Distillation-guided Optical Neural Networks with Reinforcement Learning-assisted Calibration

Kangjian Di, Fuhao Yu, Silin Chen, Jiashu Li, Andy Liu, Sen Shao, Zhiyuan Shi, Xuping Zhang, yixin zhang, W. Jiang

Optical neural networks (ONNs) promise ultra-fast and energy-efficient computing but are hampered by the critical challenge of on-chip training. Here, we propose an on-chip training distillation-guided optical neural network (DGONN) and introduce a forward distilled algorithm to address these challenges. The forward distillation algorithm mitigates the instability of blind gradient propagation in photonic chips. Our on-chip distillation scheme reduces the number of parameters compared to electronic deep neural networks, thus enhancing training efficiency and accuracy relative to traditional optical training methods. 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 photonic neural networks and enabling the potential development of energy-efficient, low-power AI systems using photonic technology.