Reproduce “Adversarial robustness does not entail human-aligned categorization” 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
Itay Reznik, Raz Lapid, Moshe Sipper, Tal Golan
The authors test whether robust ImageNet classifiers categorize deliberately generated images as people do. They optimize images to make pairs of four classifiers favor different labels, then compare predictions with human choices. On targeted two-choice trials, model-human agreement was 48.6–51.3%, near chance, versus 85.0–86.8% human predictability. This separates tested local robustness from alignment on these stimuli.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.