CoProNN: Concept-based Prototypical Nearest Neighbors for Explaining Vision Models

Domain experts can now explain AI vision models using natural language rather than requiring technical machine learning expertise. CoProNN allows experts to describe visual concepts in plain language, which are then converted into visual prototypes for explaining model predictions—making explainability more accessible across different fields.

Task-specific explanations are more useful than one-size-fits-all approaches. The paper emphasizes that good explanations should be tailored to individual tasks and relate to concepts relevant to those specific tasks, suggesting that generic XAI methods may miss important context that domain experts understand.

The method is modular and adaptable as AI technology improves. CoProNN's design allows researchers to swap in newer classification and text-to-image models without rebuilding the entire system, meaning explanations can benefit from advances in AI without requiring complete redesigns.

Visual prototypes provide a common language between AI researchers and domain experts. Because the explanations are based on visual concepts that can be easily shown and communicated, domain experts can verify whether the AI is using the right reasoning—not just whether it gets the right answer.

The approach performs competitively on real-world image classification tasks. The method shows promise on both standard coarse-grained image classification and more demanding tasks, suggesting practical applicability beyond simple use cases.