JEPA

JEPA (Joint Embedding Predictive Architecture) was systematically proposed by Yann LeCun in 2022 in the paper 'A Path Towards Autonomous Machine Intelligence'

JEPA makes predictions in an abstract embedding space rather than at the pixel or token level, avoiding computational waste.

The concept of the world model originates from the autonomous machine intelligence architecture proposed by Yann LeCun; its core idea is to enable AI to internally construct a simulator of the operating rules of the physical world in order to predict the consequences of actions.

Yann LeCun is Meta's Chief AI Scientist and regards world models as a key component on the path to artificial general intelligence.

Institutions such as DeepMind and Meta AI have produced representative world model works including DreamerV3 and JEPA

SPERA is an EEG foundation model that introduces a spherical prior and uses a Joint Embedding Predictive Architecture (JEPA) to perform predictions in latent space rather than reconstructing the observed signal.

JEPA was first systematically proposed by Yann LeCun's team, with representative implementations including I-JEPA and Video JEPA for vision.

JEPA (Joint-Embedding Predictive Architecture) is the self-supervised learning framework championed by Yann LeCun's team, which predicts the representation of another part of the input in latent space rather than predicting raw content.

JEPA (Joint Embedding Predictive Architecture), which Yann LeCun has strongly advocated for in recent years, is an extension of the world model approach.

The alternative path toward artificial general intelligence proposed by LeCun is a self-supervised learning architecture driven by a 'World Model', such as JEPA

JEPA is a direction strongly advocated by Yann LeCun, who has publicly criticized the autoregressive LLM paradigm on multiple occasions for being unable to lead to true machine intelligence.

The JEPA architecture consists of two parts—a context encoder and a target encoder—with a predictor completing prediction tasks in the embedding space between the two.

Relying solely on latent distance to guide action selection is insufficient, and new representation learning objectives or planning algorithms need to be explored.

The root cause of the action ordering defect lies in the representational capability of the JEPA architecture itself, rather than in the choice of a specific visual encoder.

The research indicates that the root cause of action-ranking failure lies in the representational capacity of the JEPA architecture itself, rather than in the choice of specific visual encoder.

The research team developed the ARC-Bench benchmark specifically to audit whether JEPA-style world models correctly rank candidate actions.

Closed-loop replanning masks the action-ranking defects of JEPA world models.

Closed-loop success rates systematically overestimate the rankability of frozen latent representations.

Closed-loop success rate has serious limitations as a single evaluation metric, and relying solely on latent distance to guide action selection is insufficient.

Yann LeCun proposed the Joint Embedding Predictive Architecture (JEPA) in 2022, arguing that true intelligence requires predicting the evolution of the world in an abstract representation space rather than pixel space.