Reproduce “The Metagame of Interpretability and Meta-Attributions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hubert Baniecki, Przemyslaw Biecek, Fabian Fumagalli
The metagame extends first-order attribution methods by measuring how other features change a fixed target’s attribution across masked contexts, then using Shapley values to assign directional effects. Each target’s individual effect and incoming effects sum to its original attribution. On an ImageNet-1k-derived interaction-recognition task, Meta-Grad-ECLIP scores .80 ± .01 versus .25 ± .00 for first-order Grad-ECLIP at four objects; this measures interaction localization, not classification accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.