Photonic neural network training method based on forward propagation

Abstract

The invention relates to a photonic neural network training method based on forward propagation. A light source is arranged at the input end of the photonic neural network, an optical detector is arranged at the output end of the photonic neural network, parallel updating of all parameters of the whole network can be achieved only through one-time forward propagation, and the method is superior toa back propagation-layer-by-layer network updating method commonly used for training the neural network in an existing electronic chip. Parameter gradient calculation can be realized only by introducing the light source at the input end and configuring the optical detector at the output end; training and reasoning are also completed through forward propagation, and training and reasoning can be synchronously carried out based on the photonic neural network chip provided by the invention; and according to the method, extra clues can be given, the training of the network parameters is helped tobe prevented from falling into some incorrect local extreme points, and therefore, a better training effect is obtained.

Description

technical field [0001] The invention relates to the technical field of photonic neural network training methods, in particular to a photonic neural network training method based on forward propagation. Background technique [0002] In 2017, researchers at the Massachusetts Institute of Technology proposed a typical photonic neural network chip (see Y.Shen, et al. "Deep learning with coherent nanophotonic circuits," Nat.Photonics 11, 441(2017).), This chip uses the SVD algorithm to decompose any matrix into two unitary matrices and one diagonal matrix, and then uses an array of Mach-Zehnder interferometers (MZI) to simulate these three matrices, thus realizing a full The calculation of optical arbitrary matrix multiplication provides a photonics solution for the calculation acceleration of fully connected neural networks. At the same time, the storage, control, nonlinear calculation and other parts that the photonic circuit is not good at are placed in the external circuit, ...

AI Technical Summary

Problems solved by technology

[0008] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a photon neural network training method based on forward propagation, which can realize the parallel update of all parameters of the entire network through only one forward propagation, which solves the existing problems The problem of increasing system complexity for photonic neural network chip training in technology

Images