Explaining Image Classifications with Near Misses, Near Hits and Prototypes: Supporting Domain Experts in Understanding Decision Boundaries
Herchenbach M, Müller D, Scheele S, Schmid U (2022)
Publication Type: Conference contribution
Publication year: 2022
Journal
Publisher: Springer Science and Business Media Deutschland GmbH
Book Volume: 13364 LNCS
Pages Range: 419-430
Conference Proceedings Title: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Event location: Paris
ISBN: 9783031092817
DOI: 10.1007/978-3-031-09282-4_35
Abstract
We propose a method for explaining the results of black box image classifiers to domain experts and end users, combining two example-based explanatory approaches: Firstly, prototypes as representative data points for classes, and secondly, contrastive example comparisons in the form of near misses and near hits. A prototype globally explains the relevant characteristics for a entire class, whereas near hit and near miss explain the local decision boundary of a specific prediction. To combine both types of explanations within one framework is novel and we propose that presenting both types of explanations is especially helpful for domain experts in visual domains. To improve the faithfulness of the explanations, we investigated an unbiased, generic embedding and a model-related (model-specific) embedding for handling the images. The proposed approaches are evaluated regarding parameter selection and suitability on two different data sets – the well-known MNIST and a real-world industrial quality control data set. Finally, it is shown how global and local example-based explanation can be combined and realized within a demonstrator.
Authors with CRIS profile
Involved external institutions
Germany (DE)How to cite
APA:
Herchenbach, M., Müller, D., Scheele, S., & Schmid, U. (2022). Explaining Image Classifications with Near Misses, Near Hits and Prototypes: Supporting Domain Experts in Understanding Decision Boundaries. In Mounîm El Yacoubi, Eric Granger, Pong Chi Yuen, Umapada Pal, Nicole Vincent (Eds.), Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (pp. 419-430). Paris, FR: Springer Science and Business Media Deutschland GmbH.
MLA:
Herchenbach, Marvin, et al. "Explaining Image Classifications with Near Misses, Near Hits and Prototypes: Supporting Domain Experts in Understanding Decision Boundaries." Proceedings of the 3rd International Conference on Pattern Recognition and Artificial Intelligence, ICPRAI 2022, Paris Ed. Mounîm El Yacoubi, Eric Granger, Pong Chi Yuen, Umapada Pal, Nicole Vincent, Springer Science and Business Media Deutschland GmbH, 2022. 419-430.
BibTeX: Download