• gradcam
• grounded Sam
• lime
• rise
• saliency map evaluation
• sidu
• woe

In recent years, Explainable AI (XAI) has gained significant attention, as it aims to make AI systems more transparent and understandable. Among the various XAI approaches, a radical shift in perspective has recently emerged in the form of hypothesis-driven XAI through a novel framework called evaluative AI. In this thesis work, we expand this framework by proposing a new approach that provides hypothesis-driven evaluations measuring the conformity of concepts with the predictions of an AI model. Specifically, the proposed solution integrates the Weight of Evidence (WoE) statistical approach with human supervision, allowing for a co-operative assessment of the alignment of human-suggested concepts with the classification of the AI model.
An open-vocabulary segmentation model, specifically the Segment Anything Model 2 (SAM2), was used to include humans in the loop, which provides a textual caption consisting of multiple concepts, that may be present in the input images and which may be relevant for humans to understand the explanations of the prediction made by the classifier. The XAI techniques used are four different saliency methods, namely GRADCAM, LIME, RISE, and SIDU. The proposed approach method supports users in the evaluation of the alignment between human concepts and deep learning model predictions.