A self-explanatory method for the black problem on discrimination part of CNN

Convolution neural networks (CNNs) have surpassed human's abilities in some specific tasks. However, they are considered difficult to understand and explain. Recently, the black box problem of CNN, especially concerning the discrimination part, has been studied by different scientific communities. Many methods were proposed for extracting an interpretable model from the discrimination part, which can explain the prediction of the part. However, it is hard for the interpretable models to approximate the discrimination part because of the tradeoff problem between interpretability performance and generalization performance of the discrimination part. We suppose the tradeoff problem is mainly attributed to the fact that the sufficient and necessary condition for the consistent convergence of the both performances is hard to be guaranteed by tradition learning algorithm. So the tradeoff problem could be solved by shrinking the distance between the interpretable model and the discrimination part. This paper firstly introduces a Markov random field model (MRF), namely Deep Cognitive Learning Model (DCLM), which explains the causal relationship between the features(the weight matrixes in the first layer of the discrimination part of CNN can capture) and the output result of the discrimination part of CNN. A greedy algorithm is proposed for initiatively extracting the DCLM from the discrimination part by solving a MAX-SAT problem. And then, a game process between two MAP inferences is implemented for shrinking an interpretation distance which can evaluate how close the discrimination part is to the DCLM. Finally, the proposed self-explanatory approach is evaluated by some contrastive experiments with certain baseline methods on some standard image processing benchmarks. These experiments indicate that the proposed method can improve the interpretability performance of the discrimination part without largely reducing its generalization performance, the generalization performance of the DCLM also can be improved and the discrimination part can be explained by the DCLM in real time during the training process after the interpretation distance converges.