LocaLeJEPA: Maximum Entropy Representations under Geometric Priors
Published in arXiv preprint arXiv:2610.28458, 2026
Overview
LocaLeJEPA keeps LeJEPA’s agreement across views while replacing its global Gaussian constraint with LocaLeReg. The regulariser forms neighbourhoods in projected embeddings, centres each neighbourhood separately, and matches its residual distribution to a covariance-normalised uniform ball. Local spreading discourages collapse while leaving neighbourhood locations free to organise.
A heat-flow perspective motivates the construction. Frozen-feature experiments evaluate recognition, transfer and controlled rotation recovery across Euclidean, hyperbolic and equivariant encoders.
Kartik Tandon and Alessio Borgi contributed equally.
Resources
- 📄 ArXiv: arXiv:2610.28458
- 🧾 BibTeX: arXiv BibTeX
- 💻 Code: LocaLeJEPA on GitHub
- 📘 Companion blog post: LocaLeJEPA: Maximum Entropy Without Flattening the Geometry
- LinkedIn: Preprint announcement
Recommended citation: Tandon, K.; Borgi, A.; Snopov, P.; García-Redondo, I.; D’Ascenzo, D.; Cultrera di Montesano, S.; Silvestri, F.; Liò, P.; Dominici, F.; Balestriero, R.; Battiloro, C. (2026). "LocaLeJEPA: Maximum Entropy Representations under Geometric Priors." arXiv:2610.28458.
Download Paper | Download Bibtex