Why do deep convolutional networks generalize so poorly to small image transformations?

Aharon Azulay, Yair Weiss

Convolutional Neural Networks (CNNs) are commonly assumed to be invariant to small image transformations: either because of the convolutional architecture or because they were trai…