Joint Embedding Predictive Architecture (JEPA)
Definition
Neural architecture solving video prediction without pixel-level generation. Both observation X and future observation Y are encoded into abstract representation spaces via encoders. Rather than predicting pixel values directly, prediction occurs in representation space: predictor takes encoded current state and predicted action, outputs predicted future representation. Encoder learns to discard unpredictable details (texture, appearance) while preserving predictable structure (physics, geometry).
Cast within
No topic-region cast yet — this appears once Joint Embedding Predictive Architecture (JEPA)'s compiled claims are aligned into a topic region's argument tree.
Claims mentioning Joint Embedding Predictive Architecture (JEPA) (20 of 24)
Your notes
Loading notes...