High-speed and versatile ONN through parametric-based nonlinear computation
- 30 May 2025
- journal article
- Published by Optica Publishing Group in Photonics Research
- Vol. 13 (6) , 1647
Abstract
Neural networks (NNs), especially electronic-based NNs, have been rapidly developed in the past few decades. However, the electronic-based NNs rely more on highly advanced and heavy power-consuming hardware, facing its bottleneck due to the slowdown of Moore’s law. Optical neural networks (ONNs), in which NNs are realized via optical components with information carried by photons at the speed of light, are drawing more attention nowadays. Despite the advantages of higher processing speed and lower system power consumption, one major challenge is to realize reliable and reusable algorithms in physical approaches, particularly nonlinear functions, for higher accuracy. In this paper, a versatile parametric-process-based ONN is demonstrated with its adaptable nonlinear computation realized using the highly nonlinear fiber (HNLF). With the specially designed mode-locked laser (MLL) and dispersive Fourier transform (DFT) algorithm, the overall computation frame rate can reach up to 40 MHz. Compared to ONNs using only linear computations, this system is able to improve the classification accuracies from 81.8% to 88.8% for the MNIST-digit dataset, and from 80.3% to 97.6% for the Vowel spoken audio dataset, without any hardware modifications.Keywords
Funding Information
- Research Grants Council of the Hong Kong Special Administrative Region of China (HKU 17212824, HKU 17210522, HKU C7074-21G, HKU R7003-21, HKU 17205321)
- Innovation and Technology Commission—Hong Kong (MHP/073/20, MHP/057/21, Health@InnoHK program)
This publication has 25 references indexed in Scilit:
- Activation functions in deep learning: A comprehensive survey and benchmarkNeurocomputing, 2022
- Electromagnetic wave-based extreme deep learning with nonlinear time-Floquet entanglementNature Communications, 2022
- Scalable optical learning operatorNature Computational Science, 2021
- Highly accurate protein structure prediction with AlphaFoldNature, 2021
- Inference in artificial intelligence with deep optics and photonicsNature, 2020
- Is your dataset big enough? Sample size requirements when using artificial neural networks for discrete choice analysisJournal of Choice Modelling, 2018
- Time stretch and its applicationsNature Photonics, 2017
- Resolving the build-up of femtosecond mode-locking with single-shot spectroscopy at 90 MHz frame rateNature Photonics, 2016
- Energy-efficient communicationNature Photonics, 2011
- Serial time-encoded amplified imaging for real-time observation of fast dynamic phenomenaNature, 2009