Stanford paper: coding agents train optical neural networks in situ from measurements alone

StanfordUniversity · hf · 2026-11-21

Stanford and EPFL researchers introduce "optics as code": a coding agent trains an optical neural network from hardware measurements alone by reparametrizing it with interpretable optical elements. It beats gradient-free methods on classification and image encoding at lower cost, approaching backprop through a calibrated model.

Related event: Coding Agents Train Optical Neural Networks In Situ, Lifting Accuracy from 35% to 96.8%(7 posts)→

Original post →

More from coding & agent

coding & agent channel →