Class-Discriminative Attention Maps for Vision Transformers

Attention maps alone are insufficient for understanding AI decisions: While Vision Transformers naturally produce attention maps showing where the model "looks," these maps don't reveal why the model makes specific predictions. CDAM addresses this by connecting attention to actual classification decisions, making explanations more task-relevant.

Existing explanation methods lack class sensitivity: The research found that seven competing importance estimators failed to adequately distinguish which features matter for different predictions. CDAM's gradient-based approach significantly outperformed alternatives on "class sensitivity" benchmarks, meaning it better identifies features specific to each decision.

Explainability methods can be tailored to different needs: Beyond explaining official predictions, CDAM can identify features associated with arbitrary concepts by measuring similarity in the model's internal representations. This flexibility allows stakeholders to ask different questions about the same model—useful for medical imaging where clinicians might want to understand both disease detection and biomarker identification.

Rigorous evaluation of explanations matters for responsible AI: The paper introduces three quantitative benchmarks (correctness, compactness, and class sensitivity) to evaluate explanation quality. This systematic approach helps prevent overconfidence in explanation methods that may appear intuitive but lack rigor—critical for high-stakes applications like medical diagnosis.

Medical imaging demonstrates real-world importance: The authors validated CDAM on malignancy and biomarker prediction in medical images, suggesting these more reliable explanations could help clinicians understand and trust AI-assisted diagnostic tools.