Reproduce “Align Once to Explain: Feature Alignment for Scalable B-cosification of Foundational Vision Transformers” 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
Raphael Maser, Siddhartha Gairola, Sukrut Rao, Bernt Schiele
ALOE converts a pretrained vision transformer into a B-cos model and aligns it to its frozen teacher using unlabeled images. It matches global and intermediate token features, allowing the backbone to provide attribution summaries. For SigLIP2 ViT-B/16, it retained 83.87% ImageNet-1k linear-probe accuracy and scored 84.4% on GridPG localization, a proxy rather than a measure of human interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.