ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
ConEx ties saliency maps to learned visual concepts, aiming to show both where a model looked and what concept drove the prediction.
The paper says ConEx discovers class-specific concepts automatically and encodes them with concept activation vectors, without manual supervision. Its masking mechanism is meant to reduce segmentation-mask noise and improve concept purity. The authors add two metrics, Vector-Concept Match and Concept-Class Match, to test whether the learned concepts align with the model’s classes. They report state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks. ArXiv · AI/CL/LG's note
The paper says ConEx discovers class-specific concepts automatically and encodes them with concept activation vectors, without manual supervision. Its masking mechanism is meant to reduce segmentation-mask noise and improve concept purity. The authors add two metrics, Vector-Concept Match and Concept-Class Match, to test whether the learned concepts align with the model’s classes. They report state-of-the-art results on faithfulness, segmentation, and concept-quality benchmarks. ArXiv · AI/CL/LG's note
score 4