Principles and Metrics of Photonic Learning Machines
M Hary1, A Skalli1, L Cardenas-Razo1, A Ermolaev1, M Marciniak2, M Gebski2, J Lott2, G Genty3, T Czyszanowski2, J M Dudley1, D Brunner1
1 Département d’Optique P. M. Duffieux, FEMTO-ST Institute, UMR CNRS 6174, Besançon, France
2 Institute of Physics, Lodz University of technology, Wólczańska 217/22190-005 Łódź, Poland, Lodz, Poland
3 Department of Physics, Tampere University of Technology, Tampere, Finland
Seminar: S12 — Optical Computing and Neural Networks
Tuesday, 7 July 2026 · 14:30 – 15:00
Abstract
Photonic systems have emerged as promising alternatives to electronic computing. Deep physical neural networks trained with physics-aware backpropagation show that nonlinear optical systems can function as neural networks, leveraging inherent parallelism, energy efficiency, and speed. Unconventional platforms such as optical fibers and semiconductor lasers demonstrate significant computing capabilities.
We characterize two such systems, a highly nonlinear fiber (HNLF) and a vertical-cavity surface-emitting laser (VCSEL), using several metrics including dimensionality, which measures the effective degrees of freedom, and consistency, which assesses the reproducibility of the response. The HNLF reaches up to 100 principal components and achieves $87\%$ accuracy on the MNIST dataset, while VCSELs exhibit similar parameter-dependent computational scaling.