Research explaining the behavior of convolutional neural networks (CNNs) has gained a lot of attention over the past few years. Although many visualization methods have been proposed to explain network predictions, most fail to provide clear correlations between the target output and the features extracted by convolutional layers. In this work, we define a concept, i.e., class-discriminative feature groups, to specify features that are extracted by groups of convolutional kernels correlated with a particular image class. We propose a detection method to detect class-discriminative feature groups and a visualization method to highlight image regions correlated with particular output and to interpret class-discriminative feature groups intuitively. The experiments showed that the proposed method can disentangle features based on image classes and shed light on what feature groups are extracted from which regions of the image. We also applied this method to visualize "lost" features in adversarial samples and features in an image containing a non-class object to demonstrate its ability to debug why the network failed or succeeded.
Keywords: Convolutional neural networks; Feature visualization; Matrix decomposition; Shapley values.
Copyright © 2020 Elsevier Ltd. All rights reserved.