Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge

For an explanation of a deep learning model to be effective, it must both provide insight into a model and suggest a corresponding action in order to achieve an objective. Too often, the litany of proposed explainable deep learning methods stop at the first step, providing practitioners with insight into a model, but no way to act on it. In this paper we propose contextual decomposition explanation penalization (CDEP), a method that enables practitioners to leverage explanations to improve the performance of a deep learning model. In particular, CDEP enables inserting domain knowledge into a model to ignore spurious correlations, correct errors, and generalize to different types of dataset shifts. We demonstrate the ability of CDEP to increase performance on an array of toy and real datasets.