Less is More: Efficient Black-box Attribution via Minimal Interpretable Subset Selection

Attribution methods help identify which parts of an image influence AI decisions: This research addresses a critical need for transparency in AI systems by developing better techniques to pinpoint which visual elements (like specific objects or regions) actually drive a model's predictions, making AI behavior more understandable to users and developers.

The new LiMA method requires examining fewer image regions while maintaining accuracy: By identifying the most important sub-regions more efficiently, LiMA achieves state-of-the-art results with 36.3% improvement in identifying critical features and 39.6% improvement in detecting non-critical ones, while running 1.6 times faster than previous approaches—demonstrating that simpler explanations can be more reliable.

The approach works across diverse AI systems and image types: Testing on eight different foundation models (including CLIP and ImageBind) and six datasets (from ImageNet to medical images) shows this method generalizes well, suggesting it could become a practical tool for explaining many different AI applications in real-world settings.

The method accounts for complex interactions between image regions: Unlike simpler attribution techniques, LiMA recognizes that image regions don't work in isolation—some regions work well together while others interfere with each other—providing more nuanced and realistic explanations of how AI systems actually process visual information.