SJEPA Learns Elegant Latent Dynamics with Hybrid Symbolic-Neural AI.

Researchers have introduced SJEPA, a novel Joint-Embedding Predictive Architecture (JEPA) that focuses on learning abstract states and their underlying dynamics. Unlike traditional JEPA models that use opaque neural maps for transitions, SJEPA employs a hybrid approach, combining a symbolic law with a neural correction mechanism. This allows the model to learn predictive representations whose induced dynamics can be described compactly and symbolically. The core principle of SJEPA is to identify the simplest adequate dynamics. It achieves this by applying representation constraints to preserve informative predictive coordinates and favoring low-complexity symbolic-neural transitions. Experiments, particularly with pendulum simulations, show that SJEPA discovers significantly simpler symbolic dynamics, leading to better long-horizon rollout error and divergence compared to post-hoc fitting methods. The neural component handles residual dynamics when the symbolic grammar is misspecified.