All-digital optical computing paradigm with ultra-high calculating precision for optical neural network applications
Lei Yue, Xilong Zeng, Zhifeng Yang, Xinlan Xiong, Hai Zhong, Yongkun Song, Jiajia Zhao · Optics Express · 2025
Photonics emerges as a paradigmatic enabler for artificial intelligence (AI) and neuromorphic computing architectures, furnishing an avenue for ultra-low-latency, high-bandwidth, and energy-efficient information processing through the manipulation of various photonic degrees of freedom with an optical signal processing procedure. However, conventional optical neural network (ONN) architectures face significant challenges in calculating precision induced by insufficient precision of optoelectronic devices such as digital-to-analog converters (DACs) and analog-to-digital converters (ADCs). Here, we propose an all-digital optical neural network (ADONN) architecture that employs digital binary signals at the optical input of the computation core, with entirely logical signal transmission and processing in the optical domain. Despite its simple architecture, the calculation precision of ONN is greatly improved even at a high computational speed, while simultaneously alleviating the rigorous requirement of optoelectronic devices. To validate our scheme, we build an ADONN architecture through numerical simulation and systematically investigate its comprehensive performance compared with traditional analog ONN architecture. The ADONN scheme achieves a root-mean-square error (RMSE) of 3 × 10 −4 at the optical signal-to-noise ratio (OSNR) of 20 dB through a basic optical multiplication test experiment, which corresponds to an 11-bit calculating precision, and is tremendously lower than that of traditional analog implementation. Meanwhile, the edge detection task for high-fidelity image processing is introduced to demonstrate the high precision of ADONN; image convolution with 8-bit resolution is performed using various convolution kernels, and simulation results exhibit that all pixel error rates (PERs) are below 1% with an all-digital computing core. Furthermore, an all-digital optical convolutional neural network (ADOCNN) is also developed and evaluated in the MNIST task with superior accuracy results. Results confirm that the proposed all-digital optical computing paradigm could provide ultra-high calculating precision with relatively low hardware complexity, which could be potentially applied in various modern ONN applications.