Training real hardware, not a digital twin: learning directly in nonlinear wave systems

bravo_abad · x · 2026-09-24

Sajnok and Matuszewski propose bypassing the "build a digital model, train it, transfer parameters back" pipeline by letting the physical system measure its own learning direction, training directly on nonlinear wave hardware.
Procedure: run the device once normally, then slightly perturb its output toward the desired answer and run again; the difference between the two physical states tells you how experimental parameters like the optical potential should change.
This avoids needing an accurate digital replica of the full nonlinear dynamics, sidestepping model-hardware mismatch that accumulates as systems grow complex.

More from Embodied