Megadose The AI curve, compressed.

Func-JEPA: Learning a Sobolev Space of Functions as Latent Vectors

· ArXiv · AI/CL/LG ·
One frozen encoder serves several mathematical tasks, with its strongest reported result in recovering exact formulas from noisy samples.

Func-JEPA learns vectors from symbolic functions, shaping their geometry to imitate the Sobolev space H¹ and learning differentiation as a latent operator. Lightweight modules reuse the encoder for decoding, composition, integration and regression from samples. Evaluations cover functions within and outside the pretraining distribution. It recovers exact formulas from noisy samples more often than the symbolic-regression methods compared, but computer algebra and task-specific models remain stronger on standard integration and differential-equation benchmarks.

ArXiv · AI/CL/LG's note

score 5

Categories: Research