On Generating Plausible Counterfactual and Semi-Factual Explanations for Deep Learning

There is a growing concern that the recent progress made in AI, especially regarding predictive competence of deep learning models, will be undermined by failure to properly explain their operation and outputs. In response this disquiet, counterfactual explanations have become very popular eXplainable AI (XAI) due asserted computational, psychological, legal benefits. contrast however, semi-factuals (which appear equally useful) surprisingly received no attention. Most methods address tabular rather than image data, partly because non-discrete nature images makes good counterfactuals difficult define; indeed, generating plausible which lie on data manifold also problematic. This paper advances novel method for black-box CNN classifiers doing computer vision. The present method, called PlausIble Exceptionality-based Contrastive Explanations (PIECE), modifies all “exceptional” features test “normal” from perspective class, generate images. Two controlled experiments compare others literature, showing PIECE generates highly (and best semi-factuals) several benchmark measures.