Finding Salient Convolutional Filters with Extreme Value Theory

Abstract

Saliency maps have attracted attention from researchers because they help people visually understand the behaviour of deep learning models. However, these maps do not necessarily reflect the pixels that lead to misclassification. In this work, we addressed this issue by focusing on the parameter space and propose an algorithm based on extreme value theory that identifies malfunctioning convolutional filters in a CNN. We describe the mathematical understanding of our method and report the empirical results showing the effectiveness of our algorithm.