Reproduce “GeRA: Label-Efficient Geometrically Regularized Alignment” 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
Dustin Klebe, Tal Shnitzer, Mikhail Yurochkin, Leonid Karlinsky, Justin Solomon
GeRA aligns pretrained image and text encoders by learning transformations from paired examples while using unpaired examples to preserve local neighborhood structure. On ImageNet, its zero-shot accuracy is nearly 9% with 1,000 paired training examples, compared with about 3% for contrastive-only alignment. These results cover tested settings and do not establish effectiveness across all encoders or modalities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.