JEPA

JEPA (Joint Embedding Predictive Architecture) is a machine learning architecture proposed by Yann LeCun in his 2022 position paper A Path Towards Autonomous Machine Intelligence, which learns by predicting abstract representations of inputs in a latent space instead of predicting raw pixels or tokens.^[1][2] The goal is to build systems that form internal models of how the world works, closer in LeCun's framing to how humans and animals learn, rather than memorising surface patterns.^[1][2]

History

LeCun introduced JEPA as a component of a broader research agenda for autonomous machine intelligence in 2022, arguing that contemporary AI, including large language models, lacks the ability to learn common-sense background knowledge efficiently from passive observation.^[2] In June 2023, Meta AI published I-JEPA (Image JEPA), the first AI model based on the architecture, which learns from a single context block of an image by predicting the representations of other target blocks in the same image, without relying on hand-crafted data augmentations.^[2][3]

The JEPA family has since expanded: MC-JEPA (motion and content), V-JEPA (video), and later V-JEPA 2, VL-JEPA (vision-language), H-JEPA (hierarchical) and LeJEPA (LeCun's later general formulation).^[1][4]

Key Concepts

JEPA operates on a simple principle: rather than predicting what comes next in pixel or token space, a JEPA predicts a representation of a future observation from a representation of a past one, in a learned latent space.^[1][2] This lets the predictor ignore unpredictable surface detail — pixel noise, lighting changes, paraphrase variation — and focus on the semantically meaningful structure of the signal.^[1]

Key concepts include:

• Latent-space prediction: predictions happen in embedding space, not raw input space, which removes the pressure to generate every pixel or token.^[1][2]
• Self-supervised learning: JEPA models are trained without labels, using the structure of the data itself as the supervisory signal.^[2][3]
• World models: LeCun positions JEPA as the architectural backbone for world models — internal simulators that let an agent predict and plan.^[1][4]
• Efficiency: I-JEPA was reported to be much more computationally efficient than widely used generative computer vision models while achieving strong performance on downstream tasks.^[2]

Applications and Impact

JEPA's most immediate applications are in computer vision, where I-JEPA and V-JEPA produce semantic representations used for image classification, video understanding and activity recognition.^[1][2][4] Because JEPA models are non-generative and self-supervised, they are seen as a candidate path toward sample-efficient learning and planning in embodied AI, robotics and autonomous driving.^[1][4] Within Meta AI, JEPA represents a research direction distinct from large language models, focused on learning common sense through observation.^[2]

>See Also

• Self-Supervised Learning
• World Models
• Transformer Architecture
• Meta AI
• Diffusion Model

References