Evaluation of the Explanatory Power Of Layer-wise Relevance Propagation using Adversarial Examples

Abstract Approaches for visualizing and explaining the decision process of convolutional neural networks (CNNs) have recently received increasing attention. Particularly popular approaches are so-called saliency methods, which aim to assign a valence each input pixel based on its importance influence classification via maps. In our paper, we contribute by novel analyzing approach build adversarial examples investigate explanatory power methods exemplified layer-wise relevance propagation (LRP). Based hypothesis that distinct decisions, such as an image’s corresponding examples, should yield dissimilar maps provide transparent rationales, break down scores images analyze them using comprehensive statistical evaluation. It turns out different decomposition rules LRP do not lead clearly distinguishable neither in terms their contour lines, nor analysis.