Approximating CNNs with Bag-of-Local-Features Models [pdf]

I consider the paper “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness” essentially mandatory follow-up for anyone interested in this. It goes further with its conclusions and introduces a potential solution. By introducing the ‘Stylized-ImageNet’ dataset, they found that they can force the network to learn shape data, improving scores even on standard ImageNet and preventing this BagNet trick from working.

https://openreview.net/forum?id=Bygh9j09KX

Open Review https://openreview.net/forum?id=SkfMWhAqYQ

Short video discussing this paper: https://www.youtube.com/watch?v=QpptSohzuDo

For anyone else who wanted it: CNN = Convolutional neural network

I somehow mis-parsed that as the news network the first time and became very confused for a moment.

It seems that texture detection is for NN the simplest and easiest way to recognize objects. I guess early animal vision system used to work this way. Then predators and prey developed camouflage and mimicry, to fool texture detection NNs. So there was strong evolutionary pressure to recognize shapes, which is much harder task, but essential for survival. Probably that's why people see in the picture a cat (predator) and ConvNet recognizes most obvious thing (elephant) in terms of statistics.

The results on this paper are 4x worse than the state of the art of a couple of years ago. Personally I think the paper oversells their results. Creating a much worse model by making it much simpler doesn't seem that interesting, to me...

I read this title as being the news network first which was much more interesting :)

Is this closer to how real biological brains process images?