Exclusive Feature Constrained Class Activation Mapping for Better Visual Explanation

Abstract

Whereas Deep Neural Network(DNN) shows wonderful performance on large scale data, lacking interpretability limits their usage in scenarios relevant to security. To make visual explanations less noisy and more class-discriminative, in this work, we propose a visual explanation method of DNN, named Exclusive Feature Constrained Class Activation Mapping(EFC-CAM). A new exclusive feature constraint is introduced to optimize the weight calculated from Grad-CAM or initialized from a constant vector. To better measure visual explanation methods, we design an effective evaluation metric which does not need bounding boxes as auxiliary information. Extensive quantitative experiments and visual inspection on ImageNet and Fashion validation set show the effectiveness of the proposed method.