ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
ConEx ties saliency maps to learned visual concepts, so explanations show both where a concept appears and how it affects a model’s prediction.
The paper says ConEx discovers class-specific concepts automatically and represents them with concept activation vectors, without manual supervision. It uses an architecture-specific masking mechanism meant to reduce segmentation-mask noise and improve concept purity. The authors also introduce VCM and CCM metrics to measure concept alignment and compare against other explanation methods. They report state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks.
Source: HF Daily Papers' note
The paper says ConEx discovers class-specific concepts automatically and represents them with concept activation vectors, without manual supervision. It uses an architecture-specific masking mechanism meant to reduce segmentation-mask noise and improve concept purity. The authors also introduce VCM and CCM metrics to measure concept alignment and compare against other explanation methods. They report state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks.
Source: HF Daily Papers' note
score 4