New probabilistic approach using Gaussian embeddings addresses multi-modal learning collapse in self-supervised representation learning

Researchers propose Gaussian Joint Embeddings (GJE) and its multi-modal extension GMJE as alternatives to deterministic prediction methods in self-supervised learning

The new approach models joint density of context and target representations, replacing black-box prediction with closed-form conditional inference under explicit probabilistic models

Method provides principled uncertainty estimates and covariance-aware objectives to control latent geometry and prevent representation collapse

Addresses key limitation of squared-loss prediction methods that collapse toward conditional averages in genuinely multi-modal inverse problems