Concept aware saliency maps improve visual model explanations
ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine Learning
Summary
Many tools show which pixels in images influence a computer vision model’s decisions, but these highlights often lack clear meaning. The authors present ConEx, a new method that automatically finds meaningful visual concepts specific to each class and links these to areas in the image. ConEx creates clearer, concept-focused maps that help people understand how different parts of an image contribute to a model's prediction. They also introduce new metrics to measure how well these concepts align with actual image features and classes. Their tests show that ConEx provides more faithful and interpretable explanations than prior methods.
What this means in practice
- •For computer vision teams: Create clearer, concept-based explanations to validate and debug image recognition models automatically.
- •For medical imaging analysts: Use concept-aware saliency maps to better interpret AI decisions in complex visual data like x-rays and MRIs.
Authors
Yehonatan Elisha, Oren Barkan, Ziv Weiss Haddad, Noam Koenigstein
Abstract
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. ConEx automatically discovers class-specific concepts and represents them through concept activation vectors (CAVs), learned without manual supervision using an architecture-specific masking mechanism that reduces noise introduced by the segmentation masks to enhance concept purity. ConEx generates faithful saliency maps that reveal where each concept appears in the image and how it contributes to the prediction. To evaluate the reliability of these learned concepts, we propose two complementary metrics, Vector-Concept Match (VCM) and Concept-Class Match (CCM), that quantify concept alignment and enable direct comparison with existing methods. Extensive experiments across diverse settings demonstrate that ConEx achieves state-of-the-art performance on faithfulness, segmentation, and concept-quality benchmarks. Overall, ConEx advances the field toward truly interpretable and concept-grounded explanations in vision models.