Evaluating Post-hoc Interpretability with Intrinsic Interpretability