Annealing-inspired training of an optical neural network with ternary weights

• DOI: 10.1038/s42005-025-01972-y
• Publication Date: 2025-02-13T12:40:28Z
• Authors: Anas Skalli; Mirko Goldmann; Nasibeh Haghighi; Stephan Reitzenstein; James A. Lott; Daniel Brunner
• Keywords: QB460-466; FOS: Computer and information sciences; Emerging Technologies (cs.ET); Physics; QC1-999; Computer Science - Emerging Technologies; Astrophysics

Abstract Artificial neural networks (ANNs) represent a fundamentally connectionist and distributed approach to computing, and as such they differ from classical computers that utilize the von Neumann architecture. This has revived research interest in new unconventional hardware for more efficient ANNs rather than emulating them on traditional machines. To fully leverage ANNs, optimization algorithms must account for hardware limitations and imperfections. Photonics offers a promising platform with scalability, speed, energy efficiency, and parallel processing capabilities. However, fully autonomous optical neural networks (ONNs) with in-situ learning are scarce. In this work, we propose and demonstrate a ternary weight high-dimensional semiconductor laser-based ONN and introduce a method for achieving ternary weights using Boolean hardware, enhancing the ONN’s information processing capabilities. Furthermore, we design an in-situ optimization algorithm that is compatible with both Boolean and ternary weights. Our algorithm results in benefits, both in terms of convergence speed and performance. Our experimental results show the ONN’s long-term inference stability, with a consistency above 99% for over 10 h. Our work is of particular relevance in the context of in-situ learning under restricted hardware resources, especially since minimizing the power consumption of auxiliary hardware is crucial to preserving efficiency gains achieved by non-von Neumann ANN implementations.