Explaining AI-based Decision Support Systems using Concept Localization Maps

Explainability is critical for trust in high-stakes AI systems: The paper emphasizes that even accurate AI decision support systems won't be trusted by experts in critical fields (like healthcare or law) unless they can explain their reasoning—highlighting that accuracy alone is insufficient for responsible AI deployment.

The method identifies which image regions influence AI decisions: Concept Localization Maps (CLMs) show not just what concepts an AI uses to make decisions, but where in an image those concepts appear, achieving over 80% localization recall for most concepts on their test dataset.

Human-understandable concepts improve AI accountability: By mapping AI decisions to human-recognizable concepts rather than opaque internal patterns, CLMs make it possible for domain experts to verify whether the system is focusing on appropriate features—essential for catching bias or errors before deployment.

New public dataset enables reproducible explainability research: The authors created and released the Simple Concept DataBase (SCDB) with annotated concepts, supporting the broader research community's ability to develop and test explainability methods rather than relying on proprietary datasets.

Visual explanations can increase practical adoption of AI systems: The paper argues that providing visual justifications through localization maps addresses a key barrier to real-world implementation—helping organizations and professionals feel confident enough to actually use AI decision support tools in practice.