Anatomically Constrained ResNets Exhibit Opponent Receptive Fields; So What?

Primate visual systems are well known to exhibit varying degrees of bottlenecks 1 in the early visual pathway. Recent works have shown that the presence of a bot2 tleneck between ‘retinal’ and ‘ventral’ parts of artificial models of visual systems, 3 simulating the optic nerve, can cause the emergence of cellular properties that have 4 been observed in primates: namely centre-surround organisation and opponency. 5 To date, however, state-of-the-art convolutional network architectures for classifica6 tion problems have not incorporated such an early bottleneck. In this paper, we ask 7 what happens if such a bottleneck is added to a ResNet-50 model trained to classify 8 the ImageNet data set. Do the emergent properties observed in simpler models 9 still appear in these considerably deeper and more complex models? Does the 10 bottleneck affect performance or robustness? Does the bottleneck bias the model 11 in some way? or, does something else happen? 12