JEPA

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

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 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

The appeal of latent world models lies in their high computational efficiency and suitability for long-horizon rollouts; the Dreamer series and methods based on the Joint Embedding Predictive Architecture (JEPA) both belong to this lineage.

JEPA has already demonstrated strong performance in the image, video, and audio domains.

In EHR sequences, high-level abstractions of patient state (such as disease progression, comorbidity burden, treatment response) lack an accepted, computable representational definition, making them difficult to convert into dense latent targets suitable for JEPA prediction.

In the study's experiments, all three JEPA variants trained stably without representation collapse.

The paper points out that the bottleneck of JEPA on EHR is not the magnitude of prediction error, but whether the predicted latent target itself has clinical significance.

On multiple downstream evaluation tasks in the study, none of the JEPA variants outperformed standard MLM.

JEPA-type methods typically employ asymmetric architectures (such as momentum encoders, stop-gradient) or regularization terms (such as VICReg's variance-covariance constraints) to prevent representation collapse.

The study suggests that the real breakthrough may lie in designing intermediate representations that can bridge clinical semantics and latent-space prediction, rather than changing the prediction target.

The study counterintuitively found that target size (rather than target heterogeneity) is the main factor driving up JEPA prediction error.

The study shows that even reducing target size cannot restore JEPA's downstream performance on EHR.

Quantum-JEPA uses a conditional quantum Born prior to learn the distribution of future representations, introducing a trainable quantum component into JEPA.

JEPA-style latent world models typically contain an encoder, a predictor, and an action-conditioning module, and their prediction target is a low-dimensional latent vector.

The idea of the LFP head originates from the world model research tradition in self-supervised learning, such as the Dreamer series and JEPA (Joint Embedding Predictive Architecture).

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.