As I recall, the most interesting feature of these networks is that they do better when learning geometric features in the presence of noise. Without the added

As I recall, the most interesting feature of these networks is that they do better when learning geometric features in the presence of noise. Without the added geometric structure baked into the network, the real neural nets overfit and 'learn the noise' to a much greater extent.