Improving the Interpretability of Deep Neural Networks with Knowledge Distillation