Discriminative Attribution from Counterfactuals

We present a method for neural network interpretability by combining feature attribution with counterfactual explanations to generate attribution maps that highlight the most discriminative features between pairs of classes. We show that this method can be used to quantitatively evaluate the performance of feature attribution methods in an objective manner, thus preventing potential observer bias. We evaluate the proposed method on three diverse datasets, including a challenging artificial dataset and real-world biological data. We show quantitatively and qualitatively that the highlighted features are substantially more discriminative than those extracted using conventional attribution methods and argue that this type of explanation is better suited for understanding fine grained class differences as learned by a deep neural network.