Hide and Seek: HIDE Concepts to SEEK explanations for black box models
arXiv cs.AI · 10 Oct 2026, 03:56 AEST
A new method called Hide and Seek (HnS) has been proposed for interpreting black-box vision models without needing access to model weights or labeled datasets. HnS retrieves concept-aligned images and measures attribution based on changes in model output, showing better agreement with ground-truth attributions compared to existing methods.
Key facts
- Interpreting machine learning model predictions is a central challenge, particularly for black-box vision models.
- Concept-based attribution methods interpret models using human-understandable concepts.
- HnS can interpret a wide range of vision models.
- We validated HnS on CelebA and COCO-Stuff.
- HnS showed better agreement with ground-truth attributions.
Read aloud
A new method called Hide and Seek HnS has been proposed for interpreting black-box vision models without needing access to model weights or labeled datasets. HnS retrieves concept-aligned images and measures attribution based on changes in model output, showing better agreement with ground-truth attributions compared to existing methods. Interpreting machine learning model predictions is a central challenge, particularly for black-box vision models. Concept-based attribution methods interpret models using human-understandable concepts. HnS can interpret a wide range of vision models. We validated HnS on CelebA and COCO-Stuff. HnS showed better agreement with ground-truth attributions.
Sources in this cluster
- Hide and Seek: HIDE Concepts to SEEK explanations for black box models
arXiv cs.AI · 10 Oct 2026, 03:56 AEST