Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

An optical neural network processor and a training method thereof

A neural network and neural network model technology, applied in biological neural network models, physical implementation, etc., can solve problems such as low accuracy of results, limited performance of optical neural network processors, slow convergence speed, etc., to achieve good test recognition. rate, faster convergence, and improved accuracy

Active Publication Date: 2019-05-21
INST OF COMPUTING TECH CHINESE ACAD OF SCI
View PDF6 Cites 17 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

However, existing optical devices are difficult to match with traditional computing theory based on Boolean logic.
Directly applying optical neurons to the traditional processor structure, the performance of the obtained optical neural network processor is limited, which makes it easy to have slow convergence speed or inaccurate output results when using such hardware for training and iteration. high case

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • An optical neural network processor and a training method thereof
  • An optical neural network processor and a training method thereof
  • An optical neural network processor and a training method thereof

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0053] According to one embodiment of the present invention, there is provided a neuron structure based on light technology (also referred to as a light neuron herein), see figure 2 As shown, the optical neuron includes a laser module 210 , a synaptic input modulation module 220 , a synaptic weight modulation module 230 , a light aggregation module 240 and a light intensity detection module 250 .

[0054] The laser module 210 includes multiple lasers for generating continuous and stable single-wavelength or multi-wavelength optical signals.

[0055] The synaptic input modulation module 220 includes a plurality of synaptic input modulators, each synapse of the optical neuron corresponds to a synaptic input modulator, and the synaptic input modulation module 220 is used to modulate the optical signal generated by the laser, for example , using intensity modulation to generate the desired light intensity distribution to carry the light signal corresponding to the input neuron va...

Embodiment 2

[0061] image 3 Further shows the optical neuron structure of a preferred embodiment of the present invention, including a laser module 310, a synaptic input modulation module 320, a synaptic weight modulation module 330, a light aggregation module 340 and a light intensity detection module 350, and figure 2 The examples differ in that, image 3 In the embodiment, the neuron structure is realized by two optical signals.

[0062] The synaptic input modulation module 320 consists of two discrete optical modulators, referred to as a positive input modulator and a negative input modulator, respectively. According to another embodiment of the present invention, the synaptic input modulation module 320 is an integrated light modulator. Correspondingly, the synaptic weight modulation module 330 is divided into a positive weight modulator and a negative weight modulator, and the light aggregation module 340 is composed of a positive light aggregator and a negative light aggregator....

Embodiment 3

[0079] In one embodiment, for the variable weight value operation mode, configure less number of neurons, the neural network processing system (combined with image 3 shown) workflow includes the following steps:

[0080] Step S410, before the system runs, the weight matrix of the neural network is decomposed into blocks, and a large matrix is ​​divided into several small block matrices, thereby adapting to the number of neurons in the system;

[0081] Step S420, decomposing each synaptic input value xi in the form of positive and negative numbers to control the positive input modulator and negative input modulator of each synapse respectively;

[0082] Step S430, for each group of neuron operations, input weight information to the synaptic weight modulation module to complete the modulation of synaptic weights;

[0083] Step S440, for each group of neuron operations, input neuron information to the synaptic input modulation module to complete the modulation of synaptic input...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention provides an optical neural network processor and a training method thereof. The processor comprises: a numerical mapping means for realizing a mapping between a numerical value and a numerical value in a positive integer domain which can be represented by an optical neuron; An optical computing device including an optical neuron for performing a corresponding calculation of a networklayer of the neural network model according to an input value and a weight value within a positive integer domain represented by the optical neuron; a photoelectric converter which is used for converting an optical signal of a calculation result of the optical calculation device into an electric signal; And a nonlinear activation device which is used for executing nonlinear activation on the electric signal of the calculation result of the corresponding network layer.

Description

technical field [0001] The invention relates to the training of neural network models, in particular to the training of neural network models using optical neuron devices. Background technique [0002] Since the dawn of the computer, humans have tried to use computing to build a new kind of intelligent machine that can respond in ways similar to human intelligence. After more than half a century of ups and downs, artificial intelligence has become an important branch of computer science. As the mainstream method in the field of artificial intelligence, the deep learning algorithm based on big data has attracted more and more attention. It has played a fundamental role in the fields of natural language processing, drones, and information security, and has created a huge economy for the society. value. [0003] An artificial neural network consists of a large number of neurons. The neuron is a device structure capable of simulating the function of a biological neuron, and t...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/067
Inventor 马恬煜臧大伟刘伯然沈华谭光明张佩珩孙凝晖
Owner INST OF COMPUTING TECH CHINESE ACAD OF SCI
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com