The Annotated JEPA
Summary
The Annotated JEPA provides an in-depth, annotated, from-scratch walkthrough of Joint Embedding Predictive Architectures (JEPA), tracing from LeCun's energy-based framing to I-JEPA for images, V-JEPA for video, and LeJEPA for distribution-regularized training. It explains predicting representations rather than reconstructing pixels, anti-collapse mechanisms (EMA teachers, stop-gradients, distribution regularizers), and the JEPA template with encoders, predictor, and a distance loss. The piece also covers implementation details, a PyTorch outline, and considerations for world models and planning.