Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs

In this paper, we challenge the common assumption that collapsing the spatial\ndimensions of a 3D (spatial-channel) tensor in a convolutional neural network\n(CNN) into a vector via global pooling removes all spatial information.\nSpecifically, we demonstrate that positional information is encoded based on\nthe ordering of the channel dimensions, while semantic information is largely\nnot. Following this demonstration, we show the real world impact of these\nfindings by applying them to two applications. First, we propose a simple yet\neffective data augmentation strategy and loss function which improves the\ntranslation invariance of a CNN's output. Second, we propose a method to\nefficiently determine which channels in the latent representation are\nresponsible for (i) encoding overall position information or (ii)\nregion-specific positions. We first show that semantic segmentation has a\nsignificant reliance on the overall position channels to make predictions. We\nthen show for the first time that it is possible to perform a `region-specific'\nattack, and degrade a network's performance in a particular part of the input.\nWe believe our findings and demonstrated applications will benefit research\nareas concerned with understanding the characteristics of CNNs.\n