EPIC: Explanation of Pretrained Image Classification Networks via Prototype

EPIC bridges two previously separate approaches to AI explainability: The method combines the practical flexibility of post-hoc explanations (which work with existing pre-trained models) with the intuitive, human-understandable prototype-based explanations traditionally only available from specially-designed models, potentially making explainability more accessible without requiring model retraining.

Prototype-based explanations offer more intuitive insights than standard visualization methods: Unlike saliency maps and other feature importance visualizations that show "coarse" decision-making information, EPIC generates explanations using representative patches from actual training data, which are easier for non-experts to understand and verify.

EPIC works with existing AI models without modification: Because EPIC operates as a post-hoc method, it can be applied to pre-trained image classification models already in use, eliminating the need to rebuild and retrain systems—a significant practical advantage for organizations implementing explainability in deployed AI systems.

The approach was validated on standard benchmark datasets: The researchers tested EPIC on established datasets like CUB-200-2011 and Stanford Cars, suggesting the method's effectiveness can be independently verified and compared against other explainability approaches using consistent evaluation standards.