Attribution methods help explain AI decisions by identifying which input regions matter most. The new arXiv paper "Less is More: Efficient Black-box Attribution via Minimal Interpretable Subset Selection" tackles this challenge for discrete inputs.
The authors propose an efficient black-box approach that selects a minimal subset of input regions that is still interpretable. The title's "less is more" suggests that focusing on a small, meaningful subset can be more effective than exhaustive attribution.
The abstract is cut off, so details of the method and experiments are not available from the source. However, the stated goal is to support trustworthy AI by improving the efficiency of attribution for discrete data.