Counterfactual attribute-based visual explanations for classification

Abstract In this paper, our aim is to provide human understandable intuitive factual and counterfactual explanations for the decisions of neural networks. Humans tend reinforce their by providing attributes counterattributes. Hence, in work, we utilize as well examples explanations. order counterexplanations make use directed perturbations arrive at counterclass attribute values doing so, explain what present absent original image. We evaluate method when images are misclassified into closer counterclasses completely different counterclasses. conducted experiments on both finegrained coarsegrained datasets. verified attribute-based quantitatively qualitatively showed that discriminating standard robust