Overview
Introduced by Yann LeCun in 2022 (“A Path Towards Autonomous Intelligence”), the Joint Embedding Predictive Architecture (JEPA) is a blueprint for self-supervised learning. Unlike generative models (like masked autoencoders) that try to predict raw pixels, JEPAs learn to predict abstract representations in a latent space.
By predicting features rather than reconstructing noisy pixels, JEPAs build a highly efficient, semantic understanding of the world.
The JEPA Family
I-JEPA (Image)
Learns by creating an internal model of the outside world, comparing abstract representations of image patches rather than comparing the pixels themselves.
V-JEPA (Video)
A collection of vision models trained solely using a feature prediction objective on 2 million videos. It operates by passively watching videos and predicting what happens next in the latent space. Crucially, it does not use text or image encoders for supervision.
EchoJEPA (Echocardiography Foundation Model)
An application of the V-JEPA architecture specifically for medical imaging. Trained on 18 million echocardiograms, it solves a massive problem in ultrasound: speckle noise. By leveraging JEPA’s latent predictive objective, EchoJEPA learns robust anatomical representations that completely ignore stochastic speckle and acquisition artifacts, outperforming prior models even when trained on just 1% of labeled data.
TODO: Add block diagram of the Joint Embedding prediction loss.