Theoretical Design and On Chip Implementation of Photonic Neural Network Based on Multi-dimensional Multiplexing Technology

Professor Wong, Kenneth Kin Yip   (Principal Investigator (PI))

24

2022-09-01

2106800

Theoretical Design and On Chip Implementation of Photonic Neural Network Based on Multi-dimensional Multiplexing Technology

Theoretical Design, On Chip Implementation of Photonic Neural Network, Multi-dimensional Multiplexing Technology

Electrical and Electronic Engineering

MHP/057/21

Mainland-Hong Kong Joint Funding Scheme (MHKJFS)

2022

On-going

This project focuses on the significant needs for high-speed data processing computingtechnology in application scenarios such as autonomous driving, medical diagnosis, andmachine control. It will unite three units in the Mainland and Hong Kong to fully play thecomprehensive advantages of theoretical research, experimental research, and on-chipintegration. It will accomplish the design of a large-scale photon neural network based ona multi-dimensional multiplexing technology photon neural network. It will realize theelectrical domain mathematical operations such as multiplication, summation, andnonlinearity on optical analog. It will develop critical technologies such as network design,fabrication, packaging, and testing. It will build a complete set of on-chip photonic neuralnetworks that can be practically applied to high-speed computing scenarios, laying anessential technical foundation for high-speed data operations.The purpose of this project is to: 1) Based on the multi-dimensional multiplexingtechnology, a large-scale photonic neural network system is designed. The maximumsingle-layer neuron node number of the network is required to be greater than 400, andthe number of network layers is greater than 5 layers; 2) Using data sets such as MNISTto complete experimental demonstrations of optical fiber-based photonic neural networksin the laboratory, designing optical waveguides and combining micro-nano processingtechnology to achieve on-chip multi-dimensional multiplexing photonic neural networksand time-lens algorithm which targets on preprocessing of the input data; 3) Combinedwith the time-lens method, the weight matrix optimization was implemented bycompressing the time signal to boost up the classification accuracy. Therefore, it issupposed the calculation speed to be at least an order of magnitude higher than that ofelectrical neural networks. The classification accuracy is close to or reaches the accuracyof the electrical neural network.