Evolutionary Generative Contribution Mappings

Abstract

Although convolutional neural networks (CNNs) have significantly evolved and demonstrated outstanding performance, their uninterpretable nature is still considered to be a major problem. In this study, we take a closer look at CNN interpretability and propose a new method called Evolutionary Generative Contribution Mappings (EGCM). In EGCM, CNN models incorporate both a classification mechanism and an interpreting mechanism in an end-to-end training process. Specifically, the network generates the class contribution maps, which indicate the discriminative regions for the model to identify a specific class. Additionally, these maps can be directly used for classification tasks; all that is needed is a global average pooling and a softmax function. The network is represented by a directed acyclic graph and optimized using a genetic algorithm. Architecture search enables EGCM to deliver reasonable classification performance while maintaining high interpretability. We apply the EGCM framework on several datasets and empirically demonstrate that the EGCM not only achieves excellent classification performance but also maintains high interpretability.