Var-JEPA Unifies LeCun's Predictive Architecture with Generative SSL via Variational Inference

arXiv·medium signal

Var-JEPA reformulates the Joint-Embedding Predictive Architecture (JEPA) as a variational inference problem, enabling JEPA-style models to both predict latent representations and generate samples—bridging the theoretical gap between LeCun's predictive SSL camp and likelihood-based generative SSL. The unification allows researchers to train a single model with properties of both paradigms. Architecturally significant for representation learning researchers deciding between SSL approaches.