CN-122021759-A - Method and system for calculating morphology of wavelength division multiplexing photon nerve

CN122021759ACN 122021759 ACN122021759 ACN 122021759ACN-122021759-A

Abstract

The invention relates to the technical field of photon calculation, in particular to a method and a system for calculating the morphology of a wavelength division multiplexing photon nerve. The method comprises the steps of obtaining a central data matrix based on an electric signal matrix and average vectors of all preprocessed electric signals, obtaining a covariance matrix based on the total number of the electric signals and the central data matrix, carrying out feature decomposition on the covariance matrix to obtain a group of feature vectors and corresponding feature values thereof, constructing a compression matrix, compressing the electric signal matrix, flattening the compressed electric signal matrix in a time division multiplexing mode, loading the flattened electric signal matrix in a wavelength division multiplexing mode to different wavelengths of a streamline photon reserve pool computing system to realize optical calculation, obtaining an output electric signal through photoelectric conversion, and decompressing the output electric signal to obtain a processing result of input data. The invention effectively improves the processing efficiency and the precision of data by fusing a compression algorithm and a wavelength division multiplexing structure, and enhances the integration of a system by a streamline hidden layer design.

Inventors

CAI DEYU
ZHANG XINGKAI
HUANG YU
LI NIANQIANG

Assignees

苏州大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. A method for calculating the morphology of a wavelength division multiplexed photonic nerve, comprising: Converting the input data into an electric signal, preprocessing the electric signal to obtain a preprocessed electric signal, and combining all the preprocessed electric signals to obtain an electric signal matrix; Averaging each column of the electric signal matrix to obtain average vectors of all the preprocessed electric signals; acquiring a central data matrix based on the electric signal matrix and the average value vectors of all the preprocessed electric signals; Acquiring a covariance matrix based on the total number of the electric signals and the central data matrix; Performing feature decomposition on the covariance matrix to obtain a group of feature vectors and corresponding feature values thereof; According to the magnitude of the characteristic values, descending order sorting is carried out on the characteristic vectors, and the characteristic vectors with the preset quantity before descending order sorting are spliced to obtain a compression matrix; compressing the electric signal matrix based on the compression matrix to obtain a compressed electric signal matrix; flattening the compressed electric signal matrix in a time division multiplexing mode, and sequentially passing through an input layer and a hidden layer of a photon storage pool computing system to obtain a plurality of independent longitudinal-mode high-dimensional nonlinear optical signals; respectively converting the high-dimensional nonlinear optical signals of a plurality of independent longitudinal modes into electric signals through corresponding photoelectric detectors in an output layer of a photon storage pool computing system; the plurality of converted electric signals are collected through an adder of an output layer of the photon reserve pool computing system, and then output electric signals are obtained; And calculating a programmable gate array of an output layer of the system by utilizing the photon storage pool, decompressing the output electric signal based on a transposed matrix of the compression matrix, and acquiring a processing result of the input data based on the decompressed output electric signal.
2. The method of claim 1, wherein the input data is any one of image data, text data, and audio data.
3. The method for calculating the morphology of the wavelength division multiplexing photonic nerve according to claim 1, wherein the formula for obtaining the center data matrix based on the electric signal matrix and the average vector of all the preprocessed electric signals is: , Wherein, the As a matrix of the central data, In the form of a matrix of electrical signals, Representing an all 1 column vector of dimension h x 1, h being the total number of electrical signals, Is the first The electrical signal after the pre-processing is applied, For the mean vector of all pre-processed electrical signals, Indexing the electrical signals.
4. The method of claim 1, wherein the formula for obtaining the covariance matrix based on the total number of electrical signals and the central data matrix is as follows: , Wherein, the For the covariance matrix, h is the total number of electrical signals, As a matrix of the central data, Is that Is a transposed matrix of (a).
5. The method for calculating the morphology of the wavelength division multiplexing photonic nerve according to claim 1, wherein the method for compressing the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix comprises the following steps: multiplying the electric signal matrix with the compression matrix to obtain a compressed electric signal matrix.
6. The method of claim 1, wherein the photon reservoir computing system is a streamlined photon reservoir computing system, and the hidden layer of the streamlined photon reservoir computing system comprises: The 1 port of the optical circulator is connected with the output end of the input layer of the streamline photon storage pool computing system and is used for receiving the composite wave optical signal; The input end of the Fabry-Perot laser is connected with the 2 port of the optical circulator and is used for receiving the combined wave optical signal, converting the combined wave optical signal into high-dimensional nonlinear optical signals with different wavelengths and independent longitudinal modes and transmitting the high-dimensional nonlinear optical signals to the optical circulator; The input end of the optical coupler is connected with the 3 port of the optical circulator and is used for receiving optical signals of independent longitudinal modes with different wavelengths and distributing the optical signals to each parallel amplifying and filtering unit; And the input ends of the amplifying and filtering units are connected with the output ends of the optical couplers, and the amplifying and filtering units are used for receiving the corresponding optical signals of independent longitudinal modes, and each amplifying and filtering unit comprises: the input end of the erbium-doped optical fiber amplifier is connected with the output end of the optical coupler and is used for receiving the optical signals distributed by the optical coupler and amplifying the power; The input end of the optical filter is connected with the output end of the erbium-doped optical fiber amplifier, and the optical filter is used for filtering the optical signals after power amplification, separating out corresponding high-dimensional nonlinear optical signals of independent longitudinal modes and sending the nonlinear optical signals to an output layer of the streamline photon reserve pool computing system.
7. The method for calculating the morphology of the wavelength division multiplexing photonic nerve according to claim 1, wherein the method for preprocessing the electrical signal to obtain the preprocessed electrical signal comprises the following steps: and preprocessing the electric signal based on the mask signal and the mask scaling factor to obtain a preprocessed electric signal.
8. The method for calculating the morphology of the wavelength division multiplexing photonic nerve according to claim 7, wherein the formula for preprocessing the electrical signal based on the mask signal and the mask scaling factor to obtain the preprocessed electrical signal is: , Wherein, the Representing the electrical signal after the pre-processing, Representing the electrical signal(s), The mask signal is represented by a code pattern, Representing the mask scaling factor, t represents the continuous time.
9. The method of claim 7, wherein the masking signal is any one of a binary masking signal, a multi-valued masking signal, and a chaotic masking signal.
10. A wavelength division multiplexed photonic neuromorphic computing system, comprising: the preprocessing module is used for converting input data into electric signals, preprocessing the electric signals to obtain preprocessed electric signals; The average value vector acquisition module is used for solving the average value of each column of the electric signal matrix to obtain average value vectors of all the preprocessed electric signals; the central data matrix acquisition module is used for acquiring a central data matrix based on the electric signal matrix and the average value vectors of all the preprocessed electric signals; The covariance matrix acquisition module is used for acquiring a covariance matrix based on the total number of the electric signals and the central data matrix; the feature decomposition module is used for carrying out feature decomposition on the covariance matrix to obtain a group of feature vectors and corresponding feature values thereof; The compression matrix acquisition module is used for carrying out descending order sequencing on the feature vectors according to the magnitude of the feature values, and splicing the feature vectors with the preset number before the descending order sequencing to obtain a compression matrix; The compression module is used for compressing the electric signal matrix based on the compression matrix to obtain a compressed electric signal matrix; The optical brain-like calculation module is used for flattening the compressed electric signal matrix in a time division multiplexing mode, and sequentially passing through an input layer and a hidden layer of the photon storage pool calculation system to obtain a plurality of independent longitudinal-mode high-dimensional nonlinear optical signals; The photoelectric conversion module is used for respectively converting the high-dimensional nonlinear optical signals of the plurality of independent longitudinal modes into electric signals through corresponding photoelectric detectors in the output layer of the photon storage pool computing system; the summarizing module is used for summarizing the plurality of converted electric signals through an adder of an output layer of the photon reserve pool computing system to obtain output electric signals; The output module is used for calculating a programmable gate array of an output layer of the system by utilizing the photon reserve pool, decompressing the output electric signal based on the transposed matrix of the compression matrix, and acquiring a processing result of the input data based on the decompressed output electric signal.

Description

Method and system for calculating morphology of wavelength division multiplexing photon nerve Technical Field The invention relates to the technical field of data processing, in particular to a method and a system for calculating the morphology of a wavelength division multiplexing photon nerve. Background With the rapid evolution of artificial intelligence technology to large-scale and high-precision directions, the fields of natural language processing, computer vision and the like mainly depend on building a large-scale complex computing architecture, so that more requirements are put on the bandwidth, speed and energy efficiency of a computing system. However, due to the adoption of a design manner of memory separation, the traditional computing architecture based on von neumann system can generate significant delay and high energy consumption when performing data processing computing tasks such as images, texts, audio and the like, which makes the traditional computing system face serious challenges in terms of energy efficiency and processing speed, and is difficult to meet the real-time processing requirements of large-scale image, text and audio data. In order to break through the bottlenecks of high delay, high energy consumption and the like caused by the traditional von neumann system deposit and separation, the photon nerve morphology calculation is generated. The photonic nerve morphology calculation provides an innovation direction for constructing a new generation of calculation architecture by simulating a high-efficiency processing mechanism of the human brain and combining unique advantages of a photonic device in aspects of multidimensional physical fields, large bandwidth and the like. In recent years, the technology has made important progress in the fields of intelligent driving, medical diagnosis, logical reasoning and the like. In the intelligent driving field, the intelligent driving system can rapidly process massive image data such as road condition images, obstacle images and the like acquired by the vehicle-mounted cameras to realize real-time environment perception and decision, in the medical diagnosis field, can efficiently analyze image data such as medical images and the like to assist doctors in improving diagnosis precision and speed, in the voice interaction field, can rapidly process voice audio data to complete tasks such as voice recognition and semantic conversion, in the large-scale artificial intelligent large model reasoning field, can relieve the computational bottleneck of a traditional computing architecture and provides support for efficient operation of large models which rely on image, text and audio multi-type data input. In the processing process of the image, text and audio data, various data are firstly converted into electric signals and then modulated onto optical signals, the optical signals are used as carriers for data transmission and calculation, and after high-efficiency calculation is completed through a photon device, the optical signals are converted back into electric signals, and finally processing results of various data are output. On the basis, the wavelength division multiplexing photonic nerve morphology calculation combines the wavelength division multiplexing technology with the photonic nerve morphology calculation, and utilizes the multi-wavelength channel to realize parallel transmission and processing of optical signals corresponding to various data such as images, texts, audios and the like, so that the communication bandwidth and the calculation throughput of the system are further improved, and huge application potential is shown in scenes such as time sequence signal processing, image recognition, intelligent perception and the like. The wavelength division multiplexing photon reserve pool calculation is taken as a typical representative of wavelength division multiplexing photon nerve morphology calculation, shows excellent physical compatibility, and can be adapted to various hardware platforms such as electrons, photons, photoelectrons and the like. The calculation model adopts fixed random weights at the input layer and the hidden layer, and only the output layer is trained through a linear regression algorithm, so that the simplified training mechanism ensures good engineering realizability while maintaining the nonlinear characterization capability of the system. However, as the artificial intelligence technology continuously evolves to a large-scale and high-precision direction, the scale of input images, texts and audio data in the fields of computer vision, natural language processing, voice recognition and the like is exponentially expanded, and various data contain a large amount of redundant information, at the moment, the existing photonic neuromorphic computing system still directly loads the original images, texts and audio data to different dimensions of optical signals in a time division multiplexing mode, and