Hide and Seek: HIDE Concepts to SEEK explanations for black box models
Hide and Seek attributes concepts in black-box vision models without weights or labeled concept datasets.
Hide and Seek (HnS) is a concept-level attribution method for black-box vision models that does not require weights or large concept-annotated datasets. A vision-language model retrieves concept-aligned images and localizes the concept; masking that region and comparing outputs measures attribution. On CelebA and COCO-Stuff, HnS agreed better with ground-truth attributions than existing methods. The authors also applied it to image captioning and satellite image segmentation, tasks prior concept methods could not cover.
- HnS needs neither model weights nor labeled concept data
- A vision-language model localizes a concept before masking
- Attribution comes from the change in black-box output
- Validated on CelebA, COCO-Stuff, captioning, and satellite segmentation
From the original article
1 min read at arxiv.orgInterpreting machine learning model predictions is a central challenge, particularly for black-box vision models. Concept-based attribution methods interpret models using human-understandable concepts. Existing concept-based methods can only interpret classification and regression models.
By Bhanu Tokas · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.