A self-interpretable module for deep image classification on small data

Abstract Deep neural networks are the driving force of recent explosion machine learning applications in everyday life. However, they usually require a lot training data to work well, and act as black-boxes, making predictions without any explanation about them. This paper presents Memory Wrap, module (i.e, set layers) that can be added deep models improve their performance interpretability settings where few available. Wrap adopts sparse content-attention mechanism between input some memories past samples. We show adding standard improves when learn from limited data, allows them reach comparable full dataset. discuss how analysis its structure weights helps get insights decision process makes more interpretable, compared same Wrap. test our approach on image classification tasks using several three different datasets, namely CIFAR10, SVHN, CINIC10.