Total: 1
Explanations for the outputs of deep neural network classifiers are essential in promoting trust and comprehension among users. Conventional methods often offer explanations only for one single class in the output and neglect other classes with high probabilities, resulting in a limited view of the model's behaviors. In this paper, we propose a holistic explanation method for image classification. It not only facilitates an overall understanding of model behavior, but also provides a framework where one can examine the evidence for discriminating competing classes, and thereby yield contrastive explanations. We demonstrate the advantages of the new method over baselines in terms of both faithfulness to the model and interpretability to users. The source code will be made available to the public upon publication of the paper.