Inspiration

We wanted to see whether we could get a JEPA world model to learn underlying physical laws, such that it could generalise to show a physical phenomenon not in its training data.

What it does

We trained a JEPA model to predict a latent next frame which is then decoded into pixel data, from numerical simulations of Barkley's reaction-diffusion model.

How we built it

Trained a JEPA model on wave dynamics - excluding the spiral dynamics we hope it will learn autonomously. The JEPA model is trained to predict latents, encouraging the model to learn the underlying dynamics of the system rather than the surface behaviour.

Challenges we ran into

Time was incredibly tight, especially as we were training the full JEPA model from scratch

Accomplishments that we're proud of

We're proud that the model seemed to be able to generalise.

What we learned

We learned how to work quickly under time pressure. Additionally, neither of us had any previous experience with JEPA models, and it was interesting to get some hands-on experience with them.