Ultrafast all-optical biomimetic encoder for spiking neural networks
用于尖峰神经网络的超快全光仿生编码器
スパイクニューラルネットワーク用超高速全光生体認証エンコーダ
스파이킹 신경 네트워크를 위한 초빠른 모든 광학 생물 모형 인코더
Codificador biomimético totalmente óptico ultrarápido para redes neuronales de picos
Encodeur biomimétique ultra-rapide tout optique pour réseaux neuronaux à pics
Ультрабыстрый полнооптический биомиметический кодер для спиковых нейронных сетей
Qijun Bian ¹ ², Fangchen Hu ³, Junlian Jia ¹ ², Aolong Sun ¹ ², Sizhe Xing ¹ ², Boyu Dong ¹ ², An Yan ¹ ², Ziwei Li ¹ ², Jianyang Shi ³, Nan Chi ¹ ², Junwen Zhang ¹ ²
¹ Future Information Innovative College (FIIC), Fudan University, Shanghai 200433, China
中国 上海 复旦大学未来信息创新学院
² Key Laboratory for Information Science of Electromagnetic Waves (MoE), Fudan University, Shanghai 200433, China
中国 上海 复旦大学电磁波信息科学教育部重点实验室
³ Zhangjiang Laboratory, Shanghai 200232, China
中国 上海 张江实验室
Spiking neural networks (SNNs) represent a brain-inspired computing paradigm with inherent advantages in low latency and energy efficiency, where high-fidelity conversion of analog sensory signals into optical spike trains remains a critical bottleneck for photonic neuromorphic deployment. In this work, we experimentally demonstrate an all-optical biomimetic spike encoder based on optically injected Fabry–Perot (FP) laser neuron, enabling amplitude-to-rate linear encoding with a spike rate beyond 2.0 GHz and an encoding precision of ≈5 bits.
Different from previous FP laser spiking devices that only support binary spike generation, our scheme establishes a robust monotonic mapping between input optical intensity and spike firing rate, and exhibits excellent compatibility with standard 1550 nm fiber-optic communication infrastructure. We validate the encoding performance via handwritten digit recognition on the MNIST dataset using a digital spiking neural network, achieving a classification accuracy over 94.1%. Beyond static image encoding, the proposed encoder is capable of converting continuous physiological and audio signals into well-behaved spike sequences.
We further construct a scalable optical neuromorphic architecture compatible with time- and wavelength-division multiplexing, supporting multi-user physiological signal access and intelligent medical diagnosis including X-ray radiography, ECG arrhythmia analysis, and heart sound classification. This work not only unveils the laser dynamical mechanism of amplitude-rate encoding in injected FP lasers, but also provides a feasible route to integrate photonic spiking encoders with existing fiber-optic networks for scalable, ultrafast, low-power neuromorphic healthcare and communication applications.
