Reproduce “Auditing Where Concept Bottleneck Accuracy Comes From on CUB” 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
Sascha Saralajew, Eleonora Poeta, Eliana Pastor, Ammar Shaker
The authors audit whether concept bottleneck models expose class information through their concept scores and whether those scores use the named visual parts. On CUB-200-2011, a probe recovers 72.11% species accuracy from ResNet50 scores, versus 42.82% from BagNet-33 scores under concept-only training. These findings assess reliability, not user interventions, and apply only to CUB and its noisy, incomplete concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.