Train a convolutional neural network on the CIFAR-10 data and find image patches which maximally activate the filters
