1836.993, "text": " So what that means is that the network will engage in reasoning in order to minimize the surprise.", "speaker": "Omar Hashash"}, {"start": 1837.554, "end": 1846.143, "text": "And it's going to use those digital twins that I talked about at the beginning, which capture the model of the agent and the work model itself in order to reason.", "speaker": "Omar Hashash"}, {"start": 1846.543, "end": 1854.511, "text": "And what we want to do is that we want to reason, just like I was talking about, I was giving an example here, should I move forward faster?", "speaker": "Omar Hashash"}, {"start": 1854.531, "end": 1855.492, "text": "Should I move slower?", "speaker": "Omar Hashash"}, {"start": 1855.552, "end": 1861.498, "text": "So all that thinking together, we're going to use it in order to put down a value function.", "speaker": "Omar Hashash"}, {"start": 1862.255, "end": 1873.032, "text": " So the value function here, we call it as a delta v. So basically, I'm trying to measure which action allows me to resolve my prediction error in that specific state that I am in.", "speaker": "Omar Hashash"}, {"start": 1873.052, "end": 1881.165, "text": "And then what I want to do is that I'm going to take in that feedback from the network and send it back to the physical AI agent.", "speaker": "Omar Hashash"}, {"start": 1881.566, "end": 1882.828, "text": "The physical AI agent", "speaker": "Omar Hashash"}, {"start": 1883.804, "end": 1886.367, "text": " What it's going to do is that it's going to use this feedback.", "speaker": "Omar Hashash"}, {"start": 1886.888, "end": 1891.374, "text": "It's like the prefrontal cortex sending the signals down to the basal ganglia.", "speaker": "Omar Hashash"}, {"start": 1891.875, "end": 1895.84, "text": "So the policy is over here at the level of the agent of the vehicle here.", "speaker": "Omar Hashash"}, {"start": 1896.301, "end": 1899.986, "text": "And it's going to use it in order to reason and do inference.", "speaker": "Omar Hashash"}, {"start": 1900.006, "end": 1904.412, "text": "So it's going to scale its policy from pi node to pi node prime, for example, here.", "speaker": "Omar Hashash"}, {"start": 1904.712, "end": 1909.0, "text": " So in that specific case, the action that it's going to take is going to be different than its normal.", "speaker": "Omar Hashash"}, {"start": 1909.44, "end": 1916.193, "text": "It's going to be possibly to decrease its velocity in order to avoid an accident, for instance.", "speaker": "Omar Hashash"}, {"start": 1916.213, "end": 1927.192, "text": "And then what happens is that after this jaywalking pedestrian just passes, I go back to my regular cases where I have my prediction error resolved and the pedestrian passes.", "speaker": "Omar Hashash"}, {"start": 1927.232, "end": 1929.336, "text": "So I can go back just to my passive inventory.", "speaker": "Omar Hashash"}, {"start": 1930.227, "end": 1943.289, "text": " So what you can see here is that our world is basically a transition always between foreseen scenarios, unforeseen scenarios, resolving these unforeseen scenarios, and then going back to foreseen scenarios.", "speaker": "Omar Hashash"}, {"start": 1943.57, "end": 1945.152, "text": "This is how we survive in the world.", "speaker": "Omar Hashash"}, {"start": 1945.753, "end": 1948.438, "text": "We're always transitioning between those different states.", "speaker": "Omar Hashash"}, {"start": 1949.937, "end": 1953.542, "text": " So now I'm going to go into the system model.", "speaker": "Omar Hashash"}, {"start": 1954.222, "end": 1957.887, "text": "So as I said before, we have an agent that is in the world.", "speaker": "Omar Hashash"}, {"start": 1958.768, "end": 1959.97, "text": "It has a policy panel.", "speaker": "Omar Hashash"}, {"start": 1960.891, "end": 1964.275, "text": "And we have a world model that exists over the network in this architecture.", "speaker": "Omar Hashash"}, {"start": 1965.136, "end": 1971.945, "text": "One part of the world model is composed from these digital tools, these different alternative policies that the agent can have.", "speaker": "Omar Hashash"}, {"start": 1971.925, "end": 1973.948, "text": " or models of the agent.", "speaker": "Omar Hashash"}, {"start": 1974.749, "end": 1984.383, "text": "At the same time, we have other things that are called assets which capture the different things that exist in the world, just like the human, the houses, everything else besides the agents themselves.", "speaker": "Omar Hashash"}, {"start": 1985.344, "end": 1988.269, "text": "And what's going to happen is that we're going to take sensing observations.", "speaker": "Omar Hashash"}, {"start": 1988.849, "end": 1993.877, "text": "So it's not the agent itself that's doing the sensing, it's actually the network that's doing the sensing on behalf of the agent.", "speaker": "Omar Hashash"}, {"start": 1994.978, "end": 1997.282, "text": "And what we're going to do is that we're going to do perception.", "speaker": "Omar Hashash"}, {"start": 1997.422, "end": 2000.987, "text": "So we're basically going to map our observations to", "speaker": "Omar Hashash"}, {"start": 2001.49, "end": 2009.058, "text": " the set of states that we have trying to make sense of the states, trying to infer what's going to happen over here in terms of the states.", "speaker": "Omar Hashash"}, {"start": 2009.538, "end": 2018.588, "text": "And then once we have a surprise, so we have an unforeseen scenario, what's going to happen is that the network is going to engage in counterfactual reasoning to minimize surprise.", "speaker": "Omar Hashash"}, {"start": 2019.289, "end": 2025.575, "text": "So it's going to think about all these different alternative routes with the different policies that they may have over here, different alternative policies.", "speaker": "Omar Hashash"}, {"start": 2026.496, "end": 2031.261, "text": "And then, as I said, it's going to send back these digital twin configurations.", "speaker": "Omar Hashash"}, {"start": 2032.405, "end": 2038.804, "text": " back to the agent in order to scale the policy and generalize in the unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2039.577, "end": 2043.742, "text": " So mathematically, we can look at perception, of course, as an inference process.", "speaker": "Omar Hashash"}, {"start": 2043.802, "end": 2051.032, "text": "But to do that, we have to first define our generative world model in terms of the states and observations and the policies that we have.", "speaker": "Omar Hashash"}, {"start": 2052.494, "end": 2060.764, "text": "Of course, before we do perception, we have to predict what the expected observation is going to be.", "speaker": "Omar Hashash"}, {"start": 2061.465, "end": 2064.048, "text": "And then we're going to get our sensing observations from the world.", "speaker": "Omar Hashash"}, {"start": 2064.509, "end": 2066.952, "text": "And then we're going to measure if there's a surprise or not.", "speaker": "Omar Hashash"}, {"start": 2066.932, "end": 2070.28, "text": " If we do have a surprise, of course, this is how we measure the surprise.", "speaker": "Omar Hashash"}, {"start": 2070.32, "end": 2071.644, "text": "It's just the easiest surprise.", "speaker": "Omar Hashash"}, {"start": 2073.127, "end": 2082.19, "text": "And after that, if we do have a surprise, it means that we're going to cross some threshold epsilon in order to signal that there's an unforeseen scenario in front of us.", "speaker": "Omar Hashash"}, {"start": 2082.17, "end": 2105.005, "text": " so the first thing that should happen in that case is that we should actually calculate the posterior because basically what we predicted about the world is not right we have an error so we have the first thing we have to do is to make sense of our world again by calculating this posterior after we saw our observations which is basically an inference process of it just we can see before we engage in plan and reason", "speaker": "Omar Hashash"}, {"start": 2104.985, "end": 2111.657, "text": " So, the planning part, it constitutes these digital twins and work models that engage in counterfactual reasoning.", "speaker": "Omar Hashash"}, {"start": 2111.697, "end": 2117.447, "text": "So, basically, the what-if analysis that I was giving an example about, in order to plan the plausible future world states.", "speaker": "Omar Hashash"}, {"start": 2117.647, "end": 2122.656, "text": "And the goal, of course, of this reasoning is to actually minimize future surprise.", "speaker": "Omar Hashash"}, {"start": 2122.957, "end": 2128.742, "text": " So our goal is to see how much surprise is going to result from each of these policies over here.", "speaker": "Omar Hashash"}, {"start": 2128.762, "end": 2135.889, "text": "And you can see that it's going to be nothing but a measure of the entropy at each of those states in the future.", "speaker": "Omar Hashash"}, {"start": 2135.969, "end": 2140.933, "text": "So from t plus 1 until the end of the time horizon that we have.", "speaker": "Omar Hashash"}, {"start": 2140.953, "end": 2149.2, "text": "And then after that, what we're going to do is that we're going to use this reasoning that we did in order to build a value function.", "speaker": "Omar Hashash"}, {"start": 2149.721, "end": 2150.922, "text": "We call it delta v.", "speaker": "Omar Hashash"}, {"start": 2152.252, "end": 2154.935, "text": " And where the reward is actually minimizing surprise.", "speaker": "Omar Hashash"}, {"start": 2154.955, "end": 2160.563, "text": "So as I said at the beginning, we have a general objective in the brain, which is to survive.", "speaker": "Omar Hashash"}, {"start": 2161.364, "end": 2164.348, "text": "So basically it has a value function over here.", "speaker": "Omar Hashash"}, {"start": 2164.368, "end": 2174.901, "text": "And then what we're going to do is that we're going to take this value and we're going to send it back to the agents in the world in order to scale the policy.", "speaker": "Omar Hashash"}, {"start": 2175.37, "end": 2196.602, "text": " how that looks like it's a little bit similar to what perception did so we have the perception as inference of course planning is also as inference over here because planning is an extension of minimizing surprises in the future so it's also planning as inference and now we have also we have to model action as inference but here it's a different type of conference", "speaker": "Omar Hashash"}, {"start": 2196.582, "end": 2212.228, "text": " the fact that when we did perception as influence the evidence that happened in front of us was 100 true it happened in front of us in the world so we're confident about it but when you want to do action as influence it's a little bit different", "speaker": "Omar Hashash"}, {"start": 2212.208, "end": 2217.479, "text": " So it's based on imaginary virtual data in our brain, right?", "speaker": "Omar Hashash"}, {"start": 2217.499, "end": 2220.204, "text": "When we were thinking, we were actually generating data.", "speaker": "Omar Hashash"}, {"start": 2220.445, "end": 2221.888, "text": "That thing didn't happen in the world.", "speaker": "Omar Hashash"}, {"start": 2222.108, "end": 2223.471, "text": "It happened inside our brain.", "speaker": "Omar Hashash"}, {"start": 2224.152, "end": 2230.926, "text": "But luckily, we have Judea Pearl, who came up with this method of virtual evidence.", "speaker": "Omar Hashash"}, {"start": 2231.446, "end": 2236.833, "text": " So he has a nice technique in order to update the beliefs in that case.", "speaker": "Omar Hashash"}, {"start": 2237.474, "end": 2242.48, "text": "So we can see here, if we write it down, this is going to be something called a soft Bayesian update.", "speaker": "Omar Hashash"}, {"start": 2243.161, "end": 2249.629, "text": "So it's soft because you don't really need to do a complete inference update over here.", "speaker": "Omar Hashash"}, {"start": 2250.27, "end": 2259.462, "text": "So if you write it down, you can see here at the last line, you can write it in terms of a feed forward and reasoning or inference term.", "speaker": "Omar Hashash"}, {"start": 2259.83, "end": 2260.891, "text": " which is something very nice.", "speaker": "Omar Hashash"}, {"start": 2261.211, "end": 2275.066, "text": "Because if you really look at it, if you put that theta, which is basically your normalized surprise, if the world is mostly predictable, so your theta or surprise is basically equal to 0.", "speaker": "Omar Hashash"}, {"start": 2275.647, "end": 2284.116, "text": "And if you put that exponential part over here equal to 0, you basically result in the feedforward policy that we had in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2285.142, "end": 2289.269, "text": " The fact is that action is always an inference process.", "speaker": "Omar Hashash"}, {"start": 2289.569, "end": 2293.195, "text": "It's always modulated by that exponential term that we have over here.", "speaker": "Omar Hashash"}, {"start": 2293.676, "end": 2299.746, "text": "And this is what scales the policy once we have prediction errors and once we have an unforeseen scenario.", "speaker": "Omar Hashash"}, {"start": 2300.427, "end": 2311.065, "text": "So the policies in today's solution, they remain limited to the case where surprise is null due to the stationary assumptions about the world, just like we do in reinforcement learning.", "speaker": "Omar Hashash"}, {"start": 2311.045, "end": 2314.932, "text": " So reinforcement learning is not stationary by itself at this time.", "speaker": "Omar Hashash"}, {"start": 2315.032, "end": 2323.968, "text": "We made it stationary because we did not compensate for this factor over here in equation 5, which is clearly a special case of the equation.", "speaker": "Omar Hashash"}, {"start": 2325.03, "end": 2329.678, "text": "However, the inference at this part can be challenging if you want to really solve the problem.", "speaker": "Omar Hashash"}, {"start": 2330.03, "end": 2335.64, "text": " It can be challenging because marginalizing over all the possible states is computationally intractable in practice.", "speaker": "Omar Hashash"}, {"start": 2336.341, "end": 2341.29, "text": "So to find a better solution, we need to resort to a variational Bayesian inference solution.", "speaker": "Omar Hashash"}, {"start": 2341.53, "end": 2347.06, "text": "And I'll keep it here for Christo to present the solution.", "speaker": "Omar Hashash"}, {"start": 2349.625, "end": 2352.51, "text": "I think I'll stop sharing for Christo to share.", "speaker": "Daniel Friedman"}, {"start": 2373.854, "end": 2374.595, "text": " Thanks, Omar.", "speaker": "Daniel Friedman"}, {"start": 2376.958, "end": 2380.243, "text": "So I will go over the solution roadmap first.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2381.305, "end": 2387.193, "text": "As Omar mentioned, our solution is grounded in variational variation inference principle.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2388.815, "end": 2401.954, "text": "So this evolves as a four-step cognitive loop, starting with perception, which is about perceiving the current state by minimizing a variational free energy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2402.474, "end": 2420.297, "text": " and then uh you know we go to the future time steps where we you know estimate the expected free energy static from the inferred state and then uh you know infer policy for actions that minimize the future surprise so that's the planning phase", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2420.277, "end": 2437.392, "text": " and then in the action phase uh so there is we introduced the concept of marco blanket which acts as a statistical boundary between agent and the environment and then policy scaling will be more modeled as a gradient descent on the free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2437.372, "end": 2453.727, "text": " um and finally okay we incorporate this uh updated policy as well as uh the land posterior distributions about the states into you know updating the world model that so that again happens via vfe minimization so", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2453.707, "end": 2464.773, "text": " These four stages is grounded on the same or minimizing the same variation free energy concept.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2465.174, "end": 2467.199, "text": "That's about the solution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2467.86, "end": 2469.825, "text": "Then first, let's look at the perception.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2469.805, "end": 2479.079, "text": " So perception is modeled as inference, inferring the posterior distribution of the states S given the observations.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2479.8, "end": 2494.362, "text": "But if you expand it in terms of the Bayes rule, you can see that this is interactable, computationally interactable because of the marginal distribution of the observations which appear in the denominator here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2495.017, "end": 2509.87, "text": " so here we move to a to find a tractable posterior distribution we actually uh formulate the kl divergence between the approximate posterior q of s comma pi zero", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2509.917, "end": 2536.633, "text": " and the KL divergence to the actual or the true posterior so that's that's the concept of variational inference so where if you actually expand this KL divergence we can see that it can be written as a variation free energy times variation of free energy as well as an additional yeah plus or so basically variation free energy can be written as a KL divergence plus a surprise time which is ln p of time here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2536.985, "end": 2566.772, "text": " uh and so minimizing the scale divergence become equal to uh yeah minimizing variation and then okay once we actually once we actually find a posterior that minimizes the kl divergence along with that we all we are also computing or we are also estimating the surprise quantity which is here uh so that's the advantage of this variational variation inference formulation here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2567.528, "end": 2591.494, "text": " um yeah so yeah so once we find an optimal or optimal posterior distribution that actually means that uh the resetting variation free energy is equal to the surprise quantity uh so which surprise the estimation of the surprise is actually needed for the policy scaling that uh homer mentioned in the previous slide", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2593.162, "end": 2597.248, "text": " So now the question is how we actually go about solving this.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2597.608, "end": 2606.241, "text": "So this can be modeled as a sequential inference process, which can be formulated as a partially observed Markov decision process as shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2606.281, "end": 2610.587, "text": "And this leads to the following factor graph as is shown here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2611.048, "end": 2617.617, "text": "So where we can see that, okay, so we actually modeled the transition from the state's", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2617.597, "end": 2620.328, "text": " to the observations using the matrix A here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2620.389, "end": 2626.695, "text": "So we actually model all these transition probabilities as categorical distributions.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2626.827, "end": 2654.252, "text": " and we use the matrix b to actually model the transition from the previous state to the current state st uh yeah so now optimizing the or minimizing the variation of free energy this leads to an expression as is shown here uh so okay this basically leads to variational message passing uh wherein we can see that okay so there is the message one which is basically uh accounting for the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2654.232, "end": 2675.38, "text": " uh yeah basically a forward transition which basically depends upon the forward transition probability and message 2 which depends upon the backward transition probability which is the message 2 component here and then message 3 actually uh incorporates the uh observation likelihood distribution", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2676.136, "end": 2689.173, "text": " And further, in order to compute the posterior distribution, which is basically defined as the s variable here, we basically take a softmax over the logarithm tense here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2689.694, "end": 2695.461, "text": "So that's how we actually obtain the ln q of s or the posterior distribution here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2696.42, "end": 2721.668, "text": " and uh so this posterior distribution can be shown to be okay so as we mentioned we started with minimizing the variation free energy component so at converge at the equilibrium or yeah at the convergence this um optimal posterior distribution should um converge to that of the true posterior so it basically is minimizing the um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2722.475, "end": 2745.208, "text": " variational free energy component so basically it's a gradient descent on the um free energy um component so that's what basically this uh yeah perception component basically shows and we can actually show that this uh variational free energy can be written to be uh okay the posterior times an error on the", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2745.188, "end": 2757.477, "text": " uh posterior posterior distribution so basically minimizing the variation of free energy is equivalent to minimizing the error on the estimated posterior or inferred posterior distributions", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2759.262, "end": 2779.414, "text": " and yeah so now uh when we go to planning so what we actually do is that uh so we look at future time steps and we we want to find the policy that minimizes the average free energy uh across time so that's basically so so for that we need expected free energy component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2779.394, "end": 2797.639, "text": " and we show that this expected free energy component can be nicely decomposed into two components so one is the risk component which actually uh measures the kl divergence between uh the you know approximate uh oh basically the inferred um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2798.615, "end": 2820.285, "text": " distribution of observations given the current policy how that so the inferred uh distribute inferred observational distribution how that is actually comparing to the uh expected or preferred observations or the preferences that we have uh for the agent so that's basically the risk component", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2820.265, "end": 2845.429, "text": " plus there is an ambiguity component that uh so basically basically it's a entropy of the probability distributions which is basically the likelihood probability distribution here which so basically if you want to minimize the expected free energy you should make sure that the inferred world model is actually close to that or inferred observations are basically close to that of the preferred observations as well as", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2845.409, "end": 2855.234, "text": " we should try to reduce the ambiguity in terms of the observations that the agent observe in future time steps.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2855.314, "end": 2858.001, "text": "So that's basically the expected free energy component.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2859.297, "end": 2877.12, "text": " uh yeah so now uh the the prior over the posterior the the prior over the policies can be written as s of max of uh minus gamma g so we should yeah the policy should ensure that it should always go in the direction of minimizing the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2877.438, "end": 2897.472, "text": " and uh yeah inferred policy uh inferred action policy can be written as um yeah summation over all the policies pi uh where yeah corresponding to action a times the the prior distribution of the particular policy pi here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2897.452, "end": 2917.3, "text": " uh and we can yeah so that's that's the inferred policy and then finally as omar mentioned in the previous slides what we actually do is using this inferred policy so this inferred policy basically uh needs an estimate of the surprise so so basically inferred policy depends upon the expected free energy", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2917.28, "end": 2921.485, "text": " which basically is equivalent to the surprise quantity, as I mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2921.525, "end": 2930.417, "text": "So basically, you cannot compute an exact value for the inference policy, but you can only estimate using the surprise quantity.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2930.897, "end": 2940.79, "text": "So to estimate the surprise, we need to perform reasoning.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2940.81, "end": 2945.616, "text": "So how much reasoning that you need to do depends upon this parameter called theta here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2945.596, "end": 2952.41, "text": " So that is similar to that of the test time computing LLMs like Omar mentioned.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2952.43, "end": 2959.504, "text": "So that's how we perform the planning and do the test time scaling.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2961.273, "end": 2968.691, "text": " Yeah, so now the question is whether all these computations are really stable or not.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2969.292, "end": 2972.821, "text": "For that, we need to look at the concept of Markov blanket.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2973.021, "end": 2978.033, "text": "So an AI agent that generalizes is a random dynamical system", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2978.013, "end": 2997.574, "text": " that exist over time so for the a for this ai agent to be able to generalize across unforeseen scenarios so there should exist a marco blanket which basically means that uh so marco blanket is consisting of those sensors okay so the agent interacts with the environment through the sensory state", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 2997.723, "end": 3019.933, "text": " and so so basically environment influences the agent through the sensory state and the agent influences the environment through the active state so this active so this marco blanket compose which is composed of the active states as well as the sensory state uh that is something which you know separates the internal agent state from the external states", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3020.369, "end": 3046.14, "text": " yeah so that's how this the closed loop of planning um perception planning action and sensing happens and the key equation here is that so we can write the states x the flow for the state x just composed of this marker blanket states as shown in as is shown in here which is composed of two temps so one is the one is a great dissipative gradient time and another is the solenoid", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3046.12, "end": 3053.953, "text": " uh time which basically both of this time depends upon the log of the model evidence which is ln p of x here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3054.018, "end": 3078.979, "text": " uh so yeah for the agent to exist over time so there should be a mod uh marco blanket which also means that it should actually maximize the this log p of x quantity or the model evidence quantity which is basically equal to minimizing a free and i mean minimizing uh my uh l minus ln p of x which is basically corresponding to uh free energy estimate", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3078.959, "end": 3101.592, "text": " uh yeah so that's what basically this um concept of marco blanket tells us and also this also gives us a principled criterion for principle stopping criterion which basically means that uh so when do when do the agent actually stop reasoning so that can be written like uh yeah that can be written as depending upon this quantity called theta that i mentioned", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3101.572, "end": 3107.6, "text": " So yeah, theta basically decides how much of reasoning that the agent should solve.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3107.62, "end": 3118.954, "text": "So the stopping criteria is written as like one minus theta times the KL divergence between the feedforward policy as well as the inferred or the tested time scaled policy.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3120.176, "end": 3123.78, "text": "So yeah, so that's basically about the stopping criteria.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3124.762, "end": 3129.688, "text": "Yeah, so now we talked about perception planning and action policies.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3130.293, "end": 3138.884, "text": " So now we have to incorporate this land policies or the land posterior distributions into the world model.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3139.365, "end": 3150.84, "text": "For that, we can actually write the generative world model as is shown here, which is composed of the joint distribution of observation state policy as well as the A and B parameters here.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3150.86, "end": 3153.203, "text": "We can actually write this joint distribution.", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3153.964, "end": 3158.089, "text": "And then, yeah, so how do we actually compute this?", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3158.423, "end": 3186.237, "text": " distributions a and b so if you do variation inference we can actually write the posterior distributions for these components a as is shown here so basically what we do is yeah in the variation free energy expression we can actually filter out all those stems that are independent of a then if you actually consider this prior distribution of a to be a Dirichlet distribution so we choose Dirichlet because", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3186.217, "end": 3214.654, "text": " so as i mentioned i mentioned that this your a a is actually a categorical distribution so in order for the posterior to be tractable uh we need to find a conjugate distribution so that's the reason why we basically choose this dirichlet distribution so traditionally distribution can actually be parameterized by this quantity called alpha ij here and we actually can find a closed form expression for this alpha ij", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3214.634, "end": 3230.787, "text": " um yeah a close to home expression for this posterior distribution as is shown here so as you can see here um so what basically this expression does is that it basically okay basically counts the number of joint occurrences of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3230.767, "end": 3254.074, "text": " uh the state um j as well as the the particular observation auto equal to i so basically but yeah so basically as we get more evidence about the world observations so we actually basically update the observation model based on that so that's what basically this observation model posterior inference does similarly", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3254.054, "end": 3282.508, "text": " uh yeah i won't go into the details of these things yeah similarly we actually compute the posterior distribution for the transition model b again we use the same procedure of minimizing the variation inference and we also assume that the b also follows the dirichlet prior distribution which basically leads to a tractable posterior distribution as is shown here here again we can see that basically what we do is basically counting the number of co-occurrence of", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3282.488, "end": 3309.711, "text": " uh s store minus one and s store so that's what basically this distribution does uh so this basically leads to uh okay given that um yeah so and given that we also assume that the states um corresponding to different um physical assets are independent in the posterior distribution we can actually write the posterior distribution as a scaled factor as is shown here", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3309.961, "end": 3336.711, "text": " uh and yeah so so basically this b i j times basically compute the ij value of this b matrix uh and finally we also compute the posterior distribution for the policy again considering that the prior distribution for the policy follows a dirichlet distribution leading to a tractable posterior distribution as i mentioned uh previously so that's that's how we basically", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3336.691, "end": 3364.504, "text": " um sold the transition model observation and the policy uh for the world model updates so uh to conclude uh the entire these four stages uh perception planning action policy and finally world model update occurred by minimizing a single quantity called variational free energy uh yeah so that's that's that's about the solution so i will um", "speaker": "Christo Kurisummoottil Thomas"}, {"start": 3365.699, "end": 3367.428, "text": " give it back to Omar now.", "speaker": "Daniel Friedman"}, {"start": 3372.895, "end": 3373.498, "text": "Okay.", "speaker": "Daniel Friedman"}, {"start": 3381.393, "end": 3384.397, "text": " OK, thank you, Christo, for the solution.", "speaker": "Daniel Friedman"}, {"start": 3384.917, "end": 3395.71, "text": "So as Christo was saying, just to recap both parts of the story, the first part shows how we can actually model everything.", "speaker": "Omar Hashash"}, {"start": 3395.95, "end": 3409.206, "text": "So all the perception, planning, action, everything can be modeled as an inference process in this case, which is mathematically also equivalent to a gradient descent process, basically machine learning.", "speaker": "Omar Hashash"}, {"start": 3409.186, "end": 3425.344, "text": " At the same time, we can find tractable solutions through approximate Bayesian inference in order to do the perception, the planning, the action, and the learning at the end to enable this continual type of learning.", "speaker": "Omar Hashash"}, {"start": 3425.945, "end": 3433.033, "text": "So now what we're going to do is that we're going to see some simulation results around the jaywalking.", "speaker": "Omar Hashash"}, {"start": 3433.013, "end": 3453.52, "text": " example that we have been covering since the beginning of the presentation so we're going to simulate now at test time basically our solution which is a test time solution in comparison to other solutions like Q-learning for example and see how they're going to react in that specific scenario", "speaker": "Omar Hashash"}, {"start": 3453.5, "end": 3468.91, "text": " so what's going to happen is that we're going to train uh new solutions we're going to consider the base solution our solution to be a q learning on top of it we have this minimum reasoning active influence capability that we're talking about the reasoning to minimize free energy", "speaker": "Omar Hashash"}, {"start": 3468.89, "end": 3474.46, "text": " At the same time, we have a Q-learning solution that does not have that capability.", "speaker": "Omar Hashash"}, {"start": 3474.941, "end": 3482.855, "text": "So first of all, of course, when the world changes and now we're going to have a jaywalking pedestrian that the agent has not seen before.", "speaker": "Omar Hashash"}, {"start": 3483.115, "end": 3485.7, "text": "So for example, when you have a green traffic light,", "speaker": "Omar Hashash"}, {"start": 3486.575, "end": 3489.198, "text": " it always sees that there's no pedestrian over there.", "speaker": "Omar Hashash"}, {"start": 3489.278, "end": 3494.265, "text": "So it's kind of confident that it needs to move forward and just go towards its goal.", "speaker": "Omar Hashash"}, {"start": 3494.285, "end": 3495.346, "text": "This is what it was trained for.", "speaker": "Omar Hashash"}, {"start": 3495.786, "end": 3502.074, "text": "But at that specific moment, when the jaywalking pedestrian passes, it should have a prediction error.", "speaker": "Omar Hashash"}, {"start": 3502.094, "end": 3506.66, "text": "But the fact that it has no world model on top of its policy, it cannot do it.", "speaker": "Omar Hashash"}, {"start": 3506.68, "end": 3508.983, "text": "So this is exactly what happens here on the left-hand side.", "speaker": "Omar Hashash"}, {"start": 3510.064, "end": 3513.128, "text": "We can see here that the agent starts at distance 3.", "speaker": "Omar Hashash"}, {"start": 3513.148, "end": 3515.451, "text": "It starts to move towards distance 1.", "speaker": "Omar Hashash"}, {"start": 3516.359, "end": 3525.545, "text": " But then what's going to happen is that at step two, the agent, the jaywalking pedestrian appears.", "speaker": "Omar Hashash"}, {"start": 3526.147, "end": 3527.23, "text": "In that case,", "speaker": "Omar Hashash"}, {"start": 3528.34, "end": 3534.049, "text": " What's going to happen, as I said, because it relies just on its policy, it's going to move forward.", "speaker": "Omar Hashash"}, {"start": 3534.249, "end": 3538.896, "text": "And it's going to still accelerate, because this is what it was used to in the beginning.", "speaker": "Omar Hashash"}, {"start": 3539.397, "end": 3542.222, "text": "And it did not realize that its world changed in that case.", "speaker": "Omar Hashash"}, {"start": 3542.803, "end": 3546.368, "text": "So it starts to accelerate, and it crashes on the pedestrian, of course.", "speaker": "Omar Hashash"}, {"start": 3546.388, "end": 3551.496, "text": "This is typical with methods like Q-learning.", "speaker": "Omar Hashash"}, {"start": 3551.476, "end": 3558.445, "text": " But however, if we look at our solution, the scale policy one, we can see here that what happened is that the agent stopped at distance two.", "speaker": "Omar Hashash"}, {"start": 3558.465, "end": 3559.766, "text": "It did not move after that.", "speaker": "Omar Hashash"}, {"start": 3560.367, "end": 3568.718, "text": "So you can see here in the background, there's a change towards a yellow color, which means that the agent is now in a reasoning type of mode.", "speaker": "Omar Hashash"}, {"start": 3569.298, "end": 3573.103, "text": "So what happened is that unlike the Q-learning agent, it actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3573.724, "end": 3575.827, "text": "So the Q-learning agent actually sped up.", "speaker": "Omar Hashash"}, {"start": 3576.487, "end": 3578.53, "text": "This one actually slowed down.", "speaker": "Omar Hashash"}, {"start": 3578.51, "end": 3590.577, "text": " and kept itself in its place until the jaywalking pedestrian actually left, after which the prediction error is resolved and it starts to accelerate again.", "speaker": "Omar Hashash"}, {"start": 3590.618, "end": 3593.464, "text": "Now, that's a scenario we haven't trained it on before.", "speaker": "Omar Hashash"}, {"start": 3593.798, "end": 3602.738, "text": " But the fact that it has a world model and can reason in those specific cases, it can actually figure out a way in order to keep itself surviving.", "speaker": "Omar Hashash"}, {"start": 3603.179, "end": 3611.858, "text": "And of course, after that, it continues after that specific moment from distance two until it reaches its destination at the other end.", "speaker": "Omar Hashash"}, {"start": 3612.715, "end": 3613.877, "text": " behind that intersection.", "speaker": "Omar Hashash"}, {"start": 3614.798, "end": 3625.855, "text": "So looking here at the rewards, at least, if we look at the Q-Learning rewards, of course, what happened over here is that when the agent crashed, it had a very big negative penalty over here, of course.", "speaker": "Omar Hashash"}, {"start": 3627.037, "end": 3630.522, "text": "However, in our case, what happened is that the agent was able", "speaker": "Omar Hashash"}, {"start": 3630.502, "end": 3654.214, "text": " uh to detect that there's something wrong okay there's a surprise so it detected the surprise and you can see it here from the purple line over here so the surprise went above the threshold which is this 0.2 this epsilon threshold and after that it started to reason so you can see here what happened is that the agent is still getting some kind of reward in the background for surviving", "speaker": "Omar Hashash"}, {"start": 3655.544, "end": 3664.114, "text": " And after that, what was happening in this region between the two lines over here, the agent was actually reasoning.", "speaker": "Omar Hashash"}, {"start": 3664.815, "end": 3675.288, "text": "So it was reasoning in order to adapt and scale its policy in order to avoid the crashing scenario that it may face over here, although it lost a bit of its rewards.", "speaker": "Omar Hashash"}, {"start": 3675.929, "end": 3678.011, "text": "So we're going to see this now in a different state.", "speaker": "Omar Hashash"}, {"start": 3678.031, "end": 3682.476, "text": "We're going to compare our solution with a Q-learning solution and a Bayesian RL solution.", "speaker": "Omar Hashash"}, {"start": 3682.516, "end": 3685.5, "text": "Of course, Q-learning is a", "speaker": "Omar Hashash"}, {"start": 3685.48, "end": 3691.39, "text": " is a model-free scenario, model-free solution, and Bayesian RL is a model-based solution.", "speaker": "Omar Hashash"}, {"start": 3691.971, "end": 3694.095, "text": "So in order just to draw contrast with them.", "speaker": "Omar Hashash"}, {"start": 3694.615, "end": 3697.881, "text": "So what happens here is that we're going to train the three scenarios.", "speaker": "Omar Hashash"}, {"start": 3698.222, "end": 3703.27, "text": "We can see here that roughly they have the same training reward.", "speaker": "Omar Hashash"}, {"start": 3703.25, "end": 3708.221, "text": " And then at test time, we're going to let them see the jaywalking pedestrian.", "speaker": "Omar Hashash"}, {"start": 3709.002, "end": 3717.561, "text": "What's going to happen is that the reward will drop massively for Q-learning and Bayesian RL, and also for our solution.", "speaker": "Omar Hashash"}, {"start": 3717.581, "end": 3723.874, "text": "It's going to drop, but not that severe as the Q-learning and Bayesian RL scenarios.", "speaker": "Omar Hashash"}, {"start": 3723.854, "end": 3726.958, "text": " Although it does, it really drops.", "speaker": "Omar Hashash"}, {"start": 3727.059, "end": 3729.562, "text": "However, this comes at the expense of success.", "speaker": "Omar Hashash"}, {"start": 3730.183, "end": 3741.72, "text": "So our agent, in order to scale, it had to sacrifice its narrow rewards in order to ensure the success of its long-term objective, right?", "speaker": "Omar Hashash"}, {"start": 3741.74, "end": 3743.342, "text": "Because this is what RL is about.", "speaker": "Omar Hashash"}, {"start": 3743.362, "end": 3746.166, "text": "It's about maximizing the long-term reward.", "speaker": "Omar Hashash"}, {"start": 3746.534, "end": 3768.39, "text": " uh in order to ensure success while efficiently utilizing inference so what it did over here is that it showed us that you don't actually need to do reasoning and inference all the time you just need to do it when needed so it can be adaptive not like what a lot of people think now that if you add always add reasoning on top of the models it's always better", "speaker": "Omar Hashash"}, {"start": 3768.758, "end": 3775.127, "text": " So the efficient human-like way is that reasoning is only included once you know this is human behavior.", "speaker": "Omar Hashash"}, {"start": 3775.167, "end": 3789.366, "text": "And so it showed around 36% improvement more than above the Bayesian RL solution in terms of utilizing the computing resources in that case.", "speaker": "Omar Hashash"}, {"start": 3789.386, "end": 3795.154, "text": "So it shows a trade-off between model-based and model-free solutions.", "speaker": "Omar Hashash"}, {"start": 3795.775, "end": 3809.017, "text": " And after that, what we did is that we trained the agent that we have with our solution on a jaywalking pedestrian scenario.", "speaker": "Omar Hashash"}, {"start": 3809.037, "end": 3814.426, "text": "But we allowed it to see that scenario again and again in order to see what's going to happen.", "speaker": "Omar Hashash"}, {"start": 3815.007, "end": 3821.277, "text": "So basically what happened in that case is that the first time it saw the scenario, it had a very high surprise.", "speaker": "Omar Hashash"}, {"start": 3821.257, "end": 3845.307, "text": " after that the surprise started to decrease over time because it started to update and started to make sense that i'm always going to see a jaywalking pedestrian at that specific scenario it got used to it basically and you can see it here from the fact that it starts to predict from the other graph you can start to see in the blue graph that it starts to predict that there's a pedestrian whenever i have a green light", "speaker": "Omar Hashash"}, {"start": 3845.759, "end": 3851.13, "text": " So it starts to update the probability of the agent having a pedestrian in that case.", "speaker": "Omar Hashash"}, {"start": 3852.132, "end": 3856.861, "text": "So basically, it becomes less surprising for the agent.", "speaker": "Omar Hashash"}, {"start": 3856.881, "end": 3865.438, "text": "And at the same time, at the level of the policy itself, it also updates the probability of taking that action in that specific state.", "speaker": "Omar Hashash"}, {"start": 3865.621, "end": 3877.632, "text": " So basically what's happening is that every time it's seeing the scenario, it's reinforcing that I should see a jaywalking pedestrian at that green light in the jaywalking scenario that I'm going to see.", "speaker": "Omar Hashash"}, {"start": 3878.052, "end": 3887.841, "text": "And I know very well what my action, that specific action that I was able to get through my reasoning, I know very well what that action should be and it reinforces it in the policy.", "speaker": "Omar Hashash"}, {"start": 3888.542, "end": 3894.447, "text": "So after 50 episodes, what happens is that the decision making moves from being", "speaker": "Omar Hashash"}, {"start": 3894.427, "end": 3915.321, "text": " controlled through inference through reasoning to become feed forward because basically it becomes a part of its work model and the policy after that in order to get it basically it gets reinforced over time at test time so you can see here that even with we have reinforcement learning now at test time and this is basically how the agent can become a continual learning agent", "speaker": "Omar Hashash"}, {"start": 3916.567, "end": 3924.176, "text": " So just to conclude this now, this is just a comprehensive roadmap of wireless communication and AI working together.", "speaker": "Omar Hashash"}, {"start": 3924.216, "end": 3930.504, "text": "Basically, our networks are moving forward very fast, just like AI.", "speaker": "Omar Hashash"}, {"start": 3930.544, "end": 3939.696, "text": "And what we want is that we have this ultimate intersection between the two so that we can reach what we call as AGI native networks in the future.", "speaker": "Omar Hashash"}, {"start": 3939.716, "end": 3943.3, "text": "So we're moving from AI native networks to AGI native networks.", "speaker": "Omar Hashash"}, {"start": 3943.28, "end": 3957.675, "text": " where these things build on the different technologies and systems that we talked about today, the world model, the digital twins, our presented test time scaling, and more broadly, a big cognitive system to architecture of the brain.", "speaker": "Omar Hashash"}, {"start": 3958.376, "end": 3963.721, "text": "And with that, I'll conclude the presentation, and I'll be happy to answer any questions you may have.", "speaker": "Omar Hashash"}, {"start": 3964.642, "end": 3965.503, "text": "Crystal and I, of course.", "speaker": "Daniel Friedman"}, {"start": 3968.426, "end": 3968.787, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 3969.427, "end": 3969.928, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 3972.743, "end": 3975.365, "text": " Yeah, just one short personal comment.", "speaker": "Daniel Friedman"}, {"start": 3975.505, "end": 3983.172, "text": "I think this really builds nicely from Christo's previous presentation, which was about the semantic information theory.", "speaker": "Daniel Friedman"}, {"start": 3983.192, "end": 4002.75, "text": "And this takes it and really connects it quite well to the passive, active planning, online learning setting and how those features are being implemented in a functional sense by the different recent generations of LLM.", "speaker": "Daniel Friedman"}, {"start": 4002.73, "end": 4010.245, "text": " and sort of associated systems like the skills, tools, sub-agent dispatch, chain of thought, all those kinds of things.", "speaker": "Daniel Friedman"}, {"start": 4011.808, "end": 4012.069, "text": "All right.", "speaker": "Daniel Friedman"}, {"start": 4013.732, "end": 4014.854, "text": "Krista, do you want to add anything?", "speaker": "Daniel Friedman"}, {"start": 4015.255, "end": 4016.858, "text": "And then I'll read some questions.", "speaker": "Daniel Friedman"}, {"start": 4023.301, "end": 4023.502, "text": " Okay.", "speaker": "Daniel Friedman"}, {"start": 4023.522, "end": 4024.925, "text": "All right.", "speaker": "Daniel Friedman"}, {"start": 4025.306, "end": 4033.547, "text": "If anyone in live chat asks a question, I'll read it, but I'll start with a few questions that were submitted by email from Joel Robinson.", "speaker": "Daniel Friedman"}, {"start": 4035.89, "end": 4046.562, "text": " Okay, Jill wrote, I've been reading active inference as the test time scaling law for physical AI agents and had several questions that arise from my own work on regulatory architectures.", "speaker": "Daniel Friedman"}, {"start": 4047.323, "end": 4057.934, "text": "I'm asking these because the empirical literature suggests certain variables must exist in any physical or biological agent, and I'm trying to understand how they map onto your formulation.", "speaker": "Daniel Friedman"}, {"start": 4059.155, "end": 4059.636, "text": "So here we go.", "speaker": "Daniel Friedman"}, {"start": 4061.135, "end": 4064.099, "text": " Question one, inference capacity as a state variable.", "speaker": "Daniel Friedman"}, {"start": 4064.84, "end": 4078.016, "text": "You write that the scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time, and that that policy update is modeled as a soft Bayesian inference process.", "speaker": "Daniel Friedman"}, {"start": 4079.838, "end": 4084.324, "text": "In cognitive neuroscience, inference capacity is state dependent and collapses under load.", "speaker": "Daniel Friedman"}, {"start": 4084.878, "end": 4090.288, "text": " So where in your formulation is the agent's capacity to perform inference represented?", "speaker": "Daniel Friedman"}, {"start": 4091.069, "end": 4099.624, "text": "Is inference assumed to be always available, or is there a state variable that modulates whether the agent can actually execute the Bayesian update?", "speaker": "Daniel Friedman"}, {"start": 4100.526, "end": 4103.511, "text": "And I'm going to paste this into the Zoom chat as well.", "speaker": "Daniel Friedman"}, {"start": 4107.322, "end": 4115.953, "text": " Yeah, so I think it's the last part of the question that's really exciting.", "speaker": "Omar Hashash"}, {"start": 4116.414, "end": 4126.087, "text": "So I do agree with the question itself as being there are limitations for the architecture.", "speaker": "Omar Hashash"}, {"start": 4127.469, "end": 4133.717, "text": "But in our case, I think we're considering mostly kind of like optimal scenarios, optimal execution scenarios.", "speaker": "Omar Hashash"}, {"start": 4134.777, "end": 4140.367, "text": " So the thing is that I agree is that I think inference collapses.", "speaker": "Omar Hashash"}, {"start": 4140.847, "end": 4145.575, "text": "I think the word was collapses in certain scenarios.", "speaker": "Omar Hashash"}, {"start": 4145.596, "end": 4156.574, "text": "But in our case, we specifically highlighted the fact here that we are using our system two, sorry, our system one most of the time to take actions.", "speaker": "Omar Hashash"}, {"start": 4157.256, "end": 4160.481, "text": "So in that case, our resources for reasoning", "speaker": "Omar Hashash"}, {"start": 4160.731, "end": 4188.409, "text": " assuming that we have a lot of time possibly not instantaneous reflective reflexive like a reflex while driving that i have enough time to think about the different scenarios that i can take in that specific scenario so i'm not considering here a where like an example where an agent where a human for example needs to take a split second action i'm feeling that there needs to be enough time", "speaker": "Omar Hashash"}, {"start": 4188.389, "end": 4201.842, "text": " so that the agent thinks because assuming i'm on a highway and there's like roughly four five six seconds which i think is more than enough for a human to think in that case at least in the broad sense", "speaker": "Omar Hashash"}, {"start": 4202.21, "end": 4213.169, "text": " So these are the type of problems that we're considering over here, which in that case, I think the inference part doesn't crash specifically in that case or collapses in that case.", "speaker": "Omar Hashash"}, {"start": 4213.189, "end": 4226.091, "text": "But I do acknowledge the fact that there are situations where this kind of has to be all short circuited, I would say, and you just have to pass to just taking an instantaneous reflex.", "speaker": "Omar Hashash"}, {"start": 4228.045, "end": 4254.222, "text": " Yeah, one point I'll add there, and this is often brought up by the RxInfer and sort of message graph developers, is that with the message graph doing local free energy minimization, you can have different frequencies or different number of iterations of optimization between observations.", "speaker": "Daniel Friedman"}, {"start": 4254.202, "end": 4257.747, "text": " You could be doing an observation slower or at a faster frequency.", "speaker": "Daniel Friedman"}, {"start": 4258.188, "end": 4262.534, "text": "You could have intermittent or a fix or a variable compute budget.", "speaker": "Daniel Friedman"}, {"start": 4263.275, "end": 4277.796, "text": "And so that's like an advantage for online methods is that you're kind of just doing as good as you can in an online fashion given the amount of cognitive or computational resourcing available.", "speaker": "Daniel Friedman"}, {"start": 4279.244, "end": 4289.809, "text": " Modulo, anything like a catastrophic failure, that kind of breaks out of the box of the performance on the self-driving car itself breaking down mechanically.", "speaker": "Daniel Friedman"}, {"start": 4289.849, "end": 4291.994, "text": "It's sort of the mortal computing.", "speaker": "Daniel Friedman"}, {"start": 4292.054, "end": 4294.62, "text": "That's like the embodiment side.", "speaker": "Daniel Friedman"}, {"start": 4294.6, "end": 4319.376, "text": " and here is like really scoping out a lot of the formalisms and methods for the computational the cognitive that kind of presuppose the integrity of the functioning machine even though that becomes also a question when you're talking about these like network systems yeah i i i believe you said it in the best way possible and so uh it's uh", "speaker": "Daniel Friedman"}, {"start": 4319.71, "end": 4323.435, "text": " What you raised over here are the perfect times at that specific moment.", "speaker": "Omar Hashash"}, {"start": 4323.876, "end": 4327.781, "text": "So of course, we're working on a strict budget over here, but we're considering that.", "speaker": "Omar Hashash"}, {"start": 4329.303, "end": 4331.265, "text": "So I'm processing it in a different way.", "speaker": "Omar Hashash"}, {"start": 4331.666, "end": 4344.503, "text": "So I think our goal of this work was to lay down the foundation of how can an agent exist in the world in order to solve the problem that", "speaker": "Omar Hashash"}, {"start": 4345.023, "end": 4349.61, "text": " the problem that we were facing or possibly what we were missing before in that case.", "speaker": "Omar Hashash"}, {"start": 4350.692, "end": 4360.568, "text": "Of course, the more practical sense that you were talking about, which is different frequencies that you may be taking in samples in about the world, the different noise", "speaker": "Omar Hashash"}, {"start": 4360.835, "end": 4365.928, "text": " the noises that you may have, how much computationally feasible is that.", "speaker": "Omar Hashash"}, {"start": 4366.068, "end": 4368.515, "text": "I think these things are right on point.", "speaker": "Omar Hashash"}, {"start": 4369.056, "end": 4371.803, "text": "And they should follow exactly from this specific moment.", "speaker": "Omar Hashash"}, {"start": 4371.823, "end": 4375.232, "text": "So I think this is roughly a", "speaker": "Omar Hashash"}, {"start": 4375.887, "end": 4399.026, "text": " a i wouldn't say a sketch but i would say as a initial solution of a problem uh that we can start with and of course on top of that we can add a different uh feasibility type of thing i assume there's a lot of solution that needs to come up in this space more but the tricky part about it was uh as you were saying about the different um", "speaker": "Omar Hashash"}, {"start": 4399.006, "end": 4403.395, "text": " connecting it into the broader AI scenario, like what was really missing.", "speaker": "Omar Hashash"}, {"start": 4403.415, "end": 4407.644, "text": "Like we're seeing these different LLMs, chain of thought, reasoning, planning, agentic.", "speaker": "Omar Hashash"}, {"start": 4408.566, "end": 4412.034, "text": "But we already knew that, like we already had agents before.", "speaker": "Omar Hashash"}, {"start": 4412.054, "end": 4413.156, "text": "So what was actually missing?", "speaker": "Omar Hashash"}, {"start": 4413.176, "end": 4416.002, "text": "What did we really not see before?", "speaker": "Omar Hashash"}, {"start": 4416.269, "end": 4433.012, "text": " And it was actually that fact that I think that the new part in LLMs was mostly based on the test time part more than the regular vanilla kind of scaling the training part, the model size, the parameters.", "speaker": "Omar Hashash"}, {"start": 4432.992, "end": 4442.342, "text": " So we try to focus on that more, try to see why is it really that I have an empirical scaling law with language models, and what's the intersection over there?", "speaker": "Omar Hashash"}, {"start": 4443.183, "end": 4447.768, "text": "So yeah, it was putting down the theoretical part of the story, I think, at first.", "speaker": "Omar Hashash"}, {"start": 4448.229, "end": 4453.575, "text": "But definitely, the next steps should be based on making it more computationally tractable.", "speaker": "Omar Hashash"}, {"start": 4455.016, "end": 4455.317, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 4455.457, "end": 4457.219, "text": "All right, here's another question from Joel.", "speaker": "Daniel Friedman"}, {"start": 4457.86, "end": 4458.24, "text": "He wrote,", "speaker": "Daniel Friedman"}, {"start": 4460.06, "end": 4484.515, "text": " temporal depth and prediction window collapse you note that humans simulate counterfactual scenarios and plan future states of the world and this is central to resolving prediction error empirically temporal depth is not fixed it narrows under stress fatigue or overload is temporal depth treated as a dynamic state variable variable in your model if the prediction window collapse how does the scaling law behave", "speaker": "Daniel Friedman"}, {"start": 4488.19, "end": 4495.268, "text": " So if I got the answer correctly, you were asking about the temporal depth?", "speaker": "Omar Hashash"}, {"start": 4496.872, "end": 4499.078, "text": "Yeah, like the depth of planning.", "speaker": "Daniel Friedman"}, {"start": 4500.577, "end": 4502.921, "text": " OK, yeah.", "speaker": "Daniel Friedman"}, {"start": 4502.981, "end": 4511.875, "text": "So of course, the depth here is, of course, focused on we're having it as the least shallow layer, I would say.", "speaker": "Omar Hashash"}, {"start": 4511.935, "end": 4517.323, "text": "Of course, we're not considering that we have a hierarchical form of the world, of course, which we know about.", "speaker": "Omar Hashash"}, {"start": 4518.385, "end": 4523.513, "text": "And we're not considering the different temporal aspects of the world that can exist over there.", "speaker": "Omar Hashash"}, {"start": 4523.493, "end": 4528.862, "text": " Of course, the next steps should be that this world is composed in a hierarchical way.", "speaker": "Omar Hashash"}, {"start": 4529.262, "end": 4530.645, "text": "There's a lot of structure in there.", "speaker": "Omar Hashash"}, {"start": 4531.166, "end": 4540.982, "text": "At which level of planning should we actually limit ourselves in so that we don't fall into the kind of collapse that Joel's talking about over here?", "speaker": "Omar Hashash"}, {"start": 4541.002, "end": 4548.454, "text": "So I think even there are some interesting parts over here related to hierarchical active inference.", "speaker": "Omar Hashash"}, {"start": 4548.822, "end": 4557.631, "text": " So the hierarchical part of modeling the world, which I assume should be also our next objective over here.", "speaker": "Omar Hashash"}, {"start": 4557.932, "end": 4560.52, "text": "I'm not sure if Christo wants to add a certain part over here.", "speaker": "Daniel Friedman"}, {"start": 4564.972, "end": 4576.104, "text": " Yeah, so I think Christo, because he had some words on hierarchical abstractions and hierarchical representations.", "speaker": "Omar Hashash"}, {"start": 4576.785, "end": 4582.371, "text": "So this is a different part of the depth that we were talking about, about the world model.", "speaker": "Omar Hashash"}, {"start": 4582.391, "end": 4588.498, "text": "So we're talking about the depth of the world in terms of the aspects that it may have hierarchically.", "speaker": "Omar Hashash"}, {"start": 4588.478, "end": 4603.135, "text": " And at the same time, we're talking about the depth in terms of time going millisecond by millisecond, or going in terms of representations that can possibly last longer over time, that don't change in an instantaneous manner.", "speaker": "Omar Hashash"}, {"start": 4604.157, "end": 4610.284, "text": "Yeah, I'll just copy two more questions in from Joel.", "speaker": "Daniel Friedman"}, {"start": 4613.707, "end": 4630.551, "text": " on some related topics these are all great great comments so just so it's all um there the one one area was about generalization and again i think it speaks to this nexus of basically like", "speaker": "Daniel Friedman"}, {"start": 4630.531, "end": 4635.501, "text": " there's what is the computational, like structure learning?", "speaker": "Daniel Friedman"}, {"start": 4635.862, "end": 4639.65, "text": "There still could be novel situations that are outside your structure learning envelope.", "speaker": "Daniel Friedman"}, {"start": 4640.371, "end": 4647.105, "text": "There could be like unknown unknowns, or there could be things that were, or there could even be structural possibilities that aren't adjacencies.", "speaker": "Daniel Friedman"}, {"start": 4647.727, "end": 4650.85, "text": " Then there's the implementation of the computer resources.", "speaker": "Daniel Friedman"}, {"start": 4651.831, "end": 4655.194, "text": "The math is like the computer science is free.", "speaker": "Daniel Friedman"}, {"start": 4656.195, "end": 4666.664, "text": "And then there's these questions about the capacity and how the compute system's capacity re-enters into the test time.", "speaker": "Daniel Friedman"}, {"start": 4666.704, "end": 4675.472, "text": "And what if you develop the test time in one situation and then it has a poor drop-off", "speaker": "Daniel Friedman"}, {"start": 4675.452, "end": 4702.911, "text": " in a compute limited setting but i think that your that um point about this um training objective or constraint being an underappreciated or underutilized factor in large model training is very key and i hope that it comes through yeah so um", "speaker": "Daniel Friedman"}, {"start": 4703.498, "end": 4710.371, "text": " I'll try to hit on the part which you're talking about, the unknown unknowns, those parts.", "speaker": "Omar Hashash"}, {"start": 4710.711, "end": 4717.123, "text": "So of course, there are still room enough for a lot of things to pop up in the world.", "speaker": "Omar Hashash"}, {"start": 4717.143, "end": 4725.118, "text": "And the fact that what we were dealing with the world is we're trying to deal with the part of the world", "speaker": "Omar Hashash"}, {"start": 4726.094, "end": 4737.353, "text": " that we don't really know about so that's the whole premise of the whole story which is the fact that you want to escape from the training grounds and escape into the test time part", "speaker": "Omar Hashash"}, {"start": 4738.126, "end": 4740.21, "text": " Of course, at this time, there's always going to be capacity.", "speaker": "Omar Hashash"}, {"start": 4740.23, "end": 4753.715, "text": "I think we do have, we included some resources in our paper about like, for example, if you have a new scenario that that's actually not explained by none of our states, let's say if you're doing perception.", "speaker": "Omar Hashash"}, {"start": 4754.255, "end": 4757.061, "text": "So in that case, you would need to, let's say, for example, expand your model.", "speaker": "Omar Hashash"}, {"start": 4757.261, "end": 4762.21, "text": "So you need to go through from a Bayesian sense, you would need to look at Bayesian model expansion.", "speaker": "Omar Hashash"}, {"start": 4762.73, "end": 4767.641, "text": " But of course, these have other limitations, and you have to specify the criteria.", "speaker": "Omar Hashash"}, {"start": 4767.661, "end": 4777.363, "text": "Just like we said that we have a prediction error, and we have a certain epsilon, that if we cross that prediction error, this means that you're going to foresee the scenario.", "speaker": "Omar Hashash"}, {"start": 4777.597, "end": 4788.038, "text": " would probably be in a case where you're trying to fit in your certain state of the world into one of the states that you have, but probably none of those states are actually explaining your situation.", "speaker": "Omar Hashash"}, {"start": 4788.198, "end": 4791.925, "text": "So in that case, the most logical thing to say is that this doesn't fit anything.", "speaker": "Omar Hashash"}, {"start": 4791.966, "end": 4793.749, "text": "This is something very new.", "speaker": "Omar Hashash"}, {"start": 4793.729, "end": 4799.375, "text": " you might probably need to say that this is something like a composition of different states together.", "speaker": "Omar Hashash"}, {"start": 4799.395, "end": 4801.037, "text": "I haven't seen this before.", "speaker": "Omar Hashash"}, {"start": 4801.097, "end": 4810.668, "text": "But in our case, let's say we took the simplistic first step about this, which is the fact that just the probability distribution is changing.", "speaker": "Omar Hashash"}, {"start": 4811.509, "end": 4819.939, "text": "So in a very simple scenario that all of us, I think, see roughly in our lifetime, which is the fact that someone is jaywalking.", "speaker": "Omar Hashash"}, {"start": 4819.919, "end": 4821.822, "text": " So this is something very easy to grasp.", "speaker": "Omar Hashash"}, {"start": 4822.062, "end": 4829.594, "text": "And the fact that we already know what a human is, what a traffic light is, what a green light is, we already know the traffic rules.", "speaker": "Omar Hashash"}, {"start": 4830.115, "end": 4831.617, "text": "So this is something very well known.", "speaker": "Omar Hashash"}, {"start": 4832.258, "end": 4840.772, "text": "In that case, we just flip the rules a little bit so that rather than crossing at a red light, you're crossing at a green light to see what the behavior is.", "speaker": "Omar Hashash"}, {"start": 4840.852, "end": 4844.037, "text": "But of course, I'll keep your imagination", "speaker": "Omar Hashash"}, {"start": 4844.017, "end": 4865.805, "text": " to help you in that case to see how much still we have a lot more to deal with in the future but i think that should be the main focus of this whole thing is trying to figure out step by step what are those critical things how can they actually be posed at this time so that we found find a", "speaker": "Omar Hashash"}, {"start": 4866.477, "end": 4869.164, "text": " first principle kind of solution to these systems.", "speaker": "Omar Hashash"}, {"start": 4869.185, "end": 4874.259, "text": "So building on the fact that all of us solve that problem in that certain way.", "speaker": "Omar Hashash"}, {"start": 4875.823, "end": 4877.989, "text": "So I think that should hit directly on that point.", "speaker": "Daniel Friedman"}, {"start": 4880.585, "end": 4896.881, "text": " Yeah, one area that I see this being very relevant to is the distillation and the on-policy distillation, the teacher, the co-learner, all these kind of model transformation and training strategies.", "speaker": "Daniel Friedman"}, {"start": 4898.583, "end": 4910.355, "text": "All different policy training strategies and this as a comparable measure, even if one", "speaker": "Daniel Friedman"}, {"start": 4911.989, "end": 4930.825, "text": " um test time is well it's 30 hours for this long horizon coding agent it's eight hours for this camera image model even though they're totally different domains they're not going to have a domain benchmark that's comparable and then for the models that are multimodal", "speaker": "Daniel Friedman"}, {"start": 4930.805, "end": 4938.658, "text": " or they're used in a multimodal setting, there aren't benchmarks at the domain level that would apply.", "speaker": "Daniel Friedman"}, {"start": 4938.678, "end": 4947.952, "text": "So it ends up being a sort of weakest link question with these systems sometimes.", "speaker": "Daniel Friedman"}, {"start": 4949.956, "end": 4957.808, "text": "I think that connects to the model training strategies, which are upstream of how well these models do.", "speaker": "Daniel Friedman"}, {"start": 4958.851, "end": 4972.725, "text": " And the fact, this is actually the first point that I wanted to talk about, which is you don't really have a benchmark of what's good or bad, which is the fact that you need to figure things out.", "speaker": "Omar Hashash"}, {"start": 4973.165, "end": 4985.938, "text": "In order to do that, there needs to be a first principle kind of way to make sure at least that it's doing the minimum best possible thing that any human could do.", "speaker": "Omar Hashash"}, {"start": 4986.745, "end": 4994.597, "text": " which is the fact of why we're looking at this active inference part, which is putting it as a limiting factor, actually than putting it as an upper bound.", "speaker": "Omar Hashash"}, {"start": 4995.078, "end": 5002.069, "text": "So we're saying that this is the minimum amount of reason that you need to have in order to solve this type of problem.", "speaker": "Omar Hashash"}, {"start": 5002.45, "end": 5005.114, "text": "But this does not mean that it's the only reason.", "speaker": "Omar Hashash"}, {"start": 5005.154, "end": 5012.145, "text": "So you can add on top of that other things and you can just expand as much as possible, but at least,", "speaker": "Omar Hashash"}, {"start": 5012.463, "end": 5024.216, "text": " to exist in the world and to solve problems like the problems that we see with these autonomous vehicles, Teslas, Waymos, to think about those humanoid robots that just crash down and fall.", "speaker": "Omar Hashash"}, {"start": 5024.957, "end": 5029.081, "text": "So we're trying to at least address that type of problem.", "speaker": "Omar Hashash"}, {"start": 5029.982, "end": 5031.724, "text": "But of course, that's bare minimum.", "speaker": "Daniel Friedman"}, {"start": 5035.228, "end": 5035.588, "text": "Awesome.", "speaker": "Daniel Friedman"}, {"start": 5035.969, "end": 5037.07, "text": "Yeah, and the paper.", "speaker": "Daniel Friedman"}, {"start": 5038.248, "end": 5040.19, "text": " really lays out that direction.", "speaker": "Daniel Friedman"}, {"start": 5041.351, "end": 5042.592, "text": "Do you have any last comments?", "speaker": "Daniel Friedman"}, {"start": 5042.712, "end": 5045.774, "text": "Otherwise, I hope this has been a good entry point to the topic.", "speaker": "Daniel Friedman"}, {"start": 5047.436, "end": 5048.056, "text": "Yeah.", "speaker": "Omar Hashash"}, {"start": 5048.797, "end": 5050.498, "text": "It's actually been a pleasure.", "speaker": "Omar Hashash"}, {"start": 5050.518, "end": 5054.222, "text": "I think the questions were really great and exciting.", "speaker": "Omar Hashash"}, {"start": 5055.543, "end": 5057.424, "text": "And I really enjoyed presenting over here today.", "speaker": "Omar Hashash"}, {"start": 5058.345, "end": 5059.146, "text": "Thank you for having us.", "speaker": "Omar Hashash"}, {"start": 5060.367, "end": 5060.727, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 5061.007, "end": 5063.569, "text": "Thank you to authors and to Joel for the questions.", "speaker": "Daniel Friedman"}, {"start": 5064.21, "end": 5065.331, "text": "So, bye.", "speaker": "Daniel Friedman"}, {"start": 5066.932, "end": 5067.152, "text": "Bye-bye.", "speaker": "Daniel Friedman"}, {"start": 5067.172, "end": 5067.453, "text": "Thank you.", "speaker": "Daniel Friedman"}, {"start": 5067.533, "end": 5067.833, "text": "Thank you.", "speaker": "Daniel Friedman"}]}] \ No newline at end of file diff --git a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt index 43c5b9c98..e42d1b4b9 100644 --- a/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt +++ b/data/video/activeinferenceinstitute/GuestStream/GuestStream_128/transcript.txt @@ -2,7 +2,7 @@ Daniel Friedman: Hello, welcome. -SPEAKER_00: +Daniel Friedman: It's July 14th, 2026. We're in active guest stream 128.1 on active inference as the test time scaling law for physical AI agents. @@ -976,7 +976,7 @@ Awesome. Thank you. -SPEAKER_00: +Daniel Friedman: Yeah, just one short personal comment. I think this really builds nicely from Christo's previous presentation, which was about the semantic information theory. @@ -998,7 +998,7 @@ Okay. All right. -SPEAKER_00: +Daniel Friedman: If anyone in live chat asks a question, I'll read it, but I'll start with a few questions that were submitted by email from Joel Robinson. Okay, Jill wrote, I've been reading active inference as the test time scaling law for physical AI agents and had several questions that arise from my own work on regulatory architectures. @@ -1044,7 +1044,7 @@ So these are the type of problems that we're considering over here, which in tha But I do acknowledge the fact that there are situations where this kind of has to be all short circuited, I would say, and you just have to pass to just taking an instantaneous reflex. -SPEAKER_00: +Daniel Friedman: Yeah, one point I'll add there, and this is often brought up by the RxInfer and sort of message graph developers, is that with the message graph doing local free energy minimization, you can have different frequencies or different number of iterations of optimization between observations. You could be doing an observation slower or at a faster frequency. @@ -1104,7 +1104,7 @@ So yeah, it was putting down the theoretical part of the story, I think, at firs But definitely, the next steps should be based on making it more computationally tractable. -SPEAKER_00: +Daniel Friedman: Awesome. All right, here's another question from Joel. @@ -1118,7 +1118,7 @@ Omar Hashash: So if I got the answer correctly, you were asking about the temporal depth? -SPEAKER_00: +Daniel Friedman: Yeah, like the depth of planning. OK, yeah. @@ -1156,7 +1156,7 @@ So we're talking about the depth of the world in terms of the aspects that it ma And at the same time, we're talking about the depth in terms of time going millisecond by millisecond, or going in terms of representations that can possibly last longer over time, that don't change in an instantaneous manner. -SPEAKER_00: +Daniel Friedman: Yeah, I'll just copy two more questions in from Joel. on some related topics these are all great great comments so just so it's all um there the one one area was about generalization and again i think it speaks to this nexus of basically like @@ -1234,7 +1234,7 @@ Daniel Friedman: So I think that should hit directly on that point. -SPEAKER_00: +Daniel Friedman: Yeah, one area that I see this being very relevant to is the distillation and the on-policy distillation, the teacher, the co-learner, all these kind of model transformation and training strategies. All different policy training strategies and this as a comparable measure, even if one @@ -1274,7 +1274,7 @@ Awesome. Yeah, and the paper. -SPEAKER_00: +Daniel Friedman: really lays out that direction. Do you have any last comments? @@ -1294,7 +1294,7 @@ And I really enjoyed presenting over here today. Thank you for having us. -SPEAKER_00: +Daniel Friedman: Thank you. Thank you to authors and to Joel for the questions.