[论文理解] Why do deep convolutional networks generalize so poorly to small image transformations?