Regularizing Black-box Models for Improved Interpretability