OPTIMUS-Prime: Minimal and Sufficient Concept Explanations for Deep Vision Models

Explanations now come with formal guarantees: Unlike previous visual explanation methods that prioritize user-friendliness without rigorous backing, OPTIMUS provides mathematical guarantees that the highlighted features actually caused the model's decision (sufficiency) and that no unnecessary features are included (minimality).

Addresses a critical gap between theory and practice: The paper identifies that existing computer vision explanation methods have sacrificed theoretical rigor for accessibility. OPTIMUS bridges this gap by grounding explanations in established mathematical theory (prime implicants) while remaining visually interpretable to non-experts.

Minimal explanations reduce confusion and bias: By ensuring explanations contain only the essential concepts needed for a decision, OPTIMUS prevents users from being misled by irrelevant highlighted features, which could otherwise create false confidence in model decisions or introduce human bias.

Visual heatmaps remain user-friendly despite added rigor: The framework maintains the practical advantage of visual explanations that end-users can understand, while adding the theoretical foundation that auditors, regulators, and AI safety researchers need to verify model behavior.

Validation demonstrates real-world applicability: Testing on visual classification benchmarks shows that OPTIMUS heatmaps naturally surface the actual decision-relevant concepts, suggesting the approach works in practice and could support more trustworthy automated decision-making in applications like medical imaging or autonomous systems.