Reproduce “Spurious Attention Heads in Vision Transformers: Mechanistic Identification of Annotation-Driven Bias in Medical Chest X-Ray Analysis” 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
Joshua Muhumuza, Cornelius Kalenzi
The authors identify L8H10, an attention head in a chest X-ray classifier, that focuses on upper-left annotations rather than lungs. Among 200 COVID-19 X-rays, it showed this behavior in 123 images (62%). Inpainting raised lung attention 1.52-fold in this group versus 0.99-fold in normal images, while head ablation did not distinguish output-label changes between groups; clinical transfer failure was not tested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.