Reproduce “LocaLeJEPA: Maximum Entropy Representations under Geometric Priors” 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
Kartik Tandon, Alessio Borgi, Pavel Snopov, Inés García-Redondo, Davide D'Ascenzo, Sebastiano Cultrera di Montesano, Fabrizio Silvestri, Pietro Liò and 3 more
LocaLeJEPA extends self-supervised image learning by replacing LeJEPA’s global Gaussian anti-collapse target with local uniform-ball matching. It groups nearby embeddings, centers each group, and regularizes its internal shape while allowing groups to retain separate positions. On ImageNet-1K, Euclidean ViT-S/8 frozen-feature linear-probe accuracy was 62.1%, versus 51.6% for matched-settings LeJEPA. These results concern tested vision settings, not universal topology preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.