How deep convolutional neural networks lose spatial information with training

A central question of machine learning is how deep nets manage to learn tasks in high dimensions. An appealing hypothesis is that they achieve this feat by building a representation of the data where information irrelevant to the task is lost. For image datasets, this view is supported by the observation that after (and not before) training, the neural representation becomes less and less sensitive to diffeomorphisms acting on images as the signal propagates through the network. This loss of sensitivity correlates with performance and surprisingly correlates with a gain of sensitivity to white noise acquired during training. Which are the mechanisms learned by convolutional neural networks (CNNs) responsible for the these phenomena? In particular, why is the sensitivity to noise heightened with training? Our approach consists of two steps. (1) Analyzing the layer-wise representations of trained CNNs, we disentangle the role of spatial pooling in contrast to channel pooling in decreasing their sensitivity to image diffeomorphisms while increasing their sensitivity to noise. (2) We introduce model scale-detection tasks, which qualitatively reproduce the phenomena reported in our empirical analysis. In these models we can assess quantitatively how spatial pooling affects these sensitivities. We find that the increased sensitivity to noise observed in deep ReLU networks is a mechanistic consequence of the perturbing noise piling up during spatial pooling, after being rectified by ReLU units. Using odd activation functions like tanh drastically reduces the CNNs' sensitivity to noise.