Truly shift-invariant convolutional neural networks
Authors: Anadi Chaman and Ivan Dokmanić Convolutional neural networks were always assumed to be shift invariant, until recently when it was shown that the classification accuracy of a trained CNN can take a serious hit with merely a 1-pixel shift in input image. One of the primary reasons for this problem is the use of downsampling (popularly known as stride) layers in the networks. In this work, we present Adaptive Polyphase Sampling (APS), an easy-to-implement non-linear downsampling scheme that completely […]
Read more