Evaluating the overall sensitivity of saliency-based explanation methods

Abstract

We address the need to generate faithful explanations of "black box" Deep Learning models. Several tests have been proposed to determine aspects of faithfulness of explanation methods, but they lack cross-domain applicability and a rigorous methodology. Hence, we select an existing test that is model agnostic and is well-suited for comparing one aspect of faithfulness (i.e., sensitivity) of multiple explanation methods, and extend it by specifying formal thresh-olds and building criteria to determine the over-all sensitivity of the explanation method. We present examples of how multiple explanation methods for Convolutional Neural Networks can be compared using this extended methodology. Finally, we discuss the relationship between sensitivity and faithfulness and consider how the test can be adapted to assess different explanation methods in other domains.