Distillation-guided optical neural networks with reinforcement learning-assisted calibration
Optica · 20.08.2026 · 13 (8) · 1423
DOI: 10.1364/optica.591264 ↗Kurz gefasst
Fachartikel in der Zeitschrift Optica, online erschienen am 22.07.2026. Verfasst von Kangjian Di und 10 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fundamental Research Funds for the Central Universities, Jiangsu Innovation Teams and AI & AI for Science Project of Nanjing University.
Abstract (Originalsprache)
Optical neural networks offer a promising route to overcome the inference-speed and energy-efficiency bottlenecks of conventional electronic neural networks, owing to their high parallelism and low-latency optical computation. However, practical on-chip training of optical neural networks remains challenging because gradient estimation, hardware nonidealities, and device-compatible parameter updates are difficult to handle directly in optical systems. Here, we propose an on-chip training distillation-guided optical neural network (DGONN) and introduce a forward distillation algorithm to address these challenges. The forward distillation algorithm avoids chain-rule-based optical backpropagation and enables hardware-aware gradient estimation through measured forward differences, thereby mitigating the instability of long-range gradient propagation in photonic chips. The block-MZI optical parameterization reduces the trainable parameter count compared with dense electronic neural networks, while layer-wise distillation recovers the accuracy loss caused by this constrained optical parameterization and improves training stability. Numerical and experimental results demonstrate the high classification accuracy of our DGONN system across various datasets [CIFAR-10, Fashion-MNIST, Handwriting-MNIST, free spoken digit dataset (FSDD), and distributed acousto-optic sensing (DAS)]. Additionally, deep reinforcement learning further enables the controllability and stability of the on-chip optical system. Our work could pave the way for efficient ONNs, offering a promising approach to overcoming the training challenges of optical neural networks and enabling the potential development of energy-efficient, low-power AI systems using photonic technology.
Förderung
- National Natural Science Foundation of China
- Fundamental Research Funds for the Central Universities
- Jiangsu Innovation Teams and AI & AI for Science Project of Nanjing University
Zitieren
APA
Di, K., Yu, F., Chen, S., Li, J., Liu, A., Shao, S., Shi, Z., Zhang, X., Zhang, Y., Jiang, W., & Zou, N. (2026). Distillation-guided optical neural networks with reinforcement learning-assisted calibration. Optica, 13(8), 1423. https://doi.org/10.1364/optica.591264
Vancouver
Di K, Yu F, Chen S, Li J, Liu A, Shao S, et al. Distillation-guided optical neural networks with reinforcement learning-assisted calibration. Optica. 2026;13(8):1423. doi:10.1364/optica.591264
BibTeX
@article{Di2026,
author = {Di, Kangjian and Yu, Fuhao and Chen, Silin and Li, Jiashu and Liu, Andy and Shao, Sen and Shi, Zhiyuan and Zhang, Xuping and Zhang, Yixin and Jiang, Wei and Zou, Ningmu},
title = {Distillation-guided optical neural networks with reinforcement learning-assisted calibration},
journal = {Optica},
year = {2026},
volume = {13},
number = {8},
pages = {1423},
doi = {10.1364/optica.591264}
}
EUVIMED