| Part of a series on |
| Machine learning and data mining |
|---|
Joint-Embedding Predictive Architecture (JEPA, pronounced "yepa") is a family of neural network architectures for self-supervised learning, proposed by Yann LeCun and collaborators at Meta. Rather than reconstructing raw inputs (as in autoencoders or diffusion models), a JEPA model learns to predict the latent representation of a target from the latent representation of a context in a learned embedding space.
The approach is non-generative: it operates entirely in representation space and therefore avoids modelling unpredictable detail. JEPA is a central component of LeCun's proposed "world model" framework for building more general artificial intelligence.
In a 2022 position paper, "A Path Towards Autonomous Machine Intelligence", LeCun argued that human-level artificial intelligence requires machines to learn predictive world models rather than relying on pattern matching alone. He proposed JEPA as a practical architecture for such predictive models. The framework splits a standard encoder–decoder into a learned embedding space and a predictive module that operates within it.
A JEPA model has two main components:
Because prediction happens in embedding space, the model only needs to capture structure that is predictable and can discard irrelevant detail.
Image-based JEPA (I-JEPA) was introduced in a 2023 CVPR paper. From a single context block of an image, it predicts the representations of multiple target blocks, learning semantic image representations without hand-crafted data augmentations.
Video-based JEPA (V-JEPA) was released by Meta in 2024, trained on more than one million hours of video, for learning visual representations from video.
V-JEPA 2 was unveiled by Meta in June 2025 as an extension of V-JEPA achieving state-of-the-art visual understanding and physical reasoning, with applications to robotics and self-driving cars. Meta also released three benchmarks for evaluating physical reasoning from video.
JEPA is an instantiation of the predictive "world model" idea set out in LeCun's 2022 position paper. The broader concept is covered at World model (artificial intelligence).