Various types of saliency methods have been proposed for explaining black-box classification. In image applications, this means highlighting the part of the image that is most relevant for the current decision. Unfortunately, the different methods may disagree and it can be hard to quantify how representative and faithful the explanation really is. We observe however that several of these methods can be seen as edge cases of a single, more general procedure based on finding a particular path through the classifier's domain. This offers additional geometric interpretation to the existing methods. We demonstrate furthermore that ablation paths can be directly used as a technique of its own right. This is able to compete with literature methods on existing benchmarks, while giving more fine-grained information and better opportunities for validation of the explanations' faithfulness.
View on arXiv