Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 3: Research with AI: Hardware, Software, Tools
AKPIK 3.4: Hauptvortrag
Dienstag, 10. März 2026, 11:45–12:15, BEY/0127
Model-free training of optical neural networks based on multimode semiconductor lasers — •Anas Skalli1, Satoshi Sunada2, Mirko Goldmann1, Marcin Gebski3, Nasibeh Haghighi4, Stephan Reitzenstein4, James A. Lott4, Tomasz Czyszanowski3, and Daniel Brunner1 — 1FEMTO-ST Institute / Optics Department, CNRS & Université Marie et Louis Pasteur, 15B avenue des Montboucons, 25030 Besançon Cedex, France. — 2Faculty of Mechanical Engineering, Institute of Science and Engineering, Kanazawa University, Kakuma-machi Kanazawa, Ishikawa 920 1192, Japan. — 3nstitute of Physics, Lodz University of technology, Wólczańska 217/22190-005 Łódź, Poland — 4Institut für Festkörperphysik, Technische Universität Berlin, Hardenbergstraße 36, 10623 Berlin, Germany
In this work, we demonstrate a fully autonomous and parallel optical neural network (ONN) based on a multimode VCSEL architecture using off-the-shelf components. The system is scalable in both network size and inference bandwidth, paving the way toward GHz-level computing. Our ONN supports in-situ learning, making it closer to true autonomous operation. To unlock the full potential of our hardware, we develop and benchmark several hardware-compatible optimization algorithms, including SPSA and PEPG, demonstrating their suitability for physical systems with limited computational resources, and showing how algorithmic choices impact convergence speed, scalability, and energy cost. Our ONN outperforms both a linear hardware baseline and a digital linear classifier on the MNIST task.
Keywords: Optical Neural Networks; Photonic Neural Networks; Online training; Black box optimization