<ul><li>Researchers propose LiMA (Less input is More faithful for Attribution), a novel black-box attribution mechanism for AI systems.</li><li>LiMA reformulates attribution of important regions as an optimization problem for submodular subset selection.</li><li>The method accurately assesses input-prediction interactions and improves optimization efficiency using a bidirectional greedy search algorithm.</li><li>Experiments show that LiMA provides faithful interpretations with fewer regions, exhibits strong generalization, and outperforms other attribution algorithms.</li></ul>

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

Discover more