A Self-supervised learning (SSL) framework proposed by Yann LeCun
How it works
Instead of predicting raw input (like pixels), a JEPA learns by predicting the representation of one part of an input from another part, entirely in latent space.
- Both context and target pass through an encoder (often shared or EMA-updated)
- A predictor maps the context embedding to the target embedding
- Loss is computed in embedding space, not pixel space
Collapse Prevention
A key challenge is representation collapse (everything maps to the same vector). JEPAs handle this by using a momentum encoder (like BYOL) for the target branch rather than contrastive negatives.