4/21/26 (virtual) Dr. Yongchao Huang, University of Aberdeen
Title: A family of joint-embedding predictive architectures
Abstract: We present a probabilistic formulation of Joint-Embedding Predictive Architectures, VJEPA, as a latent predictive world model. Unlike deterministic JEPA, VJEPA learns a full predictive distribution over future representations via a variational objective, establishing connections to predictive state representations and Bayesian filtering while avoiding observation-level likelihoods. This yields latent states that act as sufficient information states for prediction and control, with explicit uncertainty quantification and robustness to high-variance nuisance variability. From an information-theoretic perspective, VJEPA optimises a predictive information bottleneck, retains information relevant for future prediction while compressing irrelevant input structure. We further extend this framework in two directions: BiJEPA enforces bidirectional predictive consistency to capture reversible structure in data, and RiJEPA incorporates symbolic rules as energy-based constraints that geometrically shape the latent space for interpretable, neuro-symbolic representation learning. Together, they provide a unified view of predictive representation learning as probabilistic belief propagation and latent dynamics modelling.