Highlighting Non-contributing Pixels for Visual Explanation of CNNs

Abstract

Explaining the output of Convolutional Neural Networks (CNNs) is a challenging topic. A typical explanation is to identify which pixels are contributing to the output of CNN. In this paper, we propose a new approach for explaining the output of CNNs by finding pixels that are \emph{not} contributing to the output. To highlight non-contirbuting pixels, we propose optimizing a noise level so that additive noise to the input image does not change the CNN output. The experimental results on MNIST show that the proposed method can idntify non-contributing pixels adequately.