P2ExNet: Patch-based Prototype Explanation Network

Dominique Mercier, Andreas Dengel, Sheraz Ahmed

In: ArXiv e-prints International Conference on Neural Information Processing. International Conference on Neural Information Processing (ICONIP-2020) 27th International Conference on Neural Information Processing November 18-22 Bangkok Thailand Seiten 318-330 LNCS 12534 ISBN 978-3-030-63836-8 Springer 11/2020.


Deep learning methods have shown great success in severaldomains as they process a large amount of data efficiently, capable ofsolving complex classification, forecast, segmentation, and other tasks.However, they come with the inherent drawback of inexplicability lim-iting their applicability and trustworthiness. Although there exists workaddressing this perspective, most of the existing approaches are limitedto the image modality due to the intuitive and prominent concepts. Con-versely, the concepts in the time-series domain are more complex andnon-comprehensive but these and an explanation for the network deci-sion are pivotal in critical domains like medical, financial, or industry.Addressing the need for an explainable approach, we propose a novelinterpretable network scheme, designed to inherently use an explainablereasoning process inspired by the human cognition without the needof additional post-hoc explainability methods. Therefore, class-specificpatches are used as they cover local concepts relevant to the classifica-tion to reveal similarities with samples of the same class. In addition,we introduce a novel loss concerning interpretability and accuracy thatconstraints P2ExNet to provide viable explanations of the data includ-ing relevant patches, their position, class similarities, and comparisonmethods without compromising accuracy. Analysis of the results on eightpublicly available time-series datasets reveals that P2ExNet reaches com-parable performance when compared to its counterparts while inherentlyproviding understandable and traceable decisions.

Weitere Links

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence