Can Post-hoc Explanations Effectively Detect Out-of-Distribution Samples?

Today there is consensus around the importance of explainability as a mandatory feature in practical deployments of Artificial Intelligence (AI) models. Most research activity reported so far in the eXplainable AI (XAI) research arena has stressed on proposing new techniques for eliciting such explanations, together with different approaches for measuring their effectivity to increase the trustworthiness of the audience for which explanations are furnished. However, alternative uses of explanations beyond their original purpose have been very scarcely explored. In this work we investigate whether local explanations can be utilized for detecting Out-of-Distribution (OoD) test samples in machine learning classifiers, i.e., to identify whether query examples of an already trained classification model can be thought to belong to the distribution of the training data. To this end, we devise and assess the performance of a clustering-based OoD detection approach that exemplifies how heatmaps produced by well-established local explanation methods can be of further use than explaining individual predictions issued by the model under analysis. The overarching purpose of this work is not only to expose the benefits and the limits of the proposed XAI-based OoD approach, but also to point out the enormous potential of post-hoc explanations beyond easing the interpretability of black-box models themselves.