Building Interpretable Deep CNN Based on Binary Tree Structure Convolution Layers

Da Liu, Dawei Dai, Yuqi Wang, Feng Hu, Shuyin Xia, Guoyin Wang

Abstract: In recent years, Deep CNN (DCNN) models have achieved great success in the field of computer vision. However, such models are still considered to lack interpretability. One of fundamental issues underlying this problem can be noted as follows: The decision-making of a DCNN model is considered as a “black-box” operation. In this study, we propose to use binary tree structure convolution layers (TSCL) to interpret the decision-making mechanism of a DCNN model in the image recognition task. First, we design a TSCL module, in which each parent layer generates two child layers, and then integrate them into a normal DCNN. Second, we design an information coding objective to guide each two child nodes of one parent node to learn the particular information coding that we expected. Through the experiments, we can verify that: 1) the logical process of decision-making made by ResNet models can be explained well based on the "decision information flow path" formed in the TSCL module; 2) the decision-path can reasonably interpret the decision reversal mechanism (Robustness mechanism) of the DCNN model; 3) the credibility of decision-making can be measured by the matching degree between the actual and expected decision-path.