Interpretations are useful: penalizing explanations to align neural networks with prior knowledge