A Learning Paradigm for Interpretable Gradients