How Good is your Explanation? Algorithmic Stability Measures to Assess the Quality of Explanations for Deep Neural Networks

A plethora of methods have been proposed to explain how deep neural networks\nreach their decisions but comparatively, little effort has been made to ensure\nthat the explanations produced by these methods are objectively relevant. While\nseveral desirable properties for trustworthy explanations have been formulated,\nobjective measures have been harder to derive. Here, we propose two new\nmeasures to evaluate explanations borrowed from the field of algorithmic\nstability: mean generalizability MeGe and relative consistency ReCo. We conduct\nextensive experiments on different network architectures, common explainability\nmethods, and several image datasets to demonstrate the benefits of the proposed\nmeasures.In comparison to ours, popular fidelity measures are not sufficient to\nguarantee trustworthy explanations.Finally, we found that 1-Lipschitz networks\nproduce explanations with higher MeGe and ReCo than common neural networks\nwhile reaching similar accuracy. This suggests that 1-Lipschitz networks are a\nrelevant direction towards predictors that are more explainable and\ntrustworthy.\n