Photonic processor and data recognition method, device, equipment, medium, and product

Abstract

This invention discloses a photonic processor and data recognition method, apparatus, device, medium, and product, relating to the field of optical computing. The simulation optical path for the computation process includes multiple tunable photonic devices deployed according to an iterative computation process of the matrix to be computed. Two optical output channels of the cyclic iteration control optical path are respectively connected to the input end of the simulation optical path and the result reading optical path. The optical path switching controller transmits the optical compensation signals output from the first k-1 iterations to the simulation optical path for the next iteration, and transmits the optical compensation signals from the k-1th and kth iterations to the result reading optical path. The result reading optical path transmits the result of the k-1th iteration to the simulation optical path, and obtains feature data based on the two computation results. This invention solves the problem of requiring multiple repeated photoelectric / electro-optical conversions in related technologies, effectively improving data processing efficiency and saving resources.

Description

Technical Field

[0001] This invention relates to the field of optical computing, and in particular to a photonic processor and a data recognition method, apparatus, device, medium, and product. Background Technology

[0002] With the explosive growth of data volume, the computational complexity of traditional matrix feature problems has increased dramatically. In order to improve the computational efficiency of matrix data, photonic computing is currently used to replace traditional electronic computing. In the process of using photonic processors to process multiple iterative matrix operations, each round of iterative operation requires the intermediate results to be output from the optical domain to the electronic storage unit. The repeated photoelectric conversion / electro-optical conversion leads to low overall data processing efficiency. Summary of the Invention

[0003] This invention provides a photonic processor and a target data recognition method, device, electronic device, non-volatile storage medium, and computer program product, which can realize the calculation of matrix feature problems with a photonic processor with fewer photoelectric / electro-optical conversions, effectively improving data processing efficiency and saving resources.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0005] The present invention also provides a photonic processor, including a computation process simulation optical path, a cyclic iteration control optical path, and a result reading optical path. The computation process simulation optical path includes multiple tunable photonic devices deployed according to a single iterative computation process of the matrix to be calculated, and its output is connected to the input of the cyclic iteration control optical path; the matrix to be calculated obtains feature data after k iterative computations.

[0006] The first optical path output channel of the cyclic iterative control optical path is connected to the input end of the simulation optical path of the calculation process, and the second optical path output channel is connected to the result reading optical path. Its optical path switching controller transmits the optical compensation signal corresponding to the first k-1 iteration calculation to the simulation optical path of the calculation process through the first optical path output channel for the next iteration calculation, and transmits the optical compensation signal corresponding to the k-1th and kth iteration calculations to the result reading optical path through the second optical path output channel.

[0007] The result is that the reading optical path is connected to the input end of the simulation optical path of the operation process, and the first data signal corresponding to the (k-1)th iteration operation is transmitted to the simulation optical path of the operation process for the kth iteration operation; the feature data is obtained based on the first data signal and the second data signal corresponding to the kth iteration operation.

[0008] The present invention also provides a target data identification method, comprising:

[0009] When a data recognition task is received to identify target data from task data, task data related to the data recognition task is obtained.

[0010] The task data is represented in matrix form to obtain the matrix to be calculated, and the feature data of the matrix to be calculated is the target data.

[0011] When the calculation process of the feature data of the matrix to be calculated involves multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, the feature data of the matrix to be calculated is obtained by calling any of the above-mentioned photonic processors to process the matrix, and then used as the processing result of the data recognition task.

[0012] The present invention also provides a target data identification device, comprising:

[0013] The task data acquisition module is used to acquire task data related to the data recognition task when a data recognition task is received to identify target data from task data.

[0014] The data processing module is used to represent the task data in matrix form to obtain the matrix to be calculated, and the feature data of the matrix to be calculated is the target data.

[0015] The task execution module is used to process the matrix to be calculated by calling any of the above photonic processors when the calculation process of the feature data of the matrix to be calculated involves multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, so as to obtain the feature data of the matrix to be calculated as the processing result of the data recognition task.

[0016] The present invention also provides an electronic device, including a memory and a processor, wherein the processor is used to implement the steps of any of the target data recognition methods described above when executing a computer program stored in the memory.

[0017] The present invention also provides a non-volatile storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of any of the target data recognition methods described above.

[0018] Finally, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the target data recognition methods described above.

[0019] The advantage of the technical solution provided by this invention lies in that, for a matrix whose feature data is calculated through multiple iterative processes, based on the mapping relationship between a tunable photonic device and matrix elements, a set of tunable photonic devices can simulate the single iterative operation process of the matrix. The optical path is controlled through iterative cycles to compensate the optical signal output from each iteration. Based on the current iteration number, the optical signal after the previous k-1 compensations is controlled to enter the simulation optical path to participate in the next iteration. The optical compensation signals from the k-1th and kth iterations are output from the optical domain to the result reading module for photo-to-electric conversion and written to the electronic storage unit. Then, the output signal from the k-1th iteration... The input signal is converted back into an optical signal, and the optical path is controlled iteratively in the final iteration. The entire process requires only two photoelectric / electro-optical conversions, which reduces the number of photoelectric / electro-optical conversions by k-2 compared to related technologies. This not only effectively avoids the increased system latency caused by photoelectric signal conversion and improves matrix data processing efficiency, but also avoids the energy consumption caused by k-2 photoelectric / electro-optical conversions, saving resources and effectively reducing the power consumption of optical computing. This also effectively improves the computing power of photonic processors and avoids the problem of limited data transmission rate caused by the sampling rate and bandwidth limitations of photoelectric / electro-optical conversion not being able to fully match the high-speed characteristics of optical computing, further improving the efficiency of matrix data processing.

[0020] Furthermore, the present invention also provides a suitable method, apparatus, electronic device, non-volatile storage medium, and computer program product for the photonic processor, further enhancing the practicality of the photonic processor, and providing corresponding advantages for the target data identification method, apparatus, electronic device, non-volatile storage medium, and computer program product. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the present invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic diagram of the hardware framework applicable to the photonic processor provided by the present invention;

[0023] Figure 2 A schematic diagram of the structural framework of the photonic processor provided by the present invention in an exemplary embodiment;

[0024] Figure 3 A schematic diagram of the structure of the photonic processor provided by the present invention in another exemplary embodiment;

[0025] Figure 4A schematic diagram of the structure of the first type of photonic device provided by the present invention in an exemplary embodiment;

[0026] Figure 5 A schematic diagram of the matrix representation optical device array provided by the present invention in an exemplary embodiment;

[0027] Figure 6 A schematic diagram of the structure of the second type of photonic device provided by the present invention in an exemplary embodiment;

[0028] Figure 7 A schematic diagram of the optical path selector provided by the present invention in an exemplary embodiment;

[0029] Figure 8 A schematic diagram of the structure of the optical path simulating the computation process provided by the present invention in an exemplary embodiment;

[0030] Figure 9 A flowchart illustrating a target data identification method provided by the present invention;

[0031] Figure 10 This is a structural framework diagram of an exemplary embodiment of the target data recognition device provided by the present invention;

[0032] Figure 11 This is a structural diagram of an exemplary embodiment of the electronic device provided by the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. In this specification and the aforementioned drawings, the terms "first," "second," "third," "fourth," etc., are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. The term "exemplary" means "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as superior to or better than other embodiments.

[0034] With the rapid development of artificial intelligence and big data technologies, the scale of data is exploding, including high-dimensional images, large-scale graph data, and deep learning parameter matrices. Traditional matrix feature calculation methods, such as the QR (orthogonal matrix integral) algorithm, have seen a dramatic increase in computational complexity for large-scale matrix data. For example, for a matrix with dimension 1... For matrices, the complexity of classical algorithms is typically as high as [missing value]. This makes it difficult for it to meet user needs when processing millions or even larger amounts of data.

[0035] To overcome the physical limitations of traditional electronic processors due to Moore's Law and the limitations of the von Neumann architecture, and to solve the bottleneck problems of computing power and functionality in current classical computers, optical computing, a computing paradigm that uses photonic signals as the basic carrier of information transmission and processing, has emerged. Optical computing utilizes the ultra-high-speed transmission and interference characteristics of light waves to perform rapid calculations, possessing advantages such as large bandwidth, ultra-high speed, low power consumption, low latency, and high parallelism. Photonic processors based on optical computing, with their inherent high parallelism, low latency, and low energy consumption, are applied in fields with high computational complexity. For example, photonic processors can be used to perform matrix-vector multiplication operations in image classification tasks and speech recognition tasks.

[0036] In practical applications, such as data science, physical modeling, and engineering optimization, solving high-dimensional, dynamic, and large-scale problems requires multiple iterative matrix operations on the characteristic data of a matrix to meet subsequent processing needs. Its advantage lies in balancing accuracy and efficiency; that is, sacrificing some mathematical rigor for computational feasibility, resource conservation, and real-time response capabilities in practical scenarios. Currently, in implementing such iterative algorithms, photonic processors in related technologies require outputting intermediate results from the optical domain to an electronic storage unit for processing before converting them back into optical signals. The calculation of a single eigenvalue of the matrix requires multiple repeated photoelectric / electro-optical conversions. The high-speed, high-precision photoelectric conversion process typically takes nanoseconds, far exceeding the picosecond propagation speed of optical signals within the photonic processor. This makes photoelectric conversion a bottleneck for the overall data processing efficiency of the photonic processor. Furthermore, frequent photoelectric signal conversions not only increase the data processing latency of the photonic processor but also increase additional energy consumption, partially offsetting the low-power advantage of optical computing itself. Meanwhile, the sampling rate and bandwidth limitations of some photoelectric / electro-optical conversions may not be able to fully match the high-speed characteristics of optical computing, resulting in limited data transmission rates and affecting the overall computing efficiency of photonic processors.

[0037] In view of this, in order to solve the above problems, this invention is based on the mapping relationship of one tunable photonic device to one matrix element. It uses a set of tunable photonic devices to simulate the single iteration process of the matrix. By cyclically iterating and controlling the optical path, the optical signal output of each iteration calculation is compensated. The opening of two optical path output channels is controlled according to the current iteration number. The optical signal after the first k-1 compensations is transmitted into the calculation process to simulate the optical path to participate in the next iteration calculation. The optical compensation signal corresponding to the k-1 iteration calculation is transmitted to the result reading circuit. The result reading circuit converts the k-1 optical compensation signal back into an optical signal input. The cyclically iterating and controlling the optical path to participate in the last iteration is used. The optical compensation signal corresponding to the k-th iteration calculation is transmitted to the result reading circuit for result reading. The whole process only requires two photoelectric / electro-optical conversions by the result reading circuit, thereby realizing the calculation of matrix feature problems with a photonic processor with fewer photoelectric / electro-optical conversions, so as to save resources and improve computing power.

[0038] The specific application environment architecture or hardware architecture upon which the matrix feature data calculation process depends is described here. The following section will combine... Figure 1 Examples of possible application scenarios related to the technical solutions of this invention are provided below:

[0039] An artificial intelligence platform is deployed on the photonic computer 1. For artificial intelligence tasks that require calculating the feature values ​​of matrix data, such as data dimensionality reduction, pattern recognition, network analysis, and model optimization, the artificial intelligence platform is implemented by the photonic processor 10 of the photonic computer 1. The photonic computer is a new type of computer that uses optical signals to perform digital operations, logical operations, information storage, and processing. It is composed of at least optical elements and devices such as lasers, optical mirrors, lenses, and filters. Information processing is performed by the laser beam entering the array composed of mirrors and lenses, and photons are used to replace electrons to realize optical operations instead of electrical operations.

[0040] In this embodiment, for the matrix involved in the artificial intelligence task, the main eigenvalue (i.e. the eigenvalue with the largest absolute value) and the matrix operation required for the corresponding eigenvector can be recursively calculated according to the power iteration algorithm. If the matrix operation is matrix multiplication or matrix and vector multiplication, the matrix operation involved in the matrix iteration process can be implemented by the photonic processor 10. The photonic processor 10 is constructed using optical silicon-based devices and may include at least a computational process simulation optical path, a cyclic iteration control optical path, and a result reading optical path. The computational process simulation optical path includes multiple tunable photonic devices. Based on the spatial position of each photonic device corresponding to the position of matrix elements, the value of each element is determined by adjusting the phase of the corresponding photonic device. Each photonic device is deployed according to one iteration of the matrix to be calculated. The cyclic iteration control optical path compensates for transmission loss in the optical signal output from one iteration of the computational process simulation optical path and controls the transmission of the optical compensation signals corresponding to the first k-1 iterations to the computational process simulation optical path for the next iteration. It also controls the transmission of the optical compensation signals corresponding to the (k-1)th and kth iterations to the result reading optical path. The result reading optical path reads the first data signal corresponding to the (k-1)th iteration and transmits it to the computational process simulation optical path for the kth iteration. It also reads the second data signal corresponding to the kth iteration; the second data signal is the principal eigenvalue and its corresponding eigenvector of the matrix. If other eigenvalues ​​need to be calculated, based on the current matrix A and the eigenvalues ​​of matrix A... and its corresponding eigenvectors Through relational formulas A new matrix is ​​constructed. By dynamically adjusting the phase of each photonic device, the elements of matrix A are adjusted to the corresponding elements of the new matrix. This new matrix is ​​then used as the current matrix. The above process is executed using a photonic processor to obtain the second largest eigenvalue and its corresponding eigenvector. This process is repeated until all eigenvalues ​​and their corresponding eigenvectors of the matrix are obtained.

[0041] As can be seen from the above, the photonic processor in this embodiment only requires two photoelectric / electro-optical conversions to perform matrix operations involved in artificial intelligence tasks. This enables the calculation of matrix feature problems to be performed using a photonic processor with fewer photoelectric / electro-optical conversions, which can effectively save resources and improve computing power.

[0042] It should be noted that the above application scenarios are only shown to facilitate understanding of the ideas and principles of the present invention, and the embodiments of the present invention are not limited in any way. On the contrary, the embodiments of the present invention can be applied to any applicable scenario. After introducing the technical solution of the present invention, various non-limiting embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Please refer to [link to previous text] first. Figure 2 , Figure 2This is a schematic structural framework diagram of the photonic processor provided in this embodiment, which may include the following:

[0043] The photonic processor comprises a computational simulation optical path 201, a loop iteration control optical path 202, and a result reading optical path 203. All three paths are built upon optical components, utilizing the parallel propagation of optical signals and optical interference characteristics to replace electrons as the information carrier for data processing, achieving efficient parallel computation of the matrix across the entire optical domain. Its basic principle is as follows: the initial input vector or matrix is ​​encoded into multiple parallel light intensity / phase signals, which are dynamically mapped to preset matrix weights using tunable photonic devices (such as a Mach-Zehnder interferometer array). During transmission, the multipath optical signals undergo weight loading and optical field interference superposition, naturally performing multiplication and addition operations. Finally, the output, for example, can convert the multidimensional interference light intensity distribution into an electrical signal via a grating coupler array, and the computation results are read in parallel by a photodetector.

[0044] The simulation optical path 201 includes three signal input terminals. The first terminal receives the laser signal emitted by the laser. The second terminal receives the compensation optical signal from the previous iteration transmitted by the iterative control optical path 202. The third terminal receives the data from the (k-1)th read transmitted by the result reading optical path 203. The output of the simulation optical path 201 is connected to the input of the iterative control optical path 202, transmitting the output result of each iteration to the iterative control optical path 202. The iterative control optical path 202 includes two output channels: one connected to the second signal input of the simulation optical path 201, and the other connected to the input of the result reading optical path 203. One output of the result reading optical path 203 outputs the result, and the other output is connected to the third signal input of the simulation optical path 201. Different photonic devices can be connected via optical waveguides, such as optical fibers or silicon waveguides.

[0045] In this embodiment, as Figure 3As shown, the computation simulation module 201 includes multiple tunable photonic devices. The spatial position of each photonic device corresponds to the position of a matrix element. The photonic devices are deployed according to one iteration of the computation process of the matrix to be calculated. This deployment includes the position of each photonic device and the connection relationships between them. The values ​​of each element are obtained by adjusting the phase of the corresponding photonic device through thermo-optic modulation or electro-optic modulation. The matrix to be calculated is obtained by performing k iterations of feature data, where these k iterations are matrix multiplication or matrix-vector multiplication operations. The feature data of the matrix to be calculated is determined based on the outputs of the (k-1)th iteration and the kth iteration. The cyclic iteration control optical path 202 includes at least one input terminal and two output channels, namely a first optical path output channel and a second optical path output channel. The first optical path output channel of the cyclic iteration control optical path 202 is connected to the input terminal of the computation simulation optical path 201, and the second optical path output channel is connected to the result reading optical path 203. The iterative control optical path 202 also includes an optical path switching controller that controls which output channel to select. The optical path switching controller can be set manually or triggered by an automated program. The optical path switching controller transmits the optical compensation signal corresponding to the first k-1 iterations to the simulation optical path 202 for the next iteration through the first optical path output channel, and transmits the optical compensation signals corresponding to the k-1th and kth iterations to the result reading optical path 203 through the second optical path output channel. Furthermore, considering the transmission loss of optical signals between different components, such as MZI couplers and optical waveguides between different components, to avoid transmission loss in subsequent iterative processes, the transmission loss of the iterative calculation optical signal output by the simulation optical path 202 can be compensated. For ease of description, the compensated signal is defined as the optical compensation signal. Accordingly, the iterative control optical path 202 compensates for the transmission loss of the optical signal output by the simulation module 201 after completing one iterative calculation, and controls the optical compensation signal corresponding to the first k-1 iterative calculations to be transmitted to the simulation module for the next iterative calculation. It also controls the optical compensation signals corresponding to the k-1 and k-th iterative calculations to be transmitted to the result reading optical path 203. The result reading optical path 203 transmits the first data signal corresponding to the k-1 iterative calculation to the simulation optical path 201 for the k-th iterative calculation.In other words, when the simulation optical path 202 completes the (k-1)th iteration, it outputs the optical signal from the (k-1)th iteration to the loop iteration control optical path 202. The loop iteration control optical path 202 transmits the signal to the result reading optical path 203 through the second output channel. The result reading optical path 203 reads the optical signal and stores it in the electronic storage unit. For ease of description, the signal obtained after the (k-1)th iteration output, compensation, and reading processing is defined as the first data signal. The first data signal is then transmitted to the simulation module 201 for the k-th iteration. After the simulation module 201 completes the k-th iteration, it outputs the signal to the result reading optical path 203 through the second optical path output channel. The result reading optical path 203 reads the signal to obtain the second data signal. The corresponding data of the second data signal and the first data signal are processed to obtain the feature data. The obtained feature data is then output through the output terminal of the result reading optical path 203. For example, the result reading optical path 203 may include at least a plurality of photodetectors, that is, the final photodetector acquires the output to characterize the principal eigenvalues ​​of the matrix and the corresponding eigenvectors.

[0046] In the technical solution provided in this embodiment, for a matrix whose feature data is calculated through multiple iterative processes, based on the mapping relationship between one tunable photonic device and one matrix element, a set of tunable photonic devices can be used to simulate the single iteration process of the matrix. The optical path is controlled through iterative cycles to compensate the optical signal output from each iteration. Based on the current iteration number, the optical signal after the previous k-1 compensations is controlled to enter the simulation optical path to participate in the next iteration. The optical compensation signals for the k-1th and kth iterations are output from the optical domain to the electronic storage unit, and then the optical compensation signal for the k-1th iteration is converted back into an optical signal. The input signal is used to iteratively control the optical path to participate in the last iteration. The entire process only requires two photoelectric / electro-optical conversions, which reduces the number of photoelectric / electro-optical conversions by k-2 compared to related technologies. This not only effectively avoids the increased system latency caused by photoelectric signal conversion and improves matrix data processing efficiency, but also avoids the energy consumption caused by k-2 photoelectric / electro-optical conversions, saving resources and effectively reducing the power consumption of optical computing. This effectively enhances the computing power of the photonic processor and avoids the problem of limited data transmission rate caused by the sampling rate and bandwidth limitations of photoelectric / electro-optical conversion not being able to fully match the high-speed characteristics of optical computing, further improving the efficiency of matrix data processing.

[0047] The above embodiments do not limit how the optical path of the computation process simulates a single iterative operation of the matrix. This embodiment also provides an exemplary implementation. In this embodiment, the optical path of the computation process can repeatedly apply the matrix to be calculated to an arbitrary initial input array, which can be a vector or a matrix, so that the result after each iteration gradually converges to the direction of the principal eigenvector. Accordingly, the optical path 201 of the computation process may include an initial signal input encoding array and a matrix representation optical device array. The initial signal input encoding array includes multiple tunable first-type photonic devices. Each first-type photonic device is arranged according to the dimension and array element arrangement of the initial input array, so that each first-type photonic device corresponds one-to-one with the array element of the initial input array. Each first-type photonic device encodes the input light emission signal into the corresponding array element value by adjusting the phase. The matrix representation optical device array includes multiple tunable second-type photonic devices. Each second-type photonic device is arranged according to the row dimension, column dimension, and matrix element arrangement of the matrix to be calculated, so that each second-type photonic device corresponds one-to-one with the matrix element of the matrix to be calculated. The connection relationship between each type I photonic device and each type II photonic device is determined according to the multiplication operation between the initial input array and the matrix to be calculated. Each type I photonic device outputs the encoded optical signal to the corresponding type II photonic device, and each type II photonic device outputs the received optical signal to the cyclic iteration control optical path to complete one iteration operation of the feature data of the matrix to be calculated.

[0048] In this embodiment, as Figure 3 As shown, the initial signal input encoding array corresponds to the initial input array. Encode each array element Corresponding to a single type I photonic device, Here, 'n' represents the array dimension and also the total number of Type I photonic devices. The dynamic changes of array elements are controlled by modulating the phase of the Type I photonic devices. The laser emits an optical signal that enters each Type I photonic device. An external electrical signal can be used to change the phase difference between the devices, making the optical transmission rate function equal to the values ​​of the array vector elements or matrix elements. This converts the electrical signal into an optical signal input. Currently, phase control methods include electro-optic / thermo-optic modulation, which adjusts the input voltage or device temperature via an electrical signal to regulate the phase. The matrix represents the optical device array as a network array composed of Type II photonic devices, representing the matrix A to be calculated. Each element of matrix A corresponds to the optical power transmission function of a Type II photonic device. Dynamic control of the matrix elements is achieved by adjusting the phase difference. Thus, the output of the Type I photonic devices represents the initial input array. The optical signal enters the optical device array represented by the matrix A to be calculated. At this time, the optical signal output by the optical device array represents the output vector. ,like Figure 5 As shown.

[0049] The above embodiments do not impose any limitations on the structure of the first type of photonic device and the second type of photonic device, as long as the first type of photonic device and the second type of photonic device are tunable photonic components, such as any kind of interferometer. This embodiment also provides an exemplary structure of the first type of photonic device and the second type of photonic device, which may include the following:

[0050] For the first type of photonic device, such as Figure 4 As shown, the structures of the various first-type photonic devices are identical. Each first-type photonic device may include a first input coupler, a third phase shifter, a fourth phase shifter, and a first output coupler. The first input terminal of the first input coupler is connected to the laser, and the output terminal of the first input coupler is connected to one end of the third and fourth phase shifters. The other ends of the third and fourth phase shifters are both connected to the input terminal of the first output coupler, and the third output terminal of the first output coupler is connected to the input terminal of the corresponding second-type photonic device. The first input coupler divides the optical emission signal input through the first input terminal into a first optical emission signal and a second optical emission signal. For ease of... The description defines the signal received by the first type of photonic device as an optical emission signal, and the two beams of optical signals are defined as the first optical emission signal and the second optical emission signal. The first optical emission signal is input to the third phase shifter, and the second optical emission signal is input to the fourth phase shifter. By adjusting the phase difference between the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter, the corresponding array element values ​​are obtained. The first output coupler combines the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter into an optical coded signal. The combined optical signal is defined as the optical coded signal, and finally the optical coded signal is output through the third output terminal.

[0051] For example, the adjustment process for array element values ​​can be as follows: For each element of the initial input array, the difference between twice the current array element value and 1 is used as the phase difference cosine value; based on the fact that twice the phase difference between the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter is the same as the phase difference cosine value, the current value of the phase difference between the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter is calculated; the phase difference between the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter is adjusted to the current value using a phase modulation method. For example, the phase difference calculation formula can be pre-stored. , This represents the phase difference between the optical emission signal output by the third phase shifter and the optical emission signal output by the fourth phase shifter, where 'a' represents the value of the current array element.

[0052] For the second type of photonic device, such as Figure 5 and Figure 6 As shown, the structures of the second type of photonic devices are identical. Each second type of photonic device includes a second input coupler, a fifth phase shifter, a sixth phase shifter, and a second output coupler. The third input terminal of the second input coupler is connected to the output terminal of the corresponding first type of photonic device. The output terminal of the second input coupler is connected to one end of the fifth and sixth phase shifters. The other ends of the fifth and sixth phase shifters are both connected to the input terminal of the second output coupler. The fifth output terminal of the second output coupler is connected to the input terminal of the corresponding signal amplifier. The second input coupler divides the optically encoded signal input through the third input terminal into a first optically encoded signal and a second optically encoded signal. The first optically encoded signal is input to the fifth phase shifter, and the second optically encoded signal is input to the sixth phase shifter. By adjusting the phase difference between the optically encoded signals output by the fifth and sixth phase shifters, the corresponding matrix element values ​​are obtained. The second output coupler synthesizes the optically encoded signals output by the fifth and sixth phase shifters into an optical operation signal and outputs the optical operation signal through the fifth output terminal.

[0053] For example, the process of adjusting the matrix element values ​​can be as follows: For each element of the matrix to be calculated, the difference between twice the current element value and 1 is taken as the target phase difference value; the phase difference value between the optically encoded signals output by the fifth and sixth phase shifters is calculated based on the fact that twice the cosine of the phase difference between the optically encoded signals output by the fifth and sixth phase shifters is the same as the target phase difference value; the phase difference between the optically encoded signals output by the fifth and sixth phase shifters is adjusted to the target phase difference value using a phase modulation method. For example, the phase difference calculation formula can be pre-stored. , represents the phase difference between the optically encoded signal output by the fifth phase shifter and the optically encoded signal output by the sixth phase shifter, and b represents the value of the current matrix element.

[0054] Based on the above embodiments, the result reading optical path 203 detects a first data signal. The first data signal includes at least an intermediate vector and a vector magnitude with the same dimension as the initial input array. In order to realize the k-th iteration operation, the first data signal needs to be converted into a corresponding optical signal by controlling the initial signal input encoding array of the optical path 202 through cyclic iteration to participate in the k-th iteration operation. Accordingly, the vector elements of the intermediate vector correspond one-to-one with each of the first type of photonic devices in the initial signal input encoding array. The phase difference of the corresponding first type of photonic device is adjusted according to the vector element value of the intermediate vector. For example, the phase value of each first type of photonic device is adjusted by electro-optic modulation or thermo-optic modulation so that the optical power transfer function value of the first type of photonic device is the same as the vector element value of the intermediate vector. In this way, each first type of photonic device inputs the adjusted optical signal to the corresponding second type of photonic device. When each second type of photonic device outputs the corresponding optical signal, the k-th iteration operation is completed.

[0055] As can be seen from the above, this embodiment simulates matrix operations by using the first type of photonic device and the second type of photonic device. The function can be switched by adjusting the phase in real time without hardware modification. It has stronger reconfigurability and versatility. It does not require energy to maintain the optical path. It only consumes energy when adjusting the phase, which reduces energy consumption. The optical signal is not affected by circuit noise, making it suitable for complex electromagnetic environments and helping to improve data processing accuracy.

[0056] The above embodiments do not limit the structure of the cyclic iterative control optical path. Based on the above embodiments, the present invention also provides an exemplary structure of the cyclic iterative control optical path, which may include the following: The cyclic iterative control optical path includes a signal amplifier group and an optical path selector group; the total number of signal amplifiers in the signal amplifier group is the same as the total number of optical path selectors in the optical path selector group; the signal amplifier group contains multiple signal amplifiers, the number of which is the same as the number of optical signals output in one iteration operation; one end of each signal amplifier is connected to the photonic device corresponding to one optical signal output from the initial signal input encoding array, and the other end is connected to the input end of one optical path selector in the optical path selector group; each optical path selector in the optical path selector group includes a first optical path output channel, a second optical path output channel, and an optical path switching controller; the first optical path output channel is connected to the simulation optical path of the operation process, the second optical path output channel is connected to the result reading optical path, and the optical path switching controller determines whether the optical compensation signal output by the corresponding signal amplifier enters the first optical path output channel or the second optical path output channel according to the current iteration operation number.

[0057] In this embodiment, the signal amplifiers and optical path selectors of the cyclic iterative control optical path are in one-to-one correspondence. Each signal amplifier and optical path selector corresponds to one optical signal output by the cyclic iterative control optical path. One optical signal represents the dimension of the output data obtained from one iteration. For example, if the initial input array has a dimension of 4×1 and the matrix to be calculated has a dimension of 4×4, then the output result has a dimension of 4×1. Correspondingly, four optical signals are output, resulting in four signal amplifiers and four optical path selectors for the cyclic iterative control optical path. In other words, the number of signal amplifiers and optical path selectors in the cyclic iterative control optical path corresponds to the dimension of the matrix or vector corresponding to the optical signal output by the cyclic iterative control optical path after one iteration.

[0058] Considering that optical signals experience transmission losses as they pass through optical waveguides, couplers, and MZI (Mach-Zehnder Interferometer), due to material absorption and scattering, insertion loss, or splitting loss, excessive losses can lead to signal strength falling below the receiver's detection threshold, resulting in increased bit error rate or communication interruption. This embodiment directly amplifies the optical signal using a signal amplifier to compensate for losses and maintain system performance. The signal amplifier can be any photonic device capable of compensating for optical signal transmission losses, such as an erbium-doped fiber amplifier (EDFA), which compensates for the transmission losses in the output optical waveguide, coupler, and interferometer.

[0059] Furthermore, the present invention also provides an exemplary result of an optical path selector, such as... Figure 7 As shown, the optical path selector may include a beam splitter, a first phase shifter, a second phase shifter, and a beam combiner. The input terminal of the beam splitter is connected to the corresponding signal amplifier, and the output terminal of the beam splitter is connected to one end of the first phase shifter and the second phase shifter. The beam splitter splits the optical compensation signal received through the input terminal into a first optical compensation signal and a second optical compensation signal, and inputs the first optical compensation signal to the first phase shifter and the second optical compensation signal to the second phase shifter. The other ends of the first phase shifter and the second phase shifter are both connected to one end of the beam combiner. The beam combiner includes a first output terminal and a second output terminal. The first output terminal is connected to the simulation optical path of the calculation process, and the second output terminal is connected to the result reading optical path. The beam combiner combines the optical signals output by the first phase shifter and the second phase shifter into a new optical signal, and outputs the new optical signal through the first output terminal or the second output terminal.

[0060] In this embodiment, the optical path selector controls whether the optical signal output from a single iteration enters the simulation optical path of the computation process or is read by the result reading optical path, thus achieving iterative computation. When the optical path selector is configured to "On," the output optical signal is allowed to enter the simulation optical path of the computation process; when configured to "Off," the optical signal is allowed to enter the photodetector for result reading. The transmission matrix corresponding to the above optical path selector can be represented as:

[0061] .

[0062] in, 'i' represents the imaginary unit. When inputting from the terminal... One port receives a compensation optical signal, while the other port receives no signal. At this point, it is possible to calculate from Input to output terminal or The output optical power transfer function is: Phase difference can be controlled by electro-optic / thermo-optic modulation, that is, by adjusting the input voltage or the temperature of the device through an electrical signal. The adjustment. When When, it indicates the "On" state, and the output signal is from The output enters the cyclic feedback loop, when When, it indicates the "Off" state, and the output signal is from The output is fed into a photodetector to obtain the result. That is, for example, for the first k-1 iterations, the phase difference between the optical compensation signal output from the first phase shifter and the optical compensation signal output from the second phase shifter is adjusted as follows: This allows a new optical signal to be input into the simulation optical path of the computation process through the first output terminal; for the (k-1)th and kth iterations, the phase difference between the optical compensation signal output by the first phase shifter and the optical compensation signal output by the second phase shifter is adjusted to 0, so that a new optical signal is input into the result reading optical path through the second output terminal.

[0063] As can be seen from the above, this embodiment compensates for the transmission loss of the optical signal by cyclically iteratively controlling the optical path, which is not limited by the transmission distance and can achieve multiple cyclic iterations. Optical path selection is achieved through phase adjustment, which not only has a fast switching speed but also extremely low energy consumption.

[0064] Considering that the computation process for the matrix to be computed is sometimes impossible to implement using a photonic processor, or that the overall cost is higher than that of an electronic processor, this invention can also use a power iteration algorithm to recursively summarize the result of k iterations of the matrix to be computed and extract the corresponding matrix operations. If the matrix operation process is a matrix multiplication operation or a matrix and vector multiplication operation, then the photonic processor of this invention is used to process the matrix to be computed, which may include the following:

[0065] The computational process simulates an optical path comprising an initial signal input encoding array and a matrix representation optical device array. The initial signal input encoding array includes a set of tunable first-type photonic devices arranged vertically, with the same number as the initial input vector dimension. Each first-type photonic device corresponds one-to-one with the vector element of the initial input vector, and the input optical signal is encoded into the corresponding vector element value by adjusting the phase. The matrix representation optical device array includes multiple tunable second-type photonic devices, which are arranged according to the row dimension, column dimension, and matrix element arrangement of the matrix to be calculated, so that each second-type photonic device corresponds one-to-one with the matrix element of the matrix to be calculated. The connection relationship between each first-type photonic device and each second-type photonic device is determined according to the multiplication operation between the initial input vector and the matrix to be calculated. Each first-type photonic device outputs the encoded optical signal to the corresponding second-type photonic device, and each second-type photonic device outputs the received optical signal to the cyclic iteration control optical path to complete one iteration operation of calculating the principal eigenvalues ​​and corresponding principal eigenvectors of the matrix to be calculated using a power iteration algorithm.

[0066] In this embodiment, the matrix to be calculated is subjected to k matrix-vector multiplication operations or matrix multiplication operations through a power iteration algorithm to obtain the principal eigenvalues ​​and corresponding principal eigenvectors. The process of deriving the principal eigenvalues ​​(i.e., the eigenvalues ​​with the largest absolute values) and their corresponding eigenvectors of this real symmetric matrix through a power iteration algorithm may include:

[0067] Iteration 0: Select the initial input vector The initial input vector is normalized to a unit vector. .

[0068] First iteration: , The vector output from the first iteration is normalized to a unit vector. .

[0069] Second iteration: , The vector output by the second iteration, based on any constant. Multiply by vector The overall modulus is equal to a constant multiple. The absolute value multiplied by the vector The modulus length, that is The relation of the second iteration Defined as Then the vector Normalized to unit vector .

[0070] 3rd iteration: , The vector output by the 3rd iteration, the vector Normalized to unit vector By continuously iterating, the result of the k-th iteration is:

[0071] .

[0072] Preset accuracy ,like When the power iteration algorithm stops iterating, it outputs the principal eigenvalue. The corresponding feature vector is The number of iterations and the precision of the power iteration algorithm are determined by the matrix properties and the required precision; their values ​​can be adjusted according to the actual application. Based on the result of the k-th iteration, the result of the k-th iteration is... This requires k matrix-vector multiplication operations, and the output of the principal eigenvalue requires the results of the (k-1)th and kth iterations. Accordingly, this embodiment modulates the input optical signal using an initial signal input encoding array, outputting an optical signal representing the initial input vector to a matrix representation optical device array, such as... Figure 8 As shown, the matrix represents the optical device array, which represents the matrix to be calculated. When the optical signal passes through the output of the matrix-represented optical device array, one iteration process is completed.

[0073] Since the power iteration algorithm is only used to determine the principal eigenvalues ​​of the matrix to be calculated. (i.e., the eigenvalue with the largest absolute value) and its corresponding eigenvector In order to determine other eigenvalues ​​and eigenvectors, this embodiment also adopts the following implementation method:

[0074] Based on the matrix to be calculated, the principal eigenvalues, and the principal eigenvectors, the target matrix is ​​determined. The target matrix is ​​then subjected to k iterations using a power iteration algorithm to obtain the second largest eigenvalue and the corresponding second eigenvector of the matrix to be calculated. The matrix elements of the target matrix correspond one-to-one with each of the second type of photonic devices in the matrix-represented optical device array. The phase difference of the corresponding second type of photonic devices is adjusted according to the matrix element values ​​of the target matrix so that the optical power transfer function values ​​of the second type of photonic devices are the same as the corresponding matrix element values ​​of the target matrix. The laser is controlled to emit an optical signal again and input it to the initial signal input encoding array. After the simulated optical path completes k iterations of calculations during the calculation process, the result of the kth iteration read by the result reading optical path is used as the second largest eigenvalue and the corresponding second eigenvector.

[0075] In this embodiment, it can be done according to the relational formula The target matrix is ​​obtained by replacing the matrix used in the previous eigenvalue calculation process. For example, when calculating the second largest eigenvalue, the target matrix replaces the matrix to be calculated in the above embodiment, and the process of calculating the principal eigenvalue and the corresponding eigenvector using the photonic processor described above is used to obtain the second largest eigenvalue of the matrix to be calculated. and the corresponding feature vectors When calculating the third eigenvalue, the target matrix replaces the target matrix in the above embodiment, and the process of calculating the second eigenvalue and corresponding eigenvector using the aforementioned photonic processor is employed to obtain the third eigenvalue and corresponding eigenvector of the matrix to be calculated. This process is repeated until the matrix... All eigenvalues ​​and their corresponding eigenvectors are extracted. The method of replacing the matrix in the previous eigenvalue calculation process with the target matrix can be as follows: the target matrix is ​​A1. The phase of each second-type photonic device in the optical device array is adjusted according to the values ​​of each matrix element of A1, representing the real symmetric matrix A1. That is, each second-type photonic device corresponds to each matrix element in the second-type photonic device matrix. The phase difference is adjusted using thermo-optic / electro-optic effects to achieve dynamic control of the matrix elements. Using the aforementioned photonic processor, the following can be obtained: The main eigenvalues ​​are The feature vector is , which is the second largest eigenvalue of the matrix to be calculated and its corresponding eigenvector. Continue constructing At this point, the aforementioned photonic processor can be used to obtain... The main eigenvalues ​​are The feature vector is This means obtaining the third largest eigenvalue of the matrix to be calculated and its corresponding eigenvector. By continuing in this way, all the eigenvalues ​​and corresponding eigenvectors of the matrix to be calculated can be obtained.

[0076] Finally, to make the implementation of the present invention clearer to those skilled in the art, the present invention also provides an illustrative photonic processor, for example... Figure 8 Taking the architecture of the photonic processor shown as an example, Figure 8 Taking a 4×4 real symmetric matrix as an example, this matrix is ​​represented by a network array of 16 second-type photonic devices cascaded together. In this embodiment, the tunable photonic devices simulating the optical path of the computation process are all implemented using silicon-based MZIs (Mach-Zehnder interferometers). Each MZI in the computation process simulates an input signal at one input terminal, while the other input terminal receives no signal, and the signal is output through an output terminal. The optical path selector for the iterative control optical path is also implemented using silicon-based MZIs. The optical signal compensation components can be erbium-doped fiber amplifiers, and the result readout optical path is a set of photoelectric converters. Each silicon-based MZI can be composed of two couplers and two phase shifters. It achieves interference of different phases by adjusting the optical path difference. When the input optical signal passes through the first coupler, it is split into two beams, which pass through two different phase shifters, generating a certain phase difference. These two beams then pass through the second coupler and merge into a new optical signal, forming an interference signal. By adjusting the phase difference, the amplitude and phase of the optical signal can be precisely controlled, thereby completing the corresponding matrix operations.

[0077] In this embodiment, the matrix to be calculated is Real symmetric matrix of dimension Its eigenvalue size is The principal eigenvalues ​​of the matrix to be calculated can be recursively derived using a power iteration algorithm. And the matrix operations corresponding to the eigenvectors, when it is determined that the matrix operation to be calculated is a matrix multiplication operation or a matrix-vector multiplication operation, the matrix calculation process of the matrix to be calculated can be implemented using the photonic processor of this embodiment. The process of recursively calculating the matrix to be calculated A using the power iteration algorithm is as follows: iteration steps First, select an initial vector and normalize it, that is, select an initial input vector. and normalize it to a unit vector. , , This represents the magnitude of the vector. Then, in each iteration, this vector is multiplied by the matrix, and the result is normalized again, i.e., the magnitude of the vector is calculated. and normalize it to a unit vector. , The loop continues until the vector change falls below a preset precision. The result of the k-th iteration can be expressed as: ,when When the time is reached, the algorithm stops iterating and outputs the principal feature value. The corresponding feature vector is The result of the kth iteration is recursively calculated based on the above algorithm steps, and the required matrix operations are extracted. Therefore, it can be seen that when using the photonic processor of this embodiment to calculate the eigenvalues ​​of the matrix to be calculated, it is only necessary to read the calculation results of the (k-1)th and kth iterations.

[0078] The process of calculating the principal eigenvalues ​​of matrix A using a photonic processor may include the following:

[0079] A1: The laser emits an optical signal that enters each silicon-based MZI, and the output, by adjusting the phase, represents the initial input vector. The optical signal will represent the initial input vector. The input optical signal represents the matrix formed by the silicon-based MZIs in the optical device array, which is a real symmetric matrix A. At this time, the output optical signal has completed one matrix-vector multiplication operation, that is... .

[0080] A2: The optical signal output from the matrix-representation optical device array is amplified by an erbium-doped fiber amplifier to compensate for the transmission losses of the optical waveguide, coupler, and MZI. The optical signal is then cyclically fed back through the MZI, which acts as an optical path selector, to the matrix-representation optical device array composed of silicon-based MZI units representing a real symmetric matrix to complete matrix-vector multiplication operations. and will represent The optical signal is then input again into the matrix representation optical device array.

[0081] A3: Repeat step A2 until the (k-1)th iteration is completed. .

[0082] A4: Will represent The optical signal is transmitted to the photodetector through the MZI switch to obtain the signal. Information, Information includes vectors The element values ​​and the magnitude of the vector The laser emits a signal again, and the output is represented by adjusting the phase. The light signal, and will represent The optical signal input represents a matrix composed of silicon-based MZIs in a real symmetric matrix within the optical device array, completing the final matrix-vector multiplication operation, i.e. .

[0083] A5: will represent The optical signal is transmitted to the photodetector through the MZI switch to obtain the signal. The information. Finally, based on... The corresponding feature vector is Then you can obtain the main feature value and the corresponding feature vector.

[0084] As can be seen from the above, this embodiment first recursively calculates the matrix operations required for the principal eigenvalues ​​and corresponding eigenvectors of the matrix using a power iteration algorithm. Then, it designs a photonic processor with fewer photoelectric / electro-optical conversions to perform the matrix operations involved in the above iteration process to complete the calculation tasks of the principal eigenvalues ​​and eigenvectors. Finally, it constructs a specific matrix and uses the aforementioned photonic processor to sequentially calculate all the eigenvalues ​​and corresponding eigenvectors of the original matrix. This allows for the processing of matrix eigenvalue problems with a photonic processor that performs fewer photoelectric / electro-optical conversions, effectively improving data processing efficiency and saving resources.

[0085] Finally, this invention also provides an implementation process for executing a practical task using the photonic processor described in the above embodiments, where the eigenvalues ​​of a computed matrix can be used as the task execution result. Please refer to [link to relevant documentation]. Figure 9 , Figure 9 This is a flowchart illustrating a target data identification method provided in this embodiment. This embodiment may include the following:

[0086] S901: When a data recognition task is received to identify target data from task data, the task data related to the data recognition task is obtained.

[0087] Data identification tasks are those that can be solved by calculating matrix eigenvalues, including but not limited to data dimensionality reduction, pattern recognition, network analysis, and model optimization. For example, Principal Component Analysis (PCA) reduces data dimensionality by solving the eigenvalue problem of the covariance matrix (the matrix to be calculated); spectral clustering relies on the eigenvalues ​​of the Laplacian matrix (the matrix to be calculated) to classify data categories; and PageRank (link analysis) uses the eigenvectors corresponding to the maximal eigenvalues ​​of the adjacency matrix (the matrix to be calculated) to rank webpage importance. Task data refers to the data needed during the data identification task analysis process, while target data is the data corresponding to the eigenvalues ​​in the task data.

[0088] S902: Represent the task data in matrix form to obtain the matrix to be calculated. The feature data of the matrix to be calculated is the target data.

[0089] The task data can be constructed using any matrix form, such as covariance matrix, Laplace matrix, or adjacency matrix. Correspondingly, the matrices to be calculated are covariance matrix, Laplace matrix, and adjacency matrix.

[0090] S903: When the calculation process of the feature data of the matrix to be calculated involves multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, the matrix to be calculated is processed by a photonic processor to obtain the feature data of the matrix to be calculated, which is then used as the processing result of the data recognition task.

[0091] The target data can be the data corresponding to the principal feature value or the data corresponding to multiple feature values. The final processing result is determined according to different tasks. For example, if the data identification task is a hot event identification task, the processing result of the data identification task is the data corresponding to the principal feature value of the matrix to be calculated. If the data identification task is a data importance ranking task, the processing result of the data corresponding to multiple feature values ​​is the ranking result. For example, obtain each feature value of the matrix to be calculated and calculate the absolute value of each feature value; sort the feature values ​​according to the absolute value from largest to smallest or smallest to largest to obtain feature ranking information; the absolute value of each feature value is proportional to the importance of the corresponding data; sort the data corresponding to each feature value according to the feature ranking information to obtain the ranking result corresponding to the data importance ranking task.

[0092] In the technical solution provided in this embodiment, the efficiency of matrix data processing is effectively improved by using a photonic processor, thereby effectively improving the efficiency of target data recognition, enhancing task response performance, saving resources required during the execution of target data recognition tasks, and enabling deployment on resource-constrained edge devices, thus improving practicality.

[0093] To enable those skilled in the art to better understand the implementation of the present invention, the present invention also provides an exemplary task implementation method. In this embodiment, the data identification task is to detect trending events on social media platforms. The process of detecting trending events may include the following:

[0094] B1: Construct the adjacency matrix corresponding to the dynamic graph based on the dynamic social network structure. t represents time.

[0095] In this embodiment, a hot topic is a concentrated interaction among a large number of users around a specific topic within a short period of time, such as a sudden event. The graph structure of the dynamic social network (user follow relationships, tweet retweets) changes rapidly as the event unfolds. This step constructs the matrix to be calculated based on the dynamic social network structure, and the hot topic corresponds to the adjacency matrix. Sudden changes in the principal eigenvalues, such as a sharp increase in the maximum eigenvalue, reflect a surge in network activity. This allows for the combination of dynamic graph updates and rapid feature extraction of hot events, solving the problem that traditional static graph models cannot capture such sudden changes in real time.

[0096] For example, the process of constructing an adjacency matrix may include: obtaining each entity node of the dynamic social network, each preset time step and the corresponding preset time interval, and assigning a unique index value to each entity node; constructing an initial matrix to be filled with matrix element values ​​based on the total number of entity nodes, wherein the row dimension and column dimension of the initial matrix are the same as the total number of entity nodes; within each preset time step, determining the corresponding matrix element values ​​of the initial matrix based on whether different entity nodes are related within each preset time interval of the current preset time step, and using the initial matrix to be filled with matrix element values ​​as the matrix to be calculated.

[0097] In this context, entity nodes refer to entities with dynamic attributes, such as users. The total number of nodes in the entire dynamic social network is N, and each node is assigned an index, such as 0 to N-1. For each time step t, an N×N matrix is ​​created and initialized with all zeros, as shown below. Then, for each edge active within the time interval t, where the edge represents the relationship between different entity nodes, such as user follow relationships, user interactions, or forwarding, the corresponding position in the adjacency matrix is ​​set to 1. In this way, an adjacency matrix is ​​obtained at each time step.

[0098] For example: Suppose there are 3 entity nodes {v0, v1, v2}, and the preset time intervals are divided into two: [0, 5) and [5, 10]. Entity nodes v0 and v1 interact within the time interval [0, 5), entity nodes v1 and v2 interact within the time interval [5, 10], and entity nodes v1 and v0 interact within both the time intervals [0, 5) and [5, 10]. Then, the following adjacency matrix can be obtained for the two intervals: .

[0099] B2: Use the aforementioned photonic processor to quickly obtain the adjacency matrix. The principal eigenvalues, based on the adjacency matrix The main feature value information is used to detect hot events.

[0100] It should be noted that there is no strict order of execution between the steps in this invention. As long as they conform to the logical order, these steps can be executed simultaneously or in a certain preset order. Figure 9 This is just an illustrative example and does not mean that this is the only possible execution order.

[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0102] This invention also provides a corresponding apparatus for the target data identification method, further enhancing the method's practicality. The apparatus can be described from both the perspective of functional optical paths and hardware. The target data identification apparatus provided by this invention is described below. This apparatus is used to implement the target data identification method provided by this invention. In this embodiment, the target data identification apparatus may include or be divided into one or more program optical paths. These one or more program optical paths are stored in a storage medium and executed by one or more processors to complete the target data identification method disclosed in Embodiment 1. The program optical path referred to in this embodiment refers to a series of computer program instruction segments capable of performing a specific function, which is more suitable than the program itself for describing the execution process of the target data identification apparatus in the storage medium. The following description will specifically introduce the functions of each program optical path in this embodiment. The target data identification apparatus described below can be referred to in correspondence with the target data identification method described above.

[0103] From the perspective of the functional optical path, see Figure 10 , Figure 10 This is a structural diagram of the target data identification device provided in this embodiment under a specific implementation. The device may include:

[0104] The task data acquisition module 101 is used to acquire task data related to the data recognition task when a data recognition task for identifying target data from task data is received.

[0105] The data processing module 102 is used to represent the task data in matrix form to obtain the matrix to be calculated, and the feature data of the matrix to be calculated is the target data.

[0106] The task execution module 103 is used to process the matrix to be calculated through a photonic processor when the process of calculating the feature data of the matrix to be calculated includes multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, so as to obtain the feature data of the matrix to be calculated as the processing result of the data recognition task.

[0107] For example, in some embodiments of this example, the task execution module 103 is used for hotspot event identification tasks, and the data corresponding to the main feature values ​​of the matrix to be calculated is used as the processing result of the data identification task.

[0108] As an exemplary implementation of the above embodiments, the data processing module 102 can also be used to: obtain each entity node, each preset time step, and the corresponding preset time interval of the dynamic social network, and assign a unique index value to each entity node; construct an initial matrix to be filled with matrix element values ​​based on the total number of entity nodes, wherein the row dimension and column dimension of the initial matrix are the same as the total number of entity nodes; within each preset time step, in each preset time interval of the current preset time step, determine the corresponding matrix element values ​​of the initial matrix based on whether different entity nodes are related, and use the initial matrix to be filled with matrix element values ​​as the matrix to be calculated.

[0109] For example, in some other embodiments of this embodiment, the task execution module 103 is used to perform a data importance ranking task, obtain each feature value of the matrix to be calculated, and calculate the absolute value of each feature value; sort each feature value in descending or ascending order of absolute value to obtain feature ranking information; the absolute value of each feature value is proportional to the importance of the corresponding data; sort the data corresponding to each feature value according to the feature ranking information to obtain the ranking result corresponding to the data importance ranking task.

[0110] The target data recognition device mentioned above is described from the perspective of functional optical path. Furthermore, the present invention also provides an electronic device, which is described from the perspective of hardware. Figure 11 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. The electronic device includes a memory 111 and a processor 112. The memory 111 stores a computer program, and the processor 112 is configured to run the computer program to perform the steps in any of the above-described embodiments of the target data recognition method.

[0111] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described embodiments of the target data identification method when it is run.

[0112] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0113] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described target data recognition method embodiments.

[0114] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above-described target data identification method embodiments.

[0115] The foregoing has provided a detailed description of the photonic processor and target data identification method, apparatus, electronic device, computer-readable storage medium, and computer program product provided by the present invention. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Whether the units and algorithm steps of the various examples described in the disclosed embodiments are executed in electronic hardware or computer software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, and such implementations should not be considered beyond the scope of the present invention. Several improvements and modifications can be made to the present invention without departing from the principles of the invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A photonic processor, characterized in that, This includes the optical path for simulating the computation process, the optical path for controlling the loop iteration, and the optical path for reading the results. The computation process simulates an optical path comprising multiple tunable photonic devices deployed according to a single iterative computation of the matrix to be calculated, with their outputs connected to the input of the cyclic iterative control optical path; the matrix to be calculated undergoes k iterative computations to obtain feature data. The iterative control optical path includes a signal amplifier group and an optical path selector group; the total number of signal amplifiers is the same as the total number of optical path selectors in the optical path selector group, and is also the same as the number of optical signals output in one iteration; one end of each signal amplifier is connected to the photonic device corresponding to one optical signal output from the simulated optical path of the operation process, and the other end is connected to the input end of the corresponding optical path selector; each optical path selector includes a first optical path output channel, a second optical path output channel, and an optical path switching controller; the first optical path output channel is connected to the simulated optical path of the operation process, and the second optical path output channel is connected to the result reading optical path; the optical path switching controller determines whether the optical compensation signal output by the corresponding signal amplifier enters the first optical path output channel or the second optical path output channel according to the current iteration operation number: the optical compensation signal corresponding to the first k-1 iteration operation is transmitted to the simulated optical path of the operation process through the first optical path output channel for the next iteration operation, and the optical compensation signals corresponding to the k-1th and kth iteration operations are transmitted to the result reading optical path through the second optical path output channel; The result reading optical path is connected to the input end of the simulation optical path of the calculation process. The first data signal corresponding to the (k-1)th iteration calculation is transmitted to the simulation optical path of the calculation process for the kth iteration calculation, and the feature data is obtained based on the first data signal and the second data signal corresponding to the kth iteration calculation.

2. The photonic processor according to claim 1, characterized in that, The optical path selector includes a beam splitter, a first phase shifter, a second phase shifter, and a beam combiner; The input terminal of the beam splitter is connected to the corresponding signal amplifier, and the output terminal of the beam splitter is connected to one end of the first phase shifter and the second phase shifter. The beam splitter divides the optical compensation signal received through the input terminal into a first optical compensation signal and a second optical compensation signal, and inputs the first optical compensation signal to the first phase shifter and the second optical compensation signal to the second phase shifter. The other ends of the first phase shifter and the second phase shifter are both connected to one end of the beam combiner. The beam combiner includes a first output end and a second output end. The first output end is connected to the simulation optical path of the calculation process, and the second output end is connected to the result reading optical path. The beam combiner combines the optical signals output by the first phase shifter and the second phase shifter into a new optical signal, and outputs the new optical signal through the first output end or the second output end.

3. The photonic processor according to claim 2, characterized in that, Outputting the new optical signal through the first output terminal or the second output terminal includes: For the first k-1 iterations, the phase difference between the optical compensation signal output by the first phase shifter and the optical compensation signal output by the second phase shifter is adjusted to π, so that the new optical signal is input to the simulated optical path of the operation process through the first output terminal; For the (k-1)th and kth iterations, the phase difference between the optical compensation signal output by the first phase shifter and the optical compensation signal output by the second phase shifter is adjusted to 0, so that the new optical signal is input to the result reading optical path through the second output terminal.

4. The photonic processor according to claim 1, characterized in that, The computation process simulates an optical path including an initial signal input encoding array and a matrix representation optical device array; The initial signal input encoding array includes multiple tunable first-type photonic devices. Each first-type photonic device is arranged according to the dimension and array element arrangement of the initial input array, so that each first-type photonic device corresponds one-to-one with the array element of the initial input array. Each first-type photonic device encodes the input light emission signal into the corresponding array element value by adjusting the phase. The matrix represents an optical device array comprising multiple tunable second-type photonic devices, each second-type photonic device being arranged according to the row dimension, column dimension, and matrix element arrangement of the matrix to be calculated, so that each second-type photonic device corresponds one-to-one with the matrix element of the matrix to be calculated; The connection relationship between each type of first-class photonic device and each type of second-class photonic device is determined according to the multiplication operation method between the initial input array and the matrix to be calculated. Each type of first-class photonic device outputs the encoded optical signal to the corresponding type of second-class photonic device, and each type of second-class photonic device outputs the received optical signal to the cyclic iterative control optical path to complete one iterative operation of the matrix to be calculated to calculate the feature data.

5. The photonic processor according to claim 4, characterized in that, The first type of photonic device includes a first input coupler, a third phase shifter, a fourth phase shifter, and a first output coupler; The first input terminal of the first input coupler is connected to the laser, the output terminal of the first input coupler is connected to one end of the third phase shifter and the fourth phase shifter, the other end of the third phase shifter and the fourth phase shifter are both connected to the input terminal of the first output coupler, and the third output terminal of the first output coupler is connected to the input terminal of the corresponding second type of photonic device. The first input coupler divides the optical emission signal input through the first input terminal into a first optical emission signal and a second optical emission signal, and inputs the first optical emission signal to the third phase shifter and the second optical emission signal to the fourth phase shifter. By adjusting the phase difference between the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter, the corresponding array element values ​​are obtained. The first output coupler synthesizes the optical emission signals output by the third phase shifter and the optical emission signals output by the fourth phase shifter into an optical encoded signal, and outputs the optical encoded signal through the third output terminal.

6. The photonic processor according to claim 5, characterized in that, The corresponding array element values ​​are obtained by adjusting the phase difference between the light emission signal output by the third phase shifter and the light emission signal output by the fourth phase shifter, including: For each element of the initial input array, the difference between twice the value of the current array element and 1 is taken as the phase difference cosine value; The current value of the phase difference between the light emission signals output by the third phase shifter and the light emission signals output by the fourth phase shifter is calculated based on the fact that twice the phase difference is the same as the cosine value of the phase difference. The phase difference between the optical emission signal output by the third phase shifter and the optical emission signal output by the fourth phase shifter is adjusted to the current value using a phase modulation method.

7. The photonic processor according to claim 4, characterized in that, The second type of photonic device includes a second input coupler, a fifth phase shifter, a sixth phase shifter, and a second output coupler; The third input terminal of the second input coupler is connected to the output terminal of the corresponding first type of photonic device. The output terminal of the second input coupler is connected to one end of the fifth phase shifter and the sixth phase shifter. The other ends of the fifth phase shifter and the sixth phase shifter are both connected to the input terminal of the second output coupler. The fifth output terminal of the second output coupler is connected to the input terminal of the corresponding signal amplifier. The second input coupler divides the optically encoded signal input through the third input terminal into a first optically encoded signal and a second optically encoded signal. The first optically encoded signal is input to the fifth phase shifter, and the second optically encoded signal is input to the sixth phase shifter. By adjusting the phase difference between the optically encoded signals output by the fifth phase shifter and the optically encoded signals output by the sixth phase shifter, the corresponding matrix element values ​​are obtained. The second output coupler synthesizes the optically encoded signals output by the fifth phase shifter and the optically encoded signals output by the sixth phase shifter into an optical operation signal, and outputs the optical operation signal through the fifth output terminal.

8. The photonic processor according to claim 7, characterized in that, The corresponding matrix element values ​​are obtained by adjusting the phase difference between the optically encoded signal output by the fifth phase shifter and the optically encoded signal output by the sixth phase shifter, including: For each element of the matrix to be calculated, the difference between twice the value of the current element and 1 is taken as the target phase difference. The phase difference between the optically encoded signal output by the fifth phase shifter and the optically encoded signal output by the sixth phase shifter is calculated based on the fact that twice the cosine value of the phase difference between them is the same as the target phase difference value. The phase difference between the optically encoded signal output by the fifth phase shifter and the optically encoded signal output by the sixth phase shifter is adjusted to the phase difference value using a phase modulation method.

9. The photonic processor according to claim 4, characterized in that, The first data signal includes an intermediate vector and a vector magnitude with the same dimension as the initial input array. Transmitting the first data signal to the simulated optical path of the computation process includes: The vector elements of the intermediate vector correspond one-to-one with each of the first type of photonic devices in the initial signal input encoding array. The phase difference of the corresponding first type of photonic device is adjusted according to the vector element values ​​of the intermediate vector so that the optical power transfer function value of the first type of photonic device is the same as the vector element value corresponding to the intermediate vector. Each type I photonic device inputs the adjusted optical signal to the corresponding type II photonic device. When each type II photonic device outputs the corresponding optical signal, the k-th iteration operation is completed.

10. The photonic processor according to any one of claims 1 to 8, characterized in that, The matrix to be calculated is obtained by performing k matrix-vector multiplication operations or matrix multiplication operations through a power iteration algorithm to obtain the principal eigenvalues ​​and corresponding principal eigenvectors. The calculation process simulates an optical path including an initial signal input encoding array and a matrix representation optical device array. The initial signal input encoding array includes a set of tunable first-type photonic devices arranged vertically, with the same number as the dimension of the initial input vector. Each first-type photonic device corresponds one-to-one with the vector element of the initial input vector, and the input optical signal is encoded into the corresponding vector element value by adjusting the phase. The matrix represents an optical device array comprising multiple tunable second-type photonic devices, each second-type photonic device being arranged according to the row dimension, column dimension, and matrix element arrangement of the matrix to be calculated, so that each second-type photonic device corresponds one-to-one with the matrix element of the matrix to be calculated; The connection relationship between each type I photonic device and each type II photonic device is determined according to the multiplication operation between the initial input vector and the matrix to be calculated. Each type I photonic device outputs the encoded optical signal to the corresponding type II photonic device, and each type II photonic device outputs the received optical signal to the cyclic iterative control optical path to complete one iterative operation of calculating the principal eigenvalue and corresponding principal eigenvector of the matrix to be calculated by the power iteration algorithm.

11. The photonic processor according to claim 10, characterized in that, Also includes: The target matrix is ​​determined based on the matrix to be calculated, the principal eigenvalues, and the principal eigenvectors. The target matrix is ​​obtained by performing k iterations using a power iteration algorithm to obtain the second largest eigenvalue and the corresponding second eigenvector of the matrix to be calculated. The matrix elements of the target matrix correspond one-to-one with each of the second type of photonic devices in the optical device array represented by the matrix. The phase difference of the corresponding second type of photonic devices is adjusted according to the matrix element values ​​of the target matrix so that the optical power transfer function value of the second type of photonic device is the same as the matrix element value corresponding to the target matrix. The laser is controlled to emit an optical signal again and input it into the initial signal input encoding array. After the simulated optical path completes k iterations of the calculation process, the result of the kth iteration read by the result reading optical path is used as the second largest feature value and the corresponding second feature vector.

12. A target data identification method, characterized in that, include: When a data identification task is received to identify target data from task data, task data related to the data identification task is obtained; The task data is represented in matrix form to obtain the matrix to be calculated, and the feature data of the matrix to be calculated is the target data. When it is determined that the feature data calculation process of the matrix to be calculated includes multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, the feature data of the matrix to be calculated is obtained by calling the photonic processor as described in any one of claims 1 to 11, and is used as the processing result of the data recognition task.

13. The target data identification method according to claim 12, characterized in that, The data identification task is a hotspot event identification task, and the processing result of the data identification task is the data corresponding to the main eigenvalues ​​of the matrix to be calculated.

14. The target data identification method according to claim 13, characterized in that, The task data is represented in matrix form, including: Obtain each entity node, each preset time step, and the corresponding preset time interval of the dynamic social network, and assign a unique index value to each entity node; An initial matrix is ​​constructed based on the total number of entity nodes, wherein the row and column dimensions of the initial matrix are the same as the total number of entity nodes. Within each preset time step, in each preset time interval of the current preset time step, the corresponding matrix element values ​​of the initial matrix are determined according to whether there is a relationship between different entity nodes, and the initial matrix filled with matrix element values ​​is used as the matrix to be calculated.

15. The target data identification method according to claim 12, characterized in that, The data identification task is a data importance ranking task, and the process of generating the results of the data identification task includes: Obtain the eigenvalues ​​of the matrix to be calculated, and calculate the absolute value of each eigenvalue; The feature values ​​are sorted in descending or ascending order of absolute value to obtain feature ranking information; the absolute value of each feature value is proportional to the importance of the corresponding data. For the data corresponding to each feature value, sort them according to the feature sorting information to obtain the sorting result corresponding to the data importance sorting task.

16. A target data identification device, characterized in that, include: The task data acquisition module is used to acquire task data related to the data recognition task when a data recognition task for identifying target data from task data is received. The data processing module is used to represent the task data in matrix form to obtain a matrix to be calculated, wherein the feature data of the matrix to be calculated is the target data; The task execution module is configured to, when it is determined that the feature data calculation process of the matrix to be calculated includes multiple iterations of matrix-vector multiplication or multiple iterations of matrix multiplication, call the photonic processor as described in any one of claims 1 to 11 to perform data processing on the matrix to be calculated, and obtain the feature data of the matrix to be calculated as the processing result of the data recognition task.

17. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the target data identification method as described in any one of claims 12 to 15.

18. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores a computer program, which, when executed by a processor, implements the steps of the target data identification method as described in any one of claims 12 to 15.

19. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the target data identification method according to any one of claims 12 to 15.

Images