Relevance regularization of convolutional neural network for interpretable classification

Conventional end-to-end learning algorithm considers only the final prediction output and ignores layer-wise relational reasoning during the training. In this paper, we propose to use a forward and backward interacted-activation (FBI) loss function that regularizes training a CNN so that the prediction model can provide interpretable results for classification. From our best knowledge, the proposed algorithm is the first work to use a regularization function without any prior knowledge or pre-defined terms to allow for a CNN to be more explainable. It is demonstrated with quantitative and qualitative analysis that the proposed technique can be used for efficiently train a CNN with more interpretability, applied to a well-known classification problem.