Explainable image classification with evidence counterfactual

Abstract The complexity of state-of-the-art modeling techniques for image classification impedes the ability to explain model predictions in an interpretable way. A counterfactual explanation highlights parts which, when removed, would change predicted class. Both legal scholars and data scientists are increasingly turning explanations as these provide a high degree human interpretability, reveal what minimal information needs be changed order come different prediction do not require disclosed. Our literature review shows that existing methods have strong requirements regarding access training internals, which often unrealistic. Therefore, SEDC is introduced model-agnostic instance-level method does need data. As tasks typically multiclass problems, additional contribution introduction SEDC-T allows specifying target These experimentally tested on ImageNet data, with concrete examples, we illustrate how resulting can give insights decisions. Moreover, benchmarked against methods, demonstrating stability results, computational efficiency nature explanations.