Locally Sparse Networks for Interpretable Predictions