MM-LeJEPA: Correlated Gaussian Regularization for Multimodal Learning
October 20, 2026
MM-LeJEPA extends LeJEPA to multimodal tasks by introducing a correlated Gaussian target that accounts for cross-modal structure. This enables self-supervised learning without reconstruction decoders or contrastive negatives by minimizing a weighted Fisher information functional.
HOW THIS AFFECTS YOU
●
researcherYou can implement more efficient multimodal self-supervised learning that avoids the overhead of contrastive training.