Supplementary Material:Interpretable Neural Networks Decoupling

As shown in Fig A, we collect filters in the last convolutional layer of VGG16 on ImageNet 2012 after network decoupling and present the results of the number of times they are activated. ImageNet 2012 contains 50,000 validation images, so the number of times each filter is activated falls in the interval [0, 50000]. The leftmost bar represents the number of filters that never been activated (i.e., silent filters), and the rightmost bar represents the number of filters that are activated every time (i.e., energetic filters). The middle bars represent the number of filters, which respond to specific inputs (i.e., dynamic filters). For instance, the rightmost bar represents that there are 60 filters, which are activated 50,000 times during evaluating on ImageNet 2012, in the last convolutional layer of