A wavelength division multiplexing photonic neuromorphic computing method and system

Abstract

The present application relates to the technical field of photon computing, and especially relates to a wavelength division multiplexing photon neuromorphic computing method and system. A center data matrix is obtained based on an electric signal matrix and mean vectors of all preprocessed electric signals; a covariance matrix is obtained based on the total number of electric signals and the center data matrix; the covariance matrix is subjected to eigenvalue decomposition to obtain a group of eigenvectors and corresponding eigenvalues, a compression matrix is constructed, and the electric signal matrix is compressed; the compressed electric signal matrix is flattened through time division multiplexing and loaded into different wavelengths of a streamlined photon reserve pool computing system through wavelength division multiplexing to realize optical computing, and output electric signals are obtained through photoelectric conversion; the output electric signals are decompressed to obtain the processing result of input data. The present application effectively improves the processing efficiency and accuracy of data by fusing compression algorithms and wavelength division multiplexing structures, and enhances the integrability of the system through the streamlined hidden layer design.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a wavelength division multiplexing photonic neuromorphic computing method and system. Background Technology

[0002] As artificial intelligence technology rapidly evolves towards large-scale and high-precision applications, fields such as natural language processing and computer vision rely heavily on building large-scale, complex computing architectures. This places higher demands on the bandwidth, speed, and energy efficiency of computing systems. However, traditional computing architectures based on the von Neumann architecture, due to their in-memory / compute separation design, exhibit significant latency and high energy consumption when performing data processing tasks such as image, text, and audio processing. This poses a severe challenge to traditional computing systems in terms of energy efficiency and processing speed, making it difficult to meet the real-time processing needs of large-scale image, text, and audio data. To overcome the bottlenecks of high latency and high energy consumption caused by the in-memory / compute separation of traditional von Neumann architectures, photonic neuromorphic computing has emerged.

[0003] Photonic neuromorphic computing, by simulating the efficient processing mechanisms of the human brain and combining the unique advantages of photonic devices in multidimensional physical fields and high bandwidth, provides an innovative direction for building next-generation computing architectures. In recent years, this technology has made significant progress in fields such as intelligent driving, medical diagnosis, and logical reasoning. In intelligent driving, it can rapidly process massive amounts of image data, including road condition and obstacle images, collected by in-vehicle cameras, enabling real-time environmental perception and decision-making. In medical diagnosis, it can efficiently analyze medical images, assisting doctors in improving diagnostic accuracy and speed. In voice interaction, it can rapidly process human voice audio data, completing tasks such as speech recognition and semantic conversion. In large-scale artificial intelligence model inference, it can alleviate the computational bottleneck of traditional computing architectures, providing support for the efficient operation of large models that rely on multiple types of data input, including images, text, and audio. In the processing of the aforementioned image, text, and audio data, various types of data are first converted into electrical signals, then modulated onto optical signals. The optical signals serve as the carrier for data transmission and computation. After efficient computation by photonic devices, they are converted back into electrical signals, ultimately outputting the processing results of various data types.

[0004] Building upon this foundation, wavelength division multiplexing (WDM) photonic neuromorphic computing combines WDM technology with photonic neuromorphic computing, utilizing multi-wavelength channels to achieve parallel transmission and processing of optical signals corresponding to various data types such as images, text, and audio. This further enhances the system's communication bandwidth and computational throughput, demonstrating significant application potential in scenarios such as time-series signal processing, image recognition, and intelligent sensing. WDM photonic reservoir computing, as a typical representative of WDM photonic neuromorphic computing, exhibits excellent physical compatibility, adapting to various hardware platforms including electronics, photonics, and optoelectronics. This computational model employs fixed random weights in the input and hidden layers, requiring only linear regression algorithm training for the output layer. This simplified training mechanism maintains the system's nonlinear representation capabilities while ensuring good engineering feasibility.

[0005] However, as artificial intelligence technology continues to evolve towards large-scale and high-precision processing, the scale of input image, text, and audio data in fields such as computer vision, natural language processing, and speech recognition is expanding exponentially. Furthermore, these data types inherently contain a large amount of redundant information. Currently, existing photonic neuromorphic computing systems still directly load the original image, text, and audio data onto different dimensions of the optical signal using time-division multiplexing, directly entering the hidden layer for high-dimensional nonlinear mapping. This processing method significantly increases the time overhead in optical computing and the processing pressure on the photoelectric detection module, ultimately resulting in poor information processing efficiency for photonic neuromorphic computing systems when handling large-scale image, text, and audio data, making it difficult to meet the demands for efficient, real-time processing of large-scale, multi-type data. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of existing photonic neuromorphic computing systems in terms of poor information processing efficiency when processing large-scale image, text and audio data, and inability to adapt to the needs of efficient and real-time processing of large-scale multi-type data.

[0007] To address the aforementioned technical problems, this invention provides a wavelength division multiplexing photonic neuromorphic computing method, comprising:

[0008] The input data is converted into electrical signals, the electrical signals are preprocessed to obtain preprocessed electrical signals, and all preprocessed electrical signals are merged to obtain an electrical signal matrix.

[0009] The mean vector of all preprocessed electrical signals is obtained by averaging each column of the electrical signal matrix.

[0010] The central data matrix is ​​obtained based on the electrical signal matrix and the mean vector of all preprocessed electrical signals.

[0011] Based on the total number of electrical signals and the central data matrix, the covariance matrix is ​​obtained.

[0012] The covariance matrix is ​​decomposed into eigenvalues ​​to obtain a set of eigenvectors and their corresponding eigenvalues.

[0013] Based on the magnitude of the eigenvalues, the eigenvectors are sorted in descending order. The first predetermined number of eigenvectors after descending sorting are then combined to obtain a compressed matrix.

[0014] Based on the compression matrix, the electrical signal matrix is ​​compressed to obtain the compressed electrical signal matrix;

[0015] The compressed electrical signal matrix is ​​flattened by time-division multiplexing and then passed through the input layer and hidden layer of the photonic reservoir computing system to obtain high-dimensional nonlinear optical signals with multiple independent longitudinal modes.

[0016] Multiple independent longitudinal mode high-dimensional nonlinear optical signals are converted into electrical signals by passing them through the corresponding photodetectors in the output layer of the photonic reservoir computing system.

[0017] The multiple converted electrical signals are combined by the adder in the output layer of the photon reservoir computing system to obtain the output electrical signal;

[0018] Using a programmable gated array at the output layer of a photonic reservoir computing system, the output electrical signal is decompressed based on the transpose of the compression matrix; the processing result of the input data is obtained based on the decompressed output electrical signal.

[0019] Preferably, the input data is any one of image data, text data, and audio data.

[0020] Preferably, the formula for obtaining the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals is as follows:

[0021] ,

[0022] in, As the central data matrix, For electrical signal matrix, This represents a column vector of all ones with dimension h×1, where h is the total number of electrical signals. For the first A preprocessed electrical signal, This is the mean vector of all preprocessed electrical signals. For electrical signal index.

[0023] Preferably, the formula for obtaining the covariance matrix based on the total number of electrical signals and the central data matrix is ​​as follows:

[0024] ,

[0025] in, Let h be the covariance matrix, and h be the total number of electrical signals. As the central data matrix, for The transpose of .

[0026] Preferably, the method for compressing the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix includes:

[0027] Multiplying the electrical signal matrix by the compression matrix yields the compressed electrical signal matrix.

[0028] Preferably, the photon reservoir computing system is a streamlined photon reservoir computing system, and the hidden layer of the streamlined photon reservoir computing system includes:

[0029] An optical circulator, whose port 1 is connected to the output of the input layer of a streamlined photon reservoir computing system, is used to receive combined optical signals;

[0030] The Fabry-Perot laser has its input end connected to port 2 of the optical circulator to receive the combined optical signal, convert the combined optical signal into a high-dimensional nonlinear optical signal with independent longitudinal modes of different wavelengths, and send it to the optical circulator.

[0031] An optical coupler, whose input is connected to port 3 of an optical circulator, is used to receive optical signals of different wavelengths and independent longitudinal modes and distribute them to various parallel amplification and filtering units.

[0032] Multiple parallel amplification and filtering units, whose inputs are connected to the output of an optocoupler, are used to receive corresponding independent longitudinal mode optical signals. Each amplification and filtering unit includes:

[0033] An erbium-doped fiber amplifier, whose input is connected to the output of an optical coupler, is used to receive the optical signal distributed by the optical coupler and amplify its power.

[0034] An optical filter, whose input is connected to the output of an erbium-doped fiber amplifier, is used to filter the amplified optical signal, separate the corresponding independent longitudinal mode high-dimensional nonlinear optical signal, and send it to the output layer of the streamlined photon reservoir computing system.

[0035] Preferably, the method for preprocessing the electrical signal to obtain the preprocessed electrical signal includes: preprocessing the electrical signal based on the mask signal and the mask scaling factor to obtain the preprocessed electrical signal.

[0036] Preferably, the formula for preprocessing the electrical signal based on the mask signal and the mask scaling factor to obtain the preprocessed electrical signal is:

[0037] ,

[0038] in, This represents the preprocessed electrical signal. Indicates an electrical signal. Indicates the mask signal. represents the mask scaling factor, and t represents continuous time.

[0039] Preferably, the mask signal is any one of a binary mask signal, a multi-valued mask signal, and a chaotic mask signal.

[0040] This invention also provides a wavelength division multiplexing photonic neuromorphic computing system, comprising:

[0041] The preprocessing module is used to convert input data into electrical signals, preprocess the electrical signals to obtain preprocessed electrical signals, and merge all preprocessed electrical signals to obtain an electrical signal matrix.

[0042] The mean vector acquisition module is used to calculate the mean of each column of the electrical signal matrix to obtain the mean vector of all preprocessed electrical signals.

[0043] The central data matrix acquisition module is used to acquire the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals.

[0044] The covariance matrix acquisition module is used to obtain the covariance matrix based on the total number of electrical signals and the central data matrix.

[0045] The eigenvalue decomposition module is used to perform eigenvalue decomposition on the covariance matrix to obtain a set of eigenvectors and their corresponding eigenvalues.

[0046] The compression matrix acquisition module is used to sort the feature vectors in descending order according to the size of the feature values, and then combine the first preset number of feature vectors after descending sorting to obtain the compression matrix.

[0047] The compression module is used to compress the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix;

[0048] The optical neuromorphic computing module is used to flatten the compressed electrical signal matrix through time-division multiplexing, and then pass it through the input layer and hidden layer of the photon reservoir computing system to obtain high-dimensional nonlinear optical signals with multiple independent longitudinal modes.

[0049] The photoelectric conversion module is used to convert high-dimensional nonlinear optical signals with multiple independent longitudinal modes into electrical signals through the corresponding photodetectors in the output layer of the photonic reservoir computing system.

[0050] The aggregation module is used to aggregate multiple converted electrical signals through the adder of the output layer of the photon reservoir computing system to obtain the output electrical signal;

[0051] The output module is used to decompress the output electrical signal based on the transpose of the compression matrix using the programmable gated array of the output layer of the photon reservoir computing system; and to obtain the processing result of the input data based on the decompressed output electrical signal.

[0052] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0053] This invention discloses a wavelength division multiplexing (WDM) photonic neuromorphic computing method and system. It designs an adaptive compression-decompression algorithm. The mean vector of all preprocessed electrical signals is essentially the average level of each dimension of all input electrical signals. Subtracting the mean vector of all preprocessed electrical signals from the electrical signal matrix yields a central data matrix, eliminating interference caused by overall data offset. The covariance matrix calculated based on the centralized data matrix fully captures the statistical distribution characteristics of multi-wavelength electrical signals. The variance reflects the dynamic intensity differences of information in each dimension, while the covariance value clearly characterizes the degree of linear correlation between channels. This lays a data foundation for subsequent identification of signal feature space, quantitative evaluation of channel redundancy, and the implementation of adaptive data compression. By performing eigenvalue decomposition on the covariance matrix, eigenvectors carrying core effective information are selected. These selected eigenvectors are then combined to form a compression matrix, which can accurately filter out effective information from multiple electrical signals and eliminate low-weight electrical signal features corresponding to redundant information, thereby achieving data simplification and compression. This approach maximizes the retention of core effective information in the input data, avoiding information loss during compression that could affect the accuracy of the processing results. Furthermore, it enables redundancy removal in the electrical domain beforehand, significantly reducing data size and preventing the direct loading of raw redundant data into the photonic wavelength channel. This reduces the significant time overhead and photodetector processing pressure caused by time-division multiplexing from the source. Decompression is then performed based on the transpose of the compression matrix to reconstruct the information, ensuring the integrity and reliability of the final output. This effectively improves the efficiency and accuracy of wavelength division multiplexing photonic neuromorphic computing systems when processing large-scale image, text, and audio data.

[0054] Furthermore, this invention proposes a streamlined photonic reservoir computing system without feedback loops. By deeply coupling the wavelength dimension with nonlinear nodes, it avoids the physical dependence of traditional reservoir computing architectures on long-distance fiber delay feedback loops. Utilizing mode competition and gain saturation effects of multi-wavelength carriers within a single nonlinear device to generate a high-dimensional state space with rich dynamic characteristics, this system architecture eliminates the difficulty of integrating fiber delay feedback loops. While significantly improving system integrability and physical robustness, it achieves efficient, real-time processing of large-scale image, text, and audio data streams without compromising computational accuracy. Attached Figure Description

[0055] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0056] Figure 1 This is a schematic flowchart of a wavelength division multiplexing photonic neuromorphic computing method according to the present invention.

[0057] Figure 2 This is a structural diagram of a streamlined photon reservoir computing system.

[0058] Figure 3 This is the output spectrum of a Fabry-Perot laser.

[0059] Figure 4 This invention presents the image recognition results of processing handwritten digit recognition datasets and fashion item datasets. Figure 4 (a) in the figure is the image recognition result of the handwritten digit recognition dataset processed by the present invention. Figure 4 (b) in the figure is the image recognition result of the fashion item dataset processed by the present invention.

[0060] Figure 5 This invention presents a streamlined photon reservoir computing system that compares the image recognition results of a handwritten digit recognition dataset with those of a traditional photon neuromorphic computing system with a delayed feedback loop. Figure 5 (a) in the figure is the image recognition result of the handwritten digit recognition dataset processed by the present invention. Figure 5 (b) in the figure shows the image recognition results of a traditional photonic neuromorphic computing system with a delayed feedback loop processing a handwritten digit recognition dataset.

[0061] Explanation of reference numerals in the accompanying drawings: 100, Input layer; 200, Hidden layer; 300, Output layer; 11, Polarization controller; 12, Arbitrary waveform generator; 13, Modulator; 14, Wavelength division multiplexer; 21, Optical circulator; 22, Fabry-Perot laser; 23, Optical coupler; 24, Erbium-doped fiber amplifier; 25, Optical filter; 31, Photodetector; 32, Adder; 33, Programmable gate array. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0063] Reference Figure 1 As shown, this embodiment provides a wavelength division multiplexing photonic neuromorphic computing method, including:

[0064] Step S1: Convert the input data into an electrical signal, preprocess the electrical signal to obtain a preprocessed electrical signal; merge all the preprocessed electrical signals to obtain an electrical signal matrix. , ;in, Let h be the h-th preprocessed electrical signal; each row in the electrical signal matrix corresponds to one preprocessed electrical signal, and h is the total number of electrical signals;

[0065] In this embodiment, the input data is specifically any one of image data, text data, and audio data.

[0066] For example: when the input data is When there are 10 images, then each image will be... The image is converted into its corresponding electrical signal, denoted as . , ,…, , Indicates continuous time. , The total number of images, For image indexing, Show the first The electrical signals corresponding to each image are processed; all preprocessed electrical signals are merged to obtain an electrical signal matrix; the number of elements in each row of the electrical signal matrix is ​​equal to the total number of pixels in a single image.

[0067] In this embodiment, specifically, the method for preprocessing the electrical signal to obtain the preprocessed electrical signal includes:

[0068] Based on the mask signal and the mask scaling factor, the electrical signal is preprocessed to obtain the preprocessed electrical signal, as shown in the formula:

[0069] ,

[0070] in, This represents the preprocessed electrical signal. Indicates an electrical signal. Indicates the mask signal. represents the mask scaling factor, and t represents continuous time.

[0071] In this embodiment, specifically, the mask signal is any one of a binary mask signal, a multi-value mask signal, and a chaotic mask signal.

[0072] Step S2: Calculate the mean of each column of the electrical signal matrix to obtain the mean vector of all preprocessed electrical signals; where the mean vector of all preprocessed electrical signals is obtained by averaging the signal values ​​of all preprocessed electrical signals at the same time.

[0073] Step S3: The method for obtaining the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals includes:

[0074] The central data matrix is ​​obtained by subtracting the mean vector of all preprocessed electrical signals from the electrical signal matrix, as shown in the formula:

[0075] ,

[0076] in, As the central data matrix, For electrical signal matrix, This represents a column vector of all ones with dimension h×1, where h is the total number of electrical signals. For the first A preprocessed electrical signal, This is the mean vector of all preprocessed electrical signals. For electrical signal index.

[0077] Step S4: Based on the total number of electrical signals and the central data matrix, obtain the covariance matrix. The formula is:

[0078] ,

[0079] in, Let h be the covariance matrix, and h be the total number of electrical signals. As the central data matrix, for The transpose of .

[0080] Step S5: The method for performing eigenvalue decomposition on the covariance matrix to obtain a set of eigenvectors and their corresponding eigenvalues ​​includes:

[0081] Perform eigenvalue decomposition on the covariance matrix, i.e., solve the equation ,in, Represents the eigenvector matrix, It is a diagonal matrix, and the eigenvalues ​​on its diagonal represent the contribution of the variance in the direction of the corresponding eigenvector.

[0082] Step S6: The feature vectors form a compression matrix according to the variance contribution of the feature values: the feature vectors are sorted in descending order according to the size of the feature values, and the first preset number of feature vectors after descending sorting are combined. By retaining the feature vectors with larger feature values, a compression matrix is ​​obtained; the feature vectors show hierarchical and ordered contributions, and the size of their contributions is quantified by the feature values.

[0083] Step S7: Based on the compression matrix, compress the electrical signal matrix to obtain the compressed electrical signal matrix. The calculation formula is as follows:

[0084] ,

[0085] in, The compressed electrical signal matrix, For electrical signal matrix, This is a compression matrix.

[0086] Compression matrix The feature vectors are constructed by selecting those sorted by variance contribution. This greatly reduces the amount of input information through principled information compression while preserving task-related information, thereby enhancing the computational speed of wavelength division multiplexing photonic neuromorphic computing.

[0087] In this embodiment, specifically, the method for compressing the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix includes:

[0088] Multiplying the electrical signal matrix by the compression matrix yields the compressed electrical signal matrix.

[0089] Step S8: Flatten the compressed electrical signal matrix by time-division multiplexing, and pass it sequentially through the input layer 100 and hidden layer 200 of the photonic reservoir computing system to obtain high-dimensional nonlinear optical signals with multiple independent longitudinal modes;

[0090] Step S9: Convert the high-dimensional nonlinear optical signals of multiple independent longitudinal modes into electrical signals by passing them through the corresponding photodetectors 31 in the output layer 300 of the photonic reservoir computing system.

[0091] Step S10: After the multiple converted electrical signals are combined by the adder 32 of the output layer 300 of the photon reservoir computing system, the output electrical signal is obtained.

[0092] Step S11: Using the programmable gate array 33 of the output layer 300 of the photon reservoir computing system, the output electrical signal is decompressed based on the transpose of the compression matrix; based on the decompressed output electrical signal, the processing result of the input data is obtained.

[0093] like Figure 2 As shown, Figure 2 This is a structural diagram of a streamlined photon reservoir computing system.

[0094] The streamlined photon reservoir computing system includes an input layer 100, a hidden layer 200, and an output layer 300.

[0095] In this embodiment, the input layer of the photon reservoir computing system adopts wavelength division multiplexing (WDM) technology, which includes: an arbitrary waveform generator (AWG) 12, a WDM multiplexer 14, and multiple parallel units; each parallel unit includes a driving laser, a polarization controller 11, and a modulator 13 connected in sequence; the output terminal of the arbitrary waveform generator 12 is connected to the input terminal of the modulator 13 of each parallel unit, and the output terminal of the modulator 13 of each parallel unit is connected to the WDM multiplexer 14;

[0096] The compressed electrical signal matrix obtained after processing by the adaptive compression-decompression algorithm is used as the input signal of the input layer 100 of the entire photon reservoir computing system and is input to the input terminal of the arbitrary waveform generator 12.

[0097] The arbitrary waveform generator 12 performs signal adaptation on the compressed electrical signal matrix, converting each compressed electrical signal in the compressed electrical signal matrix into an adjustable driving electrical signal that adapts to the driving requirements of each parallel unit modulator 13, ensuring that the waveform, amplitude, and timing of the driving signal are accurately matched with the characteristics of the compressed electrical signal, thus providing an adapted electrical driving basis for subsequent optical modulation.

[0098] During signal processing in each parallel unit, the driving laser within each parallel unit operates independently, providing a continuous optical carrier beam of a fixed wavelength to that unit. Furthermore, the optical carrier wavelengths output by the driving lasers of different parallel units are different. By setting optical carriers of different wavelengths, parallel transmission of multiple signals can be achieved without crosstalk. Subsequently, the polarization controller 11 within the same parallel unit performs polarization state calibration on the continuous optical carrier output by the driving laser, adjusting the polarization direction of the optical signal to ensure that the optical carrier can adapt to the modulation requirements of the subsequent modulator 13, improving the accuracy and stability of optical modulation, and avoiding the impact of polarization state shift on signal transmission quality.

[0099] The optical carrier, calibrated by the polarization controller 11, is transmitted to the input of the modulator 13 in the same parallel unit. Simultaneously, the adjustable drive signal output by the arbitrary waveform generator 12, corresponding to the parallel unit, is also transmitted to the input of the modulator 13. Under the control of the drive signal, the modulator 13 loads the information carried by the compressed electrical signal onto the calibrated optical carrier, completing the conversion from electrical signal to optical signal, and obtaining a modulated optical signal carrying compressed information. The wavelength of this modulated optical signal is consistent with the wavelength of the optical carrier output by the laser driven by the corresponding parallel unit.

[0100] After the modulators 13 of each parallel unit complete the electro-optic modulation process described above, the modulated optical signals of different wavelengths output by all parallel units are transmitted to the input of the wavelength division multiplexer 14. As the core device of wavelength division multiplexing technology, the wavelength division multiplexer 14 performs multiplexing processing on multiple modulated optical signals of different wavelengths, integrating multiple modulated optical signals of different wavelengths into a single multiplexed optical signal. Through multiplexing operation, parallel transmission of multiple compressed signals is realized, effectively saving transmission channel resources, significantly improving the signal transmission efficiency of the input layer, and avoiding crosstalk between multiple optical signals.

[0101] Finally, the combined optical signal output by wavelength division multiplexer 14 is injected into the hidden layer 200 of the photon reservoir computing system as the final output signal of input layer 100.

[0102] In this embodiment, modulator 13 is a phase modulator or an intensity modulator.

[0103] In this embodiment, preferably, the modulator 13 is a Mach-Zehnder modulator. Mach-Zehnder modulators are characterized by high modulation accuracy and fast response speed, enabling them to accurately load the compressed information carried by the driving electrical signal onto the optical carrier, effectively reducing signal distortion during modulation and ensuring the integrity of the compressed electrical signal information.

[0104] In this embodiment, specifically, under the control of the driving electrical signal, the modulator 13 loads the information carried by the compressed electrical signal onto the calibrated optical carrier, and the output modulated optical signal formula is:

[0105] ,

[0106] in, To modulate the optical signal, To drive the steady-state output light intensity of the laser, The compressed electrical signal matrix, It is the imaginary unit.

[0107] Existing technologies mostly employ time-division multiplexing combined with delayed feedback, whose processing efficiency is limited by sample size and processing cycle, and introduces additional timing overhead. Although parallel computing can partially alleviate this problem, existing solutions still face inherent limitations in scalability such as crosstalk and power consumption. Therefore, developing novel photonic computing architectures that combine high-speed processing capabilities with good scalability and integrability remains a key technical challenge that needs to be addressed in this field.

[0108] To this end, the present invention introduces a Fabry-Perot laser 22. The multiple longitudinal modes of the Fabry-Perot laser 22 serve as a physical carrier of natural parallelism, and its physical characteristics are directly utilized to realize the concurrent processing of multi-channel information, thus ensuring the high parallelism of the system from a physical perspective.

[0109] like Figure 3 As shown, Figure 3 This is the output spectrum of a Fabry-Perot laser.

[0110] In this embodiment, preferably, the hidden layer 200 of the photon reservoir computing system includes: an optical circulator 21, a Fabry-Perot laser 22, an optical coupler 23, and multiple parallel amplification and filtering units. Each amplification and filtering unit includes an erbium-doped fiber amplifier 24 and an optical filter 25 connected in sequence. The Fabry-Perot laser 22 has m longitudinal modes, and the m channels formed by the m longitudinal modes are used to realize parallel computing. The output light of the Fabry-Perot laser is injected into the amplification and filtering units through the optical circulator 21 to generate multiple independent longitudinal modes of different wavelengths. Here, m represents the number of longitudinal modes of the Fabry-Perot laser 22; the Fabry-Perot laser 22 has multiple longitudinal modes of different wavelengths, which constitute the basis for multi-channel parallel processing, not only for loading algorithmically compressed information, but also significantly enhancing the system's computing speed and scalability.

[0111] The output wavelength of the driving laser in the input layer 100 needs to be configured according to the longitudinal mode wavelength of the Fabry-Perot laser 22. The driving laser is used to generate optical carriers, and the output wavelength of the optical carriers corresponds one-to-one with the longitudinal mode wavelength of the Fabry-Perot laser 22. By achieving precise mode matching and resonant excitation, the computing performance of the reservoir is optimized and improved.

[0112] The input layer 100 also includes a variable optical attenuator, which is used to adjust the optical power of the output light of the driving laser;

[0113] The input signal of the hidden layer 200 is the combined optical signal output by the wavelength division multiplexer 14 in the input layer 100. The combined optical signal is injected into the Fabry-Perot laser 22 through the optical circulator 21 to excite its m independent longitudinal modes. The m longitudinal modes correspond to m parallel channels. Each channel corresponds to one optical signal carrying compressed information, realizing the natural parallel carrying of multi-channel information without the need to build an additional parallel structure.

[0114] The optical signal containing multiple longitudinal modes (containing m independent longitudinal modes of different wavelengths, each carrying one input information) from the Fabry-Perot laser 22 is transmitted to the optical coupler 23 via the optical circulator 21. The optical coupler 23 distributes the optical signal containing multiple longitudinal modes to each parallel amplification and filtering unit. The erbium-doped fiber amplifier 24 in each hidden layer unit amplifies the optical signal output from the optical coupler to compensate for transmission loss. Subsequently, the optical filter 25 filters out clutter and separates the high-dimensional nonlinear optical signal of the independent longitudinal modes. The high-dimensional nonlinear optical signal of the independent longitudinal modes output from each hidden layer unit is transmitted to the output layer 300.

[0115] The output layer 300 of the photon reservoir computing system includes: multiple photodetectors 31, adders 32 and programmable gate array 33; the number of photodetectors 31 is consistent with the number of parallel amplification and filtering units in the hidden layer, and they are used to convert the final output signal of each amplification and filtering unit into an electrical signal.

[0116] The input terminal of adder 32 is connected to the output terminal of each photodetector 31, and is used to receive the electrical signals output by each photodetector 31, perform summation and weighting, and obtain the output electrical signal.

[0117] The input of the programmable gate array 33 is connected to the output of the adder 32 to decompress the output electrical signal and ultimately generate the output weights of the hidden layer 200. The programmable gate array 33 is configured to use either ridge regression or linear regression algorithms. This invention decompresses the high-dimensional mapping information output from the hidden layer 200 in the output layer 300 to complete information reconstruction. Compared to traditional systems, this invention achieves efficient processing of large-scale complex tasks with an extremely simplified network structure, possessing high information processing speed, ease of integration, and excellent scalability.

[0118] The adaptive compression-decompression algorithm extracts the temporal features of the input information after mask preprocessing through feature vector analysis, and realizes information compression and rate enhancement in a wavelength division multiplexing photonic neuromorphic computing system based on the adaptive compression-decompression algorithm.

[0119] The streamlined photonic reservoir computing system replaces the reliance on long-distance fiber delay feedback loops in traditional reservoir architectures. It not only eliminates the processing latency and physical size limitations caused by feedback loops, but also greatly enhances the on-chip integrability of the system, providing high-density, low-power hardware support for the realization of an all-optical chip-level neuromorphic computing platform.

[0120] In this embodiment, specifically, the method for decompressing the output electrical signal based on the transpose of the compression matrix to obtain the processing result of the input data includes:

[0121] Multiplying the output electrical signal by the transpose of the compression matrix yields the decompressed output matrix, as shown in the formula:

[0122] ,

[0123] in, This is the output matrix after decompression. To output an electrical signal, This is the transpose of the compressed matrix.

[0124] To verify that the photonic neuromorphic computing system based on the adaptive compression-decompression algorithm and Fabry-Perot laser can improve the data processing rate, its rate equation was calculated:

[0125] ,

[0126] ,

[0127] ,

[0128] in, and M and m represent the amplitude and frequency detuning of the slowly varying complex electric field of the m-th longitudinal mode of the Fabry-Perot laser 22, respectively. c These represent the total number of longitudinal modes and the central mode of the Fabry-Perot laser 22, respectively. Indicates carrier concentration. Indicates the linear enhancement factor. The transparent carrier concentration is represented by ε, and the saturation coefficient is represented by ε. and These represent the lifetimes of photons and charge carriers, respectively. Indicates the gain coefficient. This represents the light injection intensity of the m-th longitudinal mode of the Fabry-Perot laser 22. This refers to the longitudinal die spacing. β is the gain bandwidth of the laser material, and β is the spontaneous emission coefficient. It is independent Gaussian white noise with unit variance and zero mean. Intensity modulation information of the m-th longitudinal mode injected into the input layer 100 of the Fabry-Perot laser. The imaginary unit, The value represents the injected current density, and m is the vertical pattern index.

[0129] This invention uses an image recognition task to test a photonic neuromorphic computing system based on an adaptive compression-decompression algorithm. For example... Figure 4 As shown, Figure 4 This invention presents the image recognition results of processing handwritten digit recognition datasets and fashion item datasets.Figure 4 (a) in the figure is the image recognition result of the handwritten digit recognition dataset processed by the present invention. Figure 4 (b) in the figure is the image recognition result of the fashion item dataset processed by the present invention.

[0130] from Figure 4 The proposed wavelength division multiplexing photonic neuromorphic computing system, based on an adaptive compression-decompression algorithm, achieves 96.55% and 84.6% accuracy on handwritten digit recognition and fashion item datasets, respectively, with 10 virtual nodes and 8 longitudinal modes using 8 Fabry-Perot lasers. This demonstrates that hidden layer 200 possesses excellent recognition and classification capabilities. Compared to traditional photonic reservoir computing, achieving the same performance requires setting the number of nodes to 1000. The proposed system achieves a 100-fold improvement in information processing speed while maintaining high recognition accuracy. Furthermore, the system described in this invention has a simple structure, is easy to train, and possesses good integration and scalability. It realizes a simple, low-cost, easily integrated, and high-speed photonic neuromorphic computing system.

[0131] like Figure 5 As shown, Figure 5 The streamlined photon reservoir computing system designed in this invention processes image recognition results of handwritten digit recognition datasets in contrast to traditional photon neuromorphic computing systems with delayed feedback loops. Figure 5 (a) in the figure is the image recognition result of the handwritten digit recognition dataset processed by the present invention. Figure 5 (b) in the figure shows the image recognition results of a traditional photonic neuromorphic computing system with a delayed feedback loop processing a handwritten digit recognition dataset.

[0132] from Figure 5 As shown in (a) of this invention, the streamlined photon reservoir computing system can achieve 96.55% performance on the MNIST handwritten digit recognition dataset with 10 virtual nodes, while... Figure 5 (b) shows that a traditional photonic neuromorphic computing system with a delayed feedback loop can achieve a recognition accuracy of 97.05% with 1000 nodes. The system described in this invention can eliminate the large-size delayed feedback loop without significantly reducing the computing performance of the reservoir, greatly enhancing the system's integrability and physical stability.

[0133] This invention overcomes the bottlenecks in processing speed and parallelism of existing photonic neuromorphic computing by combining an adaptive compression-decompression algorithm with a streamlined photonic reservoir computing system. Specifically, this invention reveals that multiple longitudinal modes of a Fabry-Perot laser 22 can serve as naturally parallel physical carriers; by synchronously injecting compressed input signals into these longitudinal modes, concurrent processing of multi-channel information is achieved directly using their physical properties. This method ensures high parallelism of the system at the physical level, thereby achieving excellent processing speed and overall performance.

[0134] This invention addresses the flawed premise that traditional photonic neuromorphic computing suffers from additional time overhead due to time-division multiplexing, thus limiting computational speed. The adaptive compression-decompression algorithm designed in this invention constructs an information compression matrix through feature vector analysis, enabling efficient compression of information based on its temporal characteristics. Information reconstruction is achieved by decompressing the high-dimensional mapping information output from hidden layer 200. This novel compression-computation-decompression paradigm significantly improves the information processing speed for large-scale complex tasks. This invention achieves efficient processing of complex tasks with an extremely simplified network structure. Furthermore, the hidden layer 200 of this invention is designed with a streamlined structure, eliminating the fiber delay feedback structure that is difficult to integrate in traditional photonic reservoir computing systems. This makes the system easier to integrate and provides high information processing speed, physical stability, and scalability without significantly reducing reservoir computing performance.

[0135] Based on Embodiment 1, this Embodiment 2 provides a wavelength division multiplexing photonic reservoir computing system, including:

[0136] Input layer 100 includes:

[0137] The compression layer includes:

[0138] The preprocessing module is used to convert input data into electrical signals, preprocess the electrical signals to obtain preprocessed electrical signals, and merge all preprocessed electrical signals to obtain an electrical signal matrix.

[0139] The mean vector acquisition module is used to calculate the mean of each column of the electrical signal matrix to obtain the mean vector of all preprocessed electrical signals.

[0140] The central data matrix acquisition module is used to acquire the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals.

[0141] The covariance matrix acquisition module is used to obtain the covariance matrix based on the total number of electrical signals and the central data matrix.

[0142] The eigenvalue decomposition module is used to perform eigenvalue decomposition on the covariance matrix to obtain a set of eigenvectors and their corresponding eigenvalues.

[0143] The compressed matrix acquisition module is used to combine the first preset number of feature vectors after descending sorting according to the size of the feature values ​​to obtain the compressed matrix;

[0144] The compression module is used to compress the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix;

[0145] The arbitrary waveform generator 12 has its input end connected to the output end of the compression layer, and is used to convert each compressed electrical signal in the compressed electrical signal matrix into an adjustable driving electrical signal that adapts to the driving requirements of each parallel unit modulator 13.

[0146] Multiple parallel units, each of which includes:

[0147] Drive a laser to output a continuous optical carrier with a fixed wavelength;

[0148] The polarization controller 11, whose input terminal is connected to the output terminal of the driving laser, is used to perform polarization state calibration on the continuous optical carrier output by the driving laser to obtain the polarization state calibrated continuous optical carrier.

[0149] The modulator 13 has its input terminals connected to the output terminals of the polarization controller 11 and the arbitrary waveform generator 12, respectively. It is used to receive the continuous optical carrier after polarization state calibration and the adjustable driving electrical signal corresponding to the modulator 13. It is used to modulate the adjustable driving electrical signal onto the light output by the driving laser through phase modulation to obtain a modulated optical signal carrying compression information.

[0150] The wavelength division multiplexer 14 has its input terminals connected to the output terminals of the modulators 13 of each parallel unit, and is used to receive the modulated optical signals carrying compressed information output by each modulator 13, perform multiplexing processing, and obtain a multiplexed optical signal.

[0151] Hidden layer 200 includes:

[0152] The optical circulator 21 has one port connected to the output of the input layer of the streamlined photon reservoir computing system for receiving combined optical signals;

[0153] The Fabry-Perot laser 22 has its input end connected to port 2 of the optical circulator 21. It is used to receive the combined optical signal, convert the combined optical signal into optical signals of different wavelengths and independent longitudinal modes, and send them to the optical circulator 21. Each longitudinal mode carries one input compressed information, realizing the natural parallel carrying of multi-channel information.

[0154] Optical coupler 23, whose input end is connected to port 3 of optical circulator 21, is used to receive optical signals of different wavelengths and independent longitudinal modes and distribute them to each parallel amplification and filtering unit.

[0155] Multiple parallel amplification and filtering units, whose input terminals are connected to the output terminal of the optocoupler 23, are used to receive corresponding independent longitudinal mode optical signals. Each amplification and filtering unit includes:

[0156] Erbium-doped fiber amplifier 24, whose input end is connected to the output end of optical coupler 23, is used to receive the optical signal distributed by the optical coupler and amplify the power, compensate for the energy loss of the optical signal during transmission, and ensure the signal strength of subsequent processing.

[0157] Optical filter 25, whose input end is connected to the output end of erbium-doped fiber amplifier 24, is used to filter clutter from the amplified optical signal and separate the corresponding independent longitudinal mode high-dimensional nonlinear optical signal as the final output signal.

[0158] Output layer 300 includes:

[0159] Multiple photodetectors 31, whose input terminals are respectively connected to the output terminals of corresponding optical filters 25, are used to receive the final output signal output by the corresponding optical filters 25 and convert it into an electrical signal;

[0160] Adder 32, whose input terminals are respectively connected to the output terminals of each photodetector 31, is used to summarize the electrical signals output by each photodetector 31 to obtain the output electrical signal;

[0161] The programmable gate array 33, whose input is connected to the output of the adder 32, is used to decompress the output electrical signal based on the transpose of the compression matrix; and to obtain the processing result of the input data based on the decompressed output electrical signal.

[0162] The process of obtaining the input data processing result based on the decompressed output electrical signal includes: performing a weighted operation on the decompressed signal and the pre-trained readout layer weights to obtain the input data processing result.

[0163] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0164] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0165] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0166] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A wavelength division multiplexing photonic neural morphology calculation method, characterized in that, include: The input data is converted into an electrical signal, and the electrical signal is preprocessed to obtain a preprocessed electrical signal. All preprocessed electrical signals are combined to obtain an electrical signal matrix; The mean vector of all preprocessed electrical signals is obtained by averaging each column of the electrical signal matrix. The central data matrix is ​​obtained based on the electrical signal matrix and the mean vector of all preprocessed electrical signals. Based on the total number of electrical signals and the central data matrix, the covariance matrix is ​​obtained. The covariance matrix is ​​decomposed into eigenvalues ​​to obtain a set of eigenvectors and their corresponding eigenvalues. Based on the magnitude of the eigenvalues, the eigenvectors are sorted in descending order. The first predetermined number of eigenvectors after descending sorting are then combined to obtain a compressed matrix. Based on the compression matrix, the electrical signal matrix is ​​compressed to obtain the compressed electrical signal matrix; The compressed electrical signal matrix is ​​flattened by time-division multiplexing and then passed through the input layer and hidden layer of the photonic reservoir computing system to obtain high-dimensional nonlinear optical signals with multiple independent longitudinal modes. Multiple independent longitudinal mode high-dimensional nonlinear optical signals are converted into electrical signals by passing them through the corresponding photodetectors in the output layer of the photonic reservoir computing system. The multiple converted electrical signals are combined by the adder in the output layer of the photon reservoir computing system to obtain the output electrical signal; The output electrical signal is decompressed using a programmable gated array at the output layer of a photonic reservoir computing system, based on the transpose of the compression matrix. The processing result of the input data is obtained based on the decompressed output electrical signal.

2. The wavelength division multiplexing photonic neuromorphic computing method according to claim 1, characterized in that, The input data can be any one of image data, text data, or audio data.

3. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 1, characterized in that, The formula for obtaining the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals is as follows: ， in, As the central data matrix, For electrical signal matrix, This represents a column vector of all ones with dimension h×1, where h is the total number of electrical signals. For the first A preprocessed electrical signal, This is the mean vector of all preprocessed electrical signals. For electrical signal index.

4. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 1, characterized in that, The formula for obtaining the covariance matrix based on the total number of electrical signals and the central data matrix is ​​as follows: ， in, Let h be the covariance matrix, and h be the total number of electrical signals. As the central data matrix, for The transpose of .

5. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 1, characterized in that, Methods for compressing electrical signal matrices based on compression matrices to obtain compressed electrical signal matrices include: Multiplying the electrical signal matrix by the compression matrix yields the compressed electrical signal matrix.

6. The wavelength division multiplexing photonic neuromorphic computing method according to claim 1, characterized in that, The photon reservoir computing system is a streamlined photon reservoir computing system. The hidden layer of the streamlined photon reservoir computing system includes: An optical circulator, whose port 1 is connected to the output of the input layer of a streamlined photon reservoir computing system, is used to receive combined optical signals; The Fabry-Perot laser has its input end connected to port 2 of the optical circulator to receive the combined optical signal, convert the combined optical signal into a high-dimensional nonlinear optical signal with independent longitudinal modes of different wavelengths, and send it to the optical circulator. An optical coupler, whose input is connected to port 3 of an optical circulator, is used to receive optical signals of different wavelengths and independent longitudinal modes and distribute them to various parallel amplification and filtering units. Multiple parallel amplification and filtering units, whose inputs are connected to the output of an optocoupler, are used to receive corresponding independent longitudinal mode optical signals. Each amplification and filtering unit includes: An erbium-doped fiber amplifier, whose input is connected to the output of an optical coupler, is used to receive the optical signal distributed by the optical coupler and amplify its power. An optical filter, whose input is connected to the output of an erbium-doped fiber amplifier, is used to filter the amplified optical signal, separate the corresponding independent longitudinal mode high-dimensional nonlinear optical signal, and send it to the output layer of the streamlined photon reservoir computing system.

7. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 1, characterized in that, Methods for preprocessing electrical signals to obtain preprocessed electrical signals include: Based on the mask signal and the mask scaling factor, the electrical signal is preprocessed to obtain the preprocessed electrical signal.

8. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 7, characterized in that, The formula for preprocessing the electrical signal based on the mask signal and the mask scaling factor to obtain the preprocessed electrical signal is as follows: ， in, This represents the preprocessed electrical signal. Indicates an electrical signal. Indicates the mask signal. represents the mask scaling factor, and t represents continuous time.

9. The wavelength division multiplexing photonic neuromorphic calculation method according to claim 7, characterized in that, The mask signal can be any one of a binary mask signal, a multi-valued mask signal, or a chaotic mask signal.

10. A wavelength division multiplexing photonic neuromorphic computing system, characterized in that, include: The preprocessing module is used to convert input data into electrical signals, preprocess the electrical signals, and obtain preprocessed electrical signals. All preprocessed electrical signals are combined to obtain an electrical signal matrix; The mean vector acquisition module is used to calculate the mean of each column of the electrical signal matrix to obtain the mean vector of all preprocessed electrical signals. The central data matrix acquisition module is used to acquire the central data matrix based on the electrical signal matrix and the mean vector of all preprocessed electrical signals. The covariance matrix acquisition module is used to obtain the covariance matrix based on the total number of electrical signals and the central data matrix. The eigenvalue decomposition module is used to perform eigenvalue decomposition on the covariance matrix to obtain a set of eigenvectors and their corresponding eigenvalues. The compression matrix acquisition module is used to sort the feature vectors in descending order according to the size of the feature values, and then combine the first preset number of feature vectors after descending sorting to obtain the compression matrix. The compression module is used to compress the electrical signal matrix based on the compression matrix to obtain the compressed electrical signal matrix; The optical neuromorphic computing module is used to flatten the compressed electrical signal matrix through time-division multiplexing, and then pass it through the input layer and hidden layer of the photon reservoir computing system to obtain high-dimensional nonlinear optical signals with multiple independent longitudinal modes. The photoelectric conversion module is used to convert high-dimensional nonlinear optical signals with multiple independent longitudinal modes into electrical signals through the corresponding photodetectors in the output layer of the photonic reservoir computing system. The aggregation module is used to aggregate multiple converted electrical signals through the adder of the output layer of the photon reservoir computing system to obtain the output electrical signal; The output module is used to decompress the output electrical signal based on the transpose of the compression matrix using a programmable gated array of the output layer of the photon reservoir computing system. The processing result of the input data is obtained based on the decompressed output electrical signal.

Images