A Survey on Neural Network Interpretability