From Attribution Maps to Human-Understandable Explanations through Concept Relevance Propagation

CRP bridges a critical gap in AI explainability: Traditional methods either show where a model focuses (attribution maps) or what concepts it has learned (global techniques), but not both. Concept Relevance Propagation combines these perspectives to answer both questions for individual predictions, providing more complete explanations of how AI systems make decisions.

Human interpretability requires connecting features to meaningful concepts: The paper demonstrates that identifying important pixels or features is insufficient for understanding AI reasoning—users need to know what those features actually represent. CRP addresses this by linking low-level features to human-understandable concepts, making model behavior more transparent to non-experts.

Concept atlases and composition analysis enable deeper model auditing: The method generates "concept atlases" and analyzes how concepts combine in decision-making, allowing stakeholders to systematically investigate whether models rely on appropriate reasoning patterns or potentially problematic associations—important for identifying bias and ensuring responsible deployment.

Quantitative concept analysis supports rigorous model evaluation: Beyond visualization, CRP enables quantitative investigation of concept subspaces and their decision-making roles, providing measurable insights into model behavior rather than relying solely on qualitative interpretation, which is essential for accountability and governance.