What Is Yann LeCun Cooking? JEPA Explained Simply

JEPA (Joint Embedding Predictive Architecture) is explained as an AI framework that predicts abstract latent representations rather than raw pixels or tokens. Unlike LLMs that predict the next token, JEPA encodes two views of the same input and trains a predictor to match their embeddings in latent space, filtering out irrelevant noise. The architecture has three components: a context encoder, a target encoder, and a predictor. Key challenges include representation collapse, addressed historically via EMA (exponential moving average) and later through contrastive methods like SimCLR, Barlow Twins, and VICReg. The latest approach, LEGO (released November 2025), constrains embeddings to follow an isotropic Gaussian distribution, avoiding collapse without EMA. JEPA is less suited for text (already symbolic and low-noise) but shows strong promise in medical imaging, particularly echocardiography, where noise reduction in latent space aligns with clinically meaningful anatomical signals.