Original article: Emilien Dupont, Arnaud Doucet, Yee Whye Teh. "Augmented Neural ODEs."
Advances in Neural Information Processing Systems 32 (NeurIPS 2019).
https://arxiv.org/abs/1904.01681 — doi:10.48550/arXiv.1904.01681
Four supporting claims are additionally replicated from: Ricky T. Q. Chen, Yulia Rubanova,
Jesse Bettencourt, David K. Duvenaud. "Neural Ordinary Differential Equations." NeurIPS 2018.
https://arxiv.org/abs/1806.07366 — doi:10.48550/arXiv.1806.07366
PDF URL: https://github.com/VanniLeonardo/re-augmented-neural-odes/releases/download/v1.0.3/article.pdf
Metadata URL: https://raw.githubusercontent.com/VanniLeonardo/re-augmented-neural-odes/v1.0.3/paper/metadata.yaml
Code URL: https://github.com/VanniLeonardo/re-augmented-neural-odes
(archived on Zenodo — concept DOI 10.5281/zenodo.22704199; the snapshot submitted is release v1.0.3, 10.5281/zenodo.22921214)
Scientific domain: Machine learning (neural differential equations)
Programming language: Python 3.11 (PyTorch 2.5.1, torchdiffeq 0.2.5; pinned CPU and GPU containers provided)
Suggested editor: Koustuv Sinha (alternative: Georgios Detorakis)
Eight claims are tested: four from Dupont et al. and the four supporting claims of Chen et al.
on which they rest. Five reproduce, two partially, and one does not. The replication is
independent, written from the published descriptions rather than either author's code, and
every measurement carries a check that the solver actually integrated the flow.
Two earlier NeurIPS 2019 Reproducibility Challenge reports on Dupont et al. exist, both
rejected and unpublished: one reproduced the results using the authors' released
implementation, the other reimplemented and undershot CIFAR-10 by about eight points without
accounting for it. Section 1 states the difference.
The non-reproduced claim is Chen et al. Figure 3c. We traced its reported value to a counting
artefact in the original experiment: the evaluation counter is reset after the forward count is
recorded, so each logged forward count also contains the previous batch's backward count. The
original authors shared their unreleased code and logs, have reviewed the analysis and agree
with it, and plan to correct their paper. Corrected, their ratio is 0.78–0.86 rather than 0.5.
A GitHub Actions workflow builds the pinned container from scratch on every push, runs the
acceptance gate, rebuilds every figure from the committed per-seed CSVs and fails if any
committed result changes.
Original article: Emilien Dupont, Arnaud Doucet, Yee Whye Teh. "Augmented Neural ODEs."
Advances in Neural Information Processing Systems 32 (NeurIPS 2019).
https://arxiv.org/abs/1904.01681 — doi:10.48550/arXiv.1904.01681
Four supporting claims are additionally replicated from: Ricky T. Q. Chen, Yulia Rubanova,
Jesse Bettencourt, David K. Duvenaud. "Neural Ordinary Differential Equations." NeurIPS 2018.
https://arxiv.org/abs/1806.07366 — doi:10.48550/arXiv.1806.07366
PDF URL: https://github.com/VanniLeonardo/re-augmented-neural-odes/releases/download/v1.0.3/article.pdf
Metadata URL: https://raw.githubusercontent.com/VanniLeonardo/re-augmented-neural-odes/v1.0.3/paper/metadata.yaml
Code URL: https://github.com/VanniLeonardo/re-augmented-neural-odes
(archived on Zenodo — concept DOI 10.5281/zenodo.22704199; the snapshot submitted is release v1.0.3, 10.5281/zenodo.22921214)
Scientific domain: Machine learning (neural differential equations)
Programming language: Python 3.11 (PyTorch 2.5.1, torchdiffeq 0.2.5; pinned CPU and GPU containers provided)
Suggested editor: Koustuv Sinha (alternative: Georgios Detorakis)
Eight claims are tested: four from Dupont et al. and the four supporting claims of Chen et al.
on which they rest. Five reproduce, two partially, and one does not. The replication is
independent, written from the published descriptions rather than either author's code, and
every measurement carries a check that the solver actually integrated the flow.
Two earlier NeurIPS 2019 Reproducibility Challenge reports on Dupont et al. exist, both
rejected and unpublished: one reproduced the results using the authors' released
implementation, the other reimplemented and undershot CIFAR-10 by about eight points without
accounting for it. Section 1 states the difference.
The non-reproduced claim is Chen et al. Figure 3c. We traced its reported value to a counting
artefact in the original experiment: the evaluation counter is reset after the forward count is
recorded, so each logged forward count also contains the previous batch's backward count. The
original authors shared their unreleased code and logs, have reviewed the analysis and agree
with it, and plan to correct their paper. Corrected, their ratio is 0.78–0.86 rather than 0.5.
A GitHub Actions workflow builds the pinned container from scratch on every push, runs the
acceptance gate, rebuilds every figure from the committed per-seed CSVs and fails if any
committed result changes.