LeJEPA

LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

Randall Balestriero, Yann LeCun · arXiv 2025

Read paper · arXiv:2511.08544

Learn isotropic representations with Gaussianity constraints on random projections.

What you will implement

SIGReg characteristic-function statistics, numerical integration, and slice aggregation.

• 55. LeJEPA Sliced Gaussian Regularization

  Match a standard Gaussian through sliced empirical characteristic functions while retaining feature gradients.

  Given unit projection directions and a positive quadrature grid, implement Epps-Pulley statistics and their slice mean. Direction sampling, encoders, multiview prediction and distributed communication are supplied or outside scope.