ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
ConEx ties vision saliency maps to automatically discovered concepts using unsupervised concept activation vectors.
Concept-based Explanations (ConEx) combines saliency maps with automatically discovered class-specific concepts represented as concept activation vectors learned without manual labels. An architecture-specific masking step is used to reduce noise from segmentation masks and improve concept purity, and maps show where each concept appears and how it affects the prediction. The paper introduces Vector-Concept Match and Concept-Class Match metrics and reports state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks.
- ConEx discovers class-specific concepts as unsupervised concept activation vectors.
- Architecture-specific masking aims to reduce segmentation-mask noise and improve concept purity.
- New VCM and CCM metrics score concept alignment and class match.
- Authors report state-of-the-art faithfulness, segmentation, and concept-quality results.
From the original article
1 min read at huggingface.co
Many visual explanation methods in computer vision highlight pixel importance but struggle to link these low-level cues to semantically meaningful concepts, limiting their interpretability and trustworthiness. We introduce Concept-based Explanations (ConEx), a novel framework that bridges saliency visualization with concept-based reasoning to provide both faithfulness and interpretability.
By Yehonatan Elisha, Oren Barkan, Ziv Weiss Haddad · huggingface.co
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.