JEPA: Learning Representations and Dynamics in Embedded Space
Definition
A framework combining classical optimal control with modern representation learning. Instead of predicting pixel values or reconstructing inputs (generative models), JEPA encodes observations into embedding space and trains a predictor to forecast future embeddings given current embeddings and actions. This creates world models that can plan through search over action sequences rather than autoregressive generation.
Cast within
No topic-region cast yet — this appears once JEPA: Learning Representations and Dynamics in Embedded Space's compiled claims are aligned into a topic region's argument tree.
Claims mentioning JEPA: Learning Representations and Dynamics in Embedded Space (1)
Your notes
Loading notes...