Model-Agnostic Interpretability of Machine Learning