Interpretable deep learning by disentangling features using modified Siamese network

In recent years, deep learning has entailed great progresses, but this success is seen as a triumph of empiricism. Despite the efforts made in the understanding of the deep models, the gap between the theoretical interpretability and the empirical success of deep models is still large. At present, the interpretability analysis of deep learning is mostly done after the training process. In this work, a novel method is proposed to extract interpretable features during the training process. Particularly, a modified Siamese network is used to disentangle features using a specially designed loss function. After a feature extractor, a fully connected layer is used to divide the output of the Siamese network into four different parts from which the new loss function is designed. The feature extractor used in the Siamese network may be different types of CNN. It is found that the proposed method can effectively disentangle the features at the semantic level, and the disentangled features can be used to effectively improve the classification accuracy in occluded face recognition.