Gaussian Joint Embeddings provide a probabilistic alternative to deterministic SSL, eliminating the need for architectural asymmetries to prevent collapse.
March 31, 2026
Original Paper
Gaussian Joint Embeddings For Self-Supervised Representation Learning
arXiv · 2603.26799
The Takeaway
Current self-supervised learning (SSL) relies on tricks like stop-gradients or momentum encoders to avoid representation collapse. GJE uses generative joint modeling to provide principled uncertainty estimates and a covariance-aware objective, offering a more stable and theoretically grounded foundation for representation learning.
From the abstract
Self-supervised representation learning often relies on deterministic predictive architectures to align context and target views in latent space. While effective in many settings, such methods are limited in genuinely multi-modal inverse problems, where squared-loss prediction collapses towards conditional averages, and they frequently depend on architectural asymmetries to prevent representation collapse. In this work, we propose a probabilistic alternative based on generative joint modeling. W