Using Explanations to Estimate the Quality of Computer Vision Models

Explainable AI (xAI) emerged as one of the ways of addressing the interpretability issues of the so-called black-box models. Most of the xAI artifacts proposed so far were designed, as expected, for human users. In this work, we posit that such artifacts can also be used by computer systems. Specifically, we propose a set of metrics derived from LIME explanations, that can eventually be used to ascertain the quality of each output of an underlying image classification model. We validate these metrics against quantitative human feedback, and identify 4 potentially interesting metrics for this purpose. This research is particularly useful in concept drift scenarios, in which models are deployed into production and there is no new labelled data to continuously evaluate them, becoming impossible to know the current performance of the model.