Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture – Bridging Predictive and Generative Self-Supervised Learning

The Joint-Embedding Predictive Architecture (JEPA) is often seen as a non-generative alternative to likelihood-based self-supervised learning, emphasizing prediction in representation space rather than reconstruction in observation space. We argue that the resulting separation from probabilistic generative modeling is largely rhetorical rather than structural: the canonical JEPA design–coupled encoders with a context-to-target predictor–mirrors the variational posteriors and learned conditional priors obtained when variational inference is applied to a particular class of coupled latent-variable models, and standard JEPA can be viewed as a deterministic specialization in which regularization is imposed via architectural and training heuristics rather than an explicit likelihood. Building on this view, we derive the Variational JEPA (Var-JEPA), which makes the latent generative structure explicit by optimizing a single Evidence Lower Bound (ELBO). This yields meaningful representations without ad-hoc anti-collapse regularizers and allows principled uncertainty quantification in the latent space. We instantiate the framework for tabular data (Var-T-JEPA) and achieve strong representation learning and downstream performance, improving over T-JEPA across real-world tabular benchmarks while remaining competitive with strong raw-feature baselines.

For machines to truly understand the world, they need to learn from raw data without humans labeling every example. One popular recipe, called JEPA, teaches a model to look at part of an input – say, the top half of an image – and predict an abstract summary of the rest, rather than predicting every pixel. JEPA works well, but it has a famous failure mode: the model can cheat by outputting the same answer for everything, so researchers have to bolt on extra tricks to keep it honest. JEPA has also been positioned as fundamentally different from older "generative" methods that try to recreate data, which has kept these two families of techniques apart. We show that this divide is largely a matter of framing, not substance: if you write JEPA down using the math of probability, it turns out to be a special case of a well-known generative model in disguise. Building on this insight, we designed Var-JEPA, a version that naturally avoids the cheating problem and can also tell you how confident it is in its predictions. On real-world tables of data, it performs well compared to the standard approach.