PHOENIX: Photonic Distillation Transfers Electronic Knowledge to Hybrid Optical Neural Networks

Shaoqing Xu, Fang Li, Mingyuan Li, Feng Zijia, Zhecheng Zhu, Shengyin Jiang, Xing Mou, Qisheng Yang, Zhi-Xin Yang

Keywords: Optical Neural Networks, Physical Neural Networks, Electronic Neural Networks, Object Detection, Distillation

TL;DR: we proposed a general photonic distill paradigm that pioneers the successful application of Optical Neural Networks (ONNs) for industrial-grade object detection.

Abstract: As artificial intelligence (AI) systems continue to scale in both complexity and dataset size, conventional electronic hardware faces significant challenges in meeting the demands of low-latency, high-throughput, and energy-efficient processing, particularly for industrial deployments. However, sustaining such scaling is increasingly constrained by the physical and energy limitations of electronic computing. Optical Neural Networks (ONNs), leveraging the superior physical properties of photons, offer inherent advantages such as ultra-fast processing speed, massive parallelism, and near-zero power consumption, which have already demonstrated potential on simple tasks in small datasets like MNIST classification. In this work, we presented the first optoelectrically fused neural network deployment framework PHOENIX for object detection tasks, demonstrating its performance in industrial-level large datasets (e.g., COCO) and benchmark models. Compared to state-of-the-art electronic models, our solution achieved approximately 85.0% accuracy. The accuracy was further improved to 93.0% through our novel knowledge distillation strategy. Furthermore, we achieved 72.6% energy reduction and 11.3× speed acceleration compared to equivalent edge GPUs by successfully transferring spatial attention knowledge from the electronic domain to the photonic domain, making it an ideal choice for real-time, energy-critical industrial applications. This technique not only bridges the performance gap but also offers an alternative physically interpretable platform for AI. Our universal framework paves the way for extending ONN deployment to a wider range of deep learning models and applications, whether based on CNN or Transformer architectures, providing a compelling choice for real-time, energy-critical scenarios such as autonomous driving, smart surveillance, and industrial automation. Source code is available a thttps://github.com/Anon-BOTs/Distill-Hybrid-Optoelectronic.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15000