Optoelectronic fusion reconfigurable analog intelligent computing system and task learning method therefor

Abstract

The present application relates to an optoelectronic fusion reconfigurable analog intelligent computing system and a task learning method therefor. The system comprises: an optical analog computing module and an electrical analog computing module, wherein the optical analog computing module is used for performing feature extraction and dimensionality reduction processing on an inputted optical analog signal to obtain an optical feature signal that meets a preset dimension, and the electrical analog computing module is used for converting the optical feature signal into an electrical analog signal and performing computation on the electrical analog signal to obtain an analog computation result. Therefore, by establishing a simulation model, determining a training set on the basis of a current learning task, performing numerical simulation training on the optoelectronic fusion analog computing system by using machine learning and other algorithms, and continuously optimizing parameters during the training to finally successfully complete the current learning task, the present application solves the problem of the limitations of an all-optical artificial neural network and a digital computing processor in completing a machine learning task and realizes advantages of high speed, low energy consumption, high robustness, and reconfigurability.

Description

Optoelectronic fusion reconfigurable analog intelligent computing system and its task learning method

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application is based on the Chinese patent application with application number 202311828354.0 and application date December 27, 2023, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby introduced into this application as a reference. Technical Field

[0003] The present application relates to the field of optoelectronic computing and machine learning technology, and in particular to an optoelectronic fusion reconfigurable analog intelligent computing system and a task learning method thereof. Background Art

[0004] Artificial intelligence and deep learning technologies have widespread applications in autonomous driving, robotics, healthcare, and other fields. The growing demand for AI places higher demands on processor energy consumption, computing speed, and robustness against noise. However, the energy consumption and speed of digital processors are constrained by factors such as slowing transistor process iterations and limited analog-to-digital conversion bandwidth, making them unable to meet the computing power demands of the AI ​​era. Photons, with their high throughput, wide bandwidth, and ability to process information at the speed of light, hold great promise for optical computing, both complementing and replacing electronic computing.

[0005] In related technologies, the all-optical diffraction deep neural network realizes an all-optical machine learning discriminant model. This architecture is optimized through machine learning design, similar to the combination of multi-level spatial-frequency domain optical phase modulation layers and nonlinear layers of artificial neural networks, to achieve functions such as handwritten digits, fashion product image classification, and image saliency analysis. The all-optical diffraction deep neural network provides an effective and unique all-optical machine learning model that uses passive components to implement diffraction operations at the speed of light. An important advantage of it is that it can be easily expanded by using various high-throughput, large-area 3D (three-dimensional, three-dimensional graphics) manufacturing methods and wide-field-of-view optical components and detection systems. It can realize hundreds of millions of neurons and billions of connections in a scalable, low-power and cost-effective manner, and has the potential to realize various complex applications.

[0006] However, optical neural networks in related technologies still face problems such as difficulty in integrating optical nonlinearity, non-reconfigurability, the need for analog-to-digital conversion for subsequent calculations, and poor robustness to noise. This makes it impossible for existing optical computing systems or optoelectronic fusion computing systems to demonstrate advantages over electronic digital computing systems in actual machine learning and computer vision tasks.

[0007] Summary of the Invention

[0008] The present application provides an optoelectronic fusion reconfigurable analog intelligent computing system and its task learning method to address the limitations of all-optical artificial neural networks and digital computing processors in completing machine learning tasks in related technologies. The system can complete machine learning tasks such as image classification, achieves functional reconfiguration, and has the advantages of high speed, low energy consumption, and high robustness. It effectively reduces the bandwidth limitations and energy consumption costs brought by analog-to-digital conversion and has broad application value.

[0009] The first embodiment of the present application provides an optoelectronic fusion reconfigurable analog intelligent computing system, including: an optical analog computing module and an electrical analog computing module.

[0010] Among them, the optical simulation calculation module is used to perform feature extraction and dimensionality reduction processing on the input optical simulation signal to obtain an optical characteristic signal that meets the preset dimension; the electrical simulation calculation module is used to convert the optical characteristic signal into an electrical simulation signal, and calculate the electrical simulation signal to obtain a simulation calculation result.

[0011] Optionally, in some embodiments, the optical simulation calculation module includes:

[0012] An optical analog signal input component is used to input the optical analog signal; an optical diffraction phase modulator is used to extract features and reduce the dimension of the optical analog signal to obtain an optical characteristic signal that meets the preset dimension; and an optical characteristic signal output component is used to output the optical characteristic signal.

[0013] Optionally, in some embodiments, the electrical simulation calculation module includes:

[0014] A photoelectric signal conversion element is used to convert the optical characteristic signal into the electrical analog signal using a photoelectric detector; an electrical analog signal matrix operation element is used to perform matrix operation on the electrical analog signal through a preset analog circuit, and obtain the analog calculation result based on Kirchhoff's law and the matrix operation result; an electrical analog signal output element is used to output the analog calculation result.

[0015] Optionally, in some embodiments, the optoelectronic fusion reconfigurable analog intelligent computing system further includes: an electrical digital computing module, which is used to convert the electrical analog signal into a digital signal and process it in a preset digital domain to obtain a digital computing result.

[0016] Optionally, in some embodiments, the electrical digital computing module includes:

[0017] An analog-to-digital signal conversion element is used to convert the electrical analog signal into a digital electrical signal; an electrical digital signal operation element is used to operate and process the digital electrical signal to obtain the digital calculation result; and an electrical digital signal output element is used to output the digital calculation result.

[0018] Optionally, in some embodiments, the optical analog signal is a coherent light field of input data information or a light field of a natural scene.

[0019] Optionally, in some embodiments, the optical analog signal and the optical characteristic signal satisfy the following relationship: l=Wx;

[0020] Where l is the output low-dimensional optical characteristic signal light field, W is the light field transformation matrix corresponding to the free propagation process of light and the optical diffraction phase modulator, and x is the input optical simulation signal light field.

[0021] Optionally, in some embodiments, the optical characteristic signal and the simulation calculation result satisfy the following relationship: i=g(l); y=Bi;

[0022] Where i is the output current of the photoelectric signal conversion module, function g is the response characteristic of the photodetector in the photoelectric signal conversion component to the input light field, l is the low-dimensional optical characteristic signal light field input to the electrical analog calculation module, y is the output voltage signal of the electrical analog calculation module, B is the matrix parameter corresponding to the electrical analog signal matrix operation component, and i is the output current of the photoelectric signal conversion module.

[0023] Optionally, in some embodiments, the optical diffraction phase modulator is physically manufactured by 3D printing or photolithography technology.

[0024] A second aspect of the present application provides a task learning method for an optoelectronic fusion reconfigurable simulated intelligent computing system, using the optoelectronic fusion reconfigurable simulated intelligent computing system as described in any of the above embodiments, wherein the method includes the following steps:

[0025] Establishing a first numerical simulation model of the optical diffraction phase modulation layer and a second numerical simulation model of the electrical simulation matrix operation layer in the optoelectronic fusion reconfigurable analog intelligent computing system;

[0026] Obtaining a current task to be learned and determining a training set based on the current task to be learned; based on the training set, performing numerical simulation training on the optoelectronic fusion reconfigurable analog intelligent computing system using a preset backpropagation algorithm optimization until the training results meet a preset end condition, optimizing the parameters of the first numerical simulation model and the parameters of the second numerical simulation model based on the training results to obtain an optimized optoelectronic fusion reconfigurable analog intelligent computing system;

[0027] The current task to be learned is completed according to the optimized optoelectronic fusion reconfigurable simulated intelligent computing system.

[0028] According to the optoelectronic fusion reconfigurable analog intelligent computing system provided in the present application, the system includes an optical analog computing module and an electrical analog computing module, and an electrical digital computing module can also be selectively added. In the optical analog computing module, the optical diffraction phase modulation layer structure is used to extract features and reduce the dimension of the input optical analog signal to obtain a low-dimensional optical feature signal; in the electrical analog computing module, the low-dimensional optical feature signal output by the optical analog computing module is converted into an analog electrical signal, the analog electrical signal is operated, and the analog calculation result is output; in the electrical digital computing module, the input electrical analog signal is converted into a digital signal, and further operations and signal processing are completed in the digital domain to obtain a digital calculation result.

[0029] Therefore, this application can establish a simulation model, determine the training set according to the current learning task, use machine learning and other algorithms to perform numerical simulation training on the optoelectronic fusion analog computing system, and continuously optimize the parameters during the training process to finally complete the current learning task. In this way, the limitations of all-optical artificial neural networks and digital computing processors in completing machine learning tasks in related technologies are solved. The system can complete machine learning tasks such as image classification, realize functional reconfiguration, and has the advantages of high speed, low energy consumption, and high robustness, effectively reducing the bandwidth limitations and energy consumption costs brought by analog-to-digital conversion, and has broad application value.

[0030] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0032] FIG1 is a block diagram of an optoelectronic fusion reconfigurable analog intelligent computing system provided according to an embodiment of the present application;

[0033] FIG2 is a schematic diagram of an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application;

[0034] FIG3 is a schematic diagram of the structure of an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application;

[0035] FIG4 is a schematic diagram of the principle of an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application;

[0036] FIG5 is a flow chart of a task learning method for an optoelectronic fusion reconfigurable analog intelligent computing system according to an embodiment of the present application;

[0037] FIG6 is a flow chart of a method for completing a machine learning task using an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application. DETAILED DESCRIPTION

[0038] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0039] The following describes the optoelectronic fusion reconfigurable analog intelligent computing system and its task learning method of the embodiment of the present application with reference to the accompanying drawings. In response to the problem of the limitations of the all-optical artificial neural network and digital computing processor in completing machine learning tasks mentioned in the above background technology, the present application provides an optoelectronic fusion reconfigurable analog intelligent computing system, which includes: an optical analog computing module and an electrical analog computing module, wherein the optical analog computing module is used to perform feature extraction and dimensionality reduction processing on the input optical analog signal to obtain an optical feature signal that meets the preset dimension; the electrical analog computing module is used to convert the optical feature signal into an electrical analog signal, and calculate the electrical analog signal to obtain an analog calculation result. As a result, the limitations of the all-optical artificial neural network and digital computing processor in completing machine learning tasks are solved. The system can complete machine learning tasks such as image classification, realizes functional reconfiguration, and has the advantages of high speed, low energy consumption, high robustness, etc., effectively reducing the bandwidth limitation and energy consumption cost brought by analog-to-digital conversion, and has broad application value.

[0040] Specifically, FIG1 is a block diagram of the optoelectronic fusion reconfigurable analog intelligent computing system provided in an embodiment of the present application.

[0041] As shown in FIG1 , the optoelectronic fusion reconfigurable analog intelligent computing system 10 includes an optical analog computing module 100 and an electrical analog computing module 200 .

[0042] Among them, the optical simulation calculation module 100 is used to perform feature extraction and dimensionality reduction processing on the input optical simulation signal to obtain an optical characteristic signal that meets the preset dimension; the electrical simulation calculation module 200 is used to convert the optical characteristic signal into an electrical simulation signal, and calculate the electrical simulation signal to obtain a simulation calculation result.

[0043] Optionally, in some embodiments, the optical analog signal is a coherent light field of input data information or a light field of a natural scene.

[0044] It is understandable that the optical analog signal input of the embodiment of the present application can be a coherent light field loaded with input data information, or a light field of a natural scene, or an analog optical signal obtained by other means, which is not specifically limited here.

[0045] Specifically, the present application can utilize the optical analog computing module 100 in the optoelectronic fusion reconfigurable analog intelligent computing system 10 to process the input optical analog signal, that is, perform feature processing on the optical analog signal to obtain a low-dimensional optical feature signal.

[0046] Furthermore, the present application can input the low-dimensional optical characteristic signal output by the optical simulation calculation module 100 into the electrical simulation calculation module 200, and the electrical simulation calculation module 200 converts the low-dimensional optical characteristic signal into an electrical simulation signal, and performs analog calculation processing on the signal to obtain an electrical simulation output result.

[0047] As shown in Figure 2, Figure 2 is a schematic diagram of an optoelectronic fusion reconfigurable analog intelligent computing system of a specific embodiment of the present application. The optoelectronic fusion reconfigurable analog intelligent computing system 10 of the embodiment of the present application can also include an electrical digital computing module 300. This module can be added selectively and is not specifically limited here. Those skilled in the art can add this module according to actual needs.

[0048] Furthermore, in some embodiments, as shown in FIG2 , the above-mentioned optoelectronic fusion reconfigurable analog intelligent computing system 10 further includes: an electrical digital computing module 300 for converting electrical analog signals into digital signals, and processing them in a preset digital domain to obtain digital computing results.

[0049] Specifically, the analog calculation results of the electrical analog calculation module 200 of the embodiment of the present application can be used as the input of the electrical digital calculation module 300. The electrical digital calculation module 300 converts the analog electrical signal into a digital electrical signal, and completes further calculations and signal processing in the digital domain, and finally obtains a digital calculation result, which can be used as the output result of the entire optoelectronic fusion reconfigurable analog intelligent computing system 10.

[0050] Thus, the optoelectronic fusion reconfigurable analog intelligent computing system 10 of the present application achieves the completion of all machine learning tasks in the analog domain through optoelectronic hybrid methods, minimizing the bandwidth limitations and energy consumption costs caused by analog-to-digital conversion. In addition, it should be noted that the analog calculation results output by the electrical analog calculation module 200 of the embodiment of the present application can be used as the output results of the entire optoelectronic fusion reconfigurable analog intelligent computing system 10, and can also be connected to the electrical digital calculation module 300 of the embodiment of the present application for further processing, without specific limitation here.

[0051] The following will list specific embodiments and combine with the accompanying drawings to introduce in detail the structure and principle of the optoelectronic fusion reconfigurable analog intelligent computing system 10 of the present application.

[0052] Specifically, as shown in FIG3 , FIG3 is a schematic diagram of the structure of an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application.

[0053] Optionally, in some embodiments, the optical simulation calculation module 100 includes: an optical simulation signal input component 101 for inputting an optical simulation signal; an optical diffraction phase modulation component 102 for performing feature extraction and dimensionality reduction on the optical simulation signal to obtain an optical characteristic signal that meets a preset dimension; and an optical characteristic signal output component 103 for outputting an optical characteristic signal.

[0054] It can be understood that the optical analog signal input through the optical analog signal input component 101 undergoes phase modulation by the optical diffraction phase modulator 102 during the broadcast process, thereby realizing the feature extraction of the input optical analog signal. Then, the input optical analog signal is encoded and compressed into a low-dimensional space, and a low-dimensional optical characteristic signal is obtained in the optical characteristic signal output component 103, and the result is output.

[0055] Optionally, in some embodiments, the optical diffraction phase modulator 102 is physically manufactured by 3D printing or photolithography technology.

[0056] Specifically, the structure of the optical diffraction phase modulator 102 in the embodiment of the present application can be manufactured using 3D printing or photolithography technology, and the parameters of the structure of the optical diffraction phase modulator 102 can be optimized using a machine learning method.

[0057] Therefore, the present application can utilize the optical diffraction network in the optical simulation calculation module 100 to realize feature extraction and optical encoding of a large-flux input analog optical signal.

[0058] Optionally, in some embodiments, the optical analog signal and the optical characteristic signal satisfy the following relationship: l=Wx;

[0059] Among them, l is the output low-dimensional optical characteristic signal light field, W is the light field transformation matrix corresponding to the light passing through the free propagation process and the optical diffraction phase modulation layer, and x is the input optical simulation signal light field.

[0060] Optionally, in some embodiments, the electrical analog calculation module 200 includes: a photoelectric signal conversion element 201, which is used to convert the optical characteristic signal into an electrical analog signal using a photodetector; an electrical analog signal matrix operation element 202, which is used to perform matrix operations on the electrical analog signal through a preset analog circuit, and obtain analog calculation results based on Kirchhoff's law and the matrix operation results; and an electrical analog signal output element 203, which is used to output the analog calculation results.

[0061] In the embodiment of the present application, the photoelectric signal converter 201 utilizes a photodetector to convert an input low-dimensional optical characteristic signal into an electrical analog signal. The intensity of the input optical characteristic signal is then converted into current signals at different ports of the electrical analog computing chip. Furthermore, the present application can perform matrix operations on the photoelectrically converted current signal through the electrical analog signal matrix operation element 202, ultimately outputting the analog calculation results through the electrical analog signal output element 203.

[0062] It should be noted that the electrical analog signal matrix operation element 202 of the embodiment of the present application uses an analog circuit to perform matrix operations on the current signal output by the photoelectric signal conversion element 201, obtains a voltage signal based on Kirchhoff's law, and outputs the voltage signal as the operation result. The parameters of the electrical analog signal matrix operation element 202 can be optimized by machine learning methods and can be changed manually.

[0063] Optionally, in some embodiments, the optical characteristic signal and the simulation calculation result satisfy the following relationship: i = g(l); (1) y = Bi; (2)

[0064] Wherein, i is the output current of the photoelectric signal conversion module, function g is the response characteristic of the photodetector in the photoelectric signal conversion module to the input light field, l is the low-dimensional optical characteristic signal light field input to the electrical analog computing system, y is the output voltage signal of the electrical analog computing system, B is the matrix parameter corresponding to the electrical analog signal matrix operation layer, and i is the output current of the photoelectric signal conversion module.

[0065] Furthermore, the optical simulation computing module 100 and the electrical simulation computing module 200 of the embodiment of the present application use joint training during the computer simulation process, and synchronously update parameters based on the gradient descent method. When using the optoelectronic fusion reconfigurable analog intelligent computing system 10 to complete the classification task, the cross entropy function can be used as the loss function.

[0066] For example, taking the classification task as an example, the cross entropy function of the embodiment of the present application as the loss function can be defined as: L = C (S (z), G); (3)

[0067] Among them, function C is the cross entropy function, function S is the Softmax function, z is the final output result of the optoelectronic fusion reconfigurable analog intelligent computing system, and G is the true value label of the classification task.

[0068] Thus, the present application can complete high-speed matrix operations through the electrical analog computing module 200, effectively avoiding the use of analog-to-digital converters during photoelectric conversion, and thus can complete machine learning tasks such as image recognition at high speed and low energy consumption. In addition, the parameters of the electrical analog signal matrix operation element 202 in the electrical analog computing module 200 can be manually adjusted, and the electrical analog computing module 200 can be selectively connected to the electrical digital computing module 300 for processing, achieving functional reconfiguration and capable of completing various machine learning tasks.

[0069] Optionally, in some embodiments, the electrical digital computing module 300 includes: an analog-to-digital signal conversion element 301 for converting an electrical analog signal into a digital electrical signal; an electrical digital signal operation element 302 for performing operations and processing on the digital electrical signal to obtain a digital calculation result; and an electrical digital signal output element 303 for outputting the digital calculation result.

[0070] It should be noted that each module in the optoelectronic fusion reconfigurable analog intelligent computing system 10 of the present application needs to be placed in a suitable position, and this position should be determined based on the initially set system parameters and training results.

[0071] Based on the above embodiments, the present application constructs an optoelectronic fusion reconfigurable analog intelligent computing system 10 based on the optical analog computing module 100 and the electrical analog computing module 200. The system is mathematically equivalent to a two-layer fully connected network and a nonlinear layer in the middle, and has the following form: y = Bg(Wx); (4)

[0072] Where y is the output voltage signal of the electrical analog calculation module, B is the matrix parameter corresponding to the electrical analog signal matrix operation component, function g is the response characteristic of the photodetector in the photoelectric signal conversion component to the input light field, W is the light field transformation matrix corresponding to the light through the free propagation process and the optical diffraction phase modulation component, and x is the input optical analog signal light field.

[0073] Based on the above embodiments, it can be understood that the optoelectronic fusion reconfigurable analog intelligent computing system proposed in this application can be applied to multiple experiments such as handwritten digit recognition, fashion item classification, and dynamic recognition of object movement direction. The system exhibits good performance and good robustness under low-light conditions. The computing speed of the system is three orders of magnitude higher than that of the most advanced digital computing processor, showing the huge advantages of the system in completing machine learning tasks.

[0074] In order to enable those skilled in the art to further understand the optoelectronic fusion reconfigurable analog intelligent computing system of the present application, the following examples are listed and the establishment and application process of the system are described in detail in conjunction with Figure 4.

[0075] (1) Establish a mathematical simulation model of the optoelectronic fusion reconfigurable analog intelligent computing system.

[0076] As shown in Figure 4, the present application can establish a mathematical simulation model of an optoelectronic fusion analog computing system based on hyperparameters such as the number of optical diffraction phase modulation layers, the spacing between each layer, and whether to add an electrical digital computing module and other system structure design schemes. Among them, the optical analog computing module and the electrical analog computing module together constitute an optoelectronic hybrid reconfigurable analog computing system, which is mathematically equivalent to two layers of fully connected networks and a nonlinear layer in the middle, as shown in the above formula (4).

[0077] It should be noted that the matrix W in formula (4) is determined by the free propagation part of the light and the modulated part of the light passing through the diffraction phase modulation layer; the free propagation part of the light includes the process of the light field from the input plane to the diffraction phase modulation layer, between the diffraction phase modulation layers, and from the diffraction phase modulation layer to the output light field layer. The propagation of light in all free spaces and uniform media is simulated using Fresnel propagation; the process of light passing through the diffraction phase modulation layer is modeled using phase modulation; the phase modulation size at different positions of the diffraction phase modulation layer is a parameter to be optimized.

[0078] Furthermore, the electrical analog signal matrix operation layer uses an analog circuit to perform matrix operations on the current signal output by the photoelectric signal conversion unit, obtains a voltage signal based on Kirchhoff's law, and outputs the voltage signal as the operation result; the parameters of the analog matrix operation layer can be optimized by machine learning methods and can be changed manually.

[0079] Therefore, after determining the specific machine learning task, the loss function of the corresponding task is defined.

[0080] (2) Use machine learning methods to optimize system parameters.

[0081] According to the machine learning task to be completed, the training set and test set are determined. Based on the training set and test set, the optoelectronic fusion simulation computing system is numerically simulated and trained through machine learning and error back propagation algorithm. During the training process, the parameters of the all-optical phase modulation layer and the optical simulation matrix operation layer are continuously optimized and updated. By debugging hyperparameters such as the latent space dimension, the arrangement pattern of the electro-optical conversion layer, and the number of phase modulation layers, the best optimization results are obtained.

[0082] (3) Based on the parameters obtained from simulation optimization, the phase modulation layer is physically manufactured using 3D printing or photolithography technology, the parameters of the electronic simulation calculation module are adjusted, and the system is built.

[0083] By using 3D printing or photolithography technology to manufacture the optical diffraction phase modulation layer, the optoelectronic fusion simulation computing system can be built and configured according to the preset or trained parameters. The parameters of the electrical simulation matrix operation layer can be fine-tuned through adaptive algorithms, and the data can be input into the system in the form of optical simulation signals to quickly and intelligently complete the corresponding machine learning tasks.

[0084] Therefore, this application constructs an optoelectronic fusion reconfigurable analog intelligent computing system based on the above three steps. The model in this system is different from the optical neural network processor or electronic digital processor in the related art. The model of this application integrates the optical diffraction network and the analog electronic computing chip, effectively avoiding the analog-to-digital conversion immediately after the optoelectronic conversion, and achieving high-speed, low-energy, and reconfigurable completion of machine learning tasks. In addition, it should be noted that the various parameters of the optoelectronic fusion reconfigurable analog intelligent computing system of this application can be obtained by establishing an optoelectronic hybrid simulation model and optimizing it using machine learning methods, and are not specifically limited here.

[0085] According to the optoelectronic fusion reconfigurable analog intelligent computing system proposed in the embodiment of the present application, the optical analog computing module is used to perform feature extraction and dimensionality reduction processing on the input optical analog signal to obtain an optical feature signal that meets the preset dimension, and the electrical analog computing module is used to convert the optical feature signal into an electrical analog signal, and calculate the electrical analog signal to obtain an analog computing result. And an electrical digital computing module can be selectively added to convert the input electrical analog signal into a digital signal, and further operations and signal processing are completed in the digital domain to obtain a digital computing result. Therefore, the present application uses an optical diffraction network to realize feature extraction and optical encoding of large-flux input analog light signals, and completes high-speed matrix operation tasks through analog electronic computing modules, creating a high-speed, low-energy, reconfigurable, and highly robust hardware implementation platform for machine learning tasks, and can be connected with existing electrical digital computing systems, with broad application value.

[0086] Next, a task learning method for an optoelectronic fusion reconfigurable analog intelligent computing system according to an embodiment of the present application will be described with reference to the accompanying drawings. The method adopts an optoelectronic fusion reconfigurable analog intelligent computing system 10 as in any of the above embodiments.

[0087] FIG5 is a flowchart of a task learning method of an optoelectronic fusion reconfigurable analog intelligent computing system according to an embodiment of the present application.

[0088] As shown in FIG5 , the task learning method of the optoelectronic fusion reconfigurable simulated intelligent computing system includes the following steps:

[0089] In step S501, a first numerical simulation model of an optical diffraction phase modulation layer and a second numerical simulation model of an electrical simulation matrix operation layer in an optoelectronic fusion reconfigurable analog intelligent computing system are established.

[0090] Specifically, this application needs to first determine the structure of the optoelectronic fusion reconfigurable analog intelligent computing system, and then establish a mathematical model of the system based on the structure of the system. The mathematical model includes numerical simulation models of the optical diffraction phase modulation layer and the electrical simulation matrix operation layer.

[0091] In step S502, the current task to be learned is obtained, and a training set is determined based on the current task to be learned; based on the training set, a preset back propagation algorithm is used to optimize the numerical simulation training of the optoelectronic fusion reconfigurable analog intelligent computing system until the training results meet the preset end conditions, and the parameters of the first numerical simulation model and the parameters of the second numerical simulation model are optimized according to the training results to obtain an optimized optoelectronic fusion reconfigurable analog intelligent computing system.

[0092] It can be understood that after determining the current task to be learned, this application needs to obtain the training set and test set to complete the learning task, and then conduct numerical simulation training on the optoelectronic fusion reconfigurable analog intelligent computing system through machine learning or error back propagation algorithm, and continuously optimize and update the parameters of the all-optical phase modulation layer and the optical simulation matrix operation layer during the training process.

[0093] Furthermore, the present application can use 3D printing or photolithography technology for processing to manufacture an optical diffraction phase modulation layer, build and configure an optoelectronic fusion reconfigurable analog intelligent computing system according to preset or trained parameters, and can fine-tune the parameters of the electrical simulation matrix operation layer through adaptive algorithms.

[0094] In step S503, the current task to be learned is completed according to the optimized optoelectronic fusion reconfigurable simulated intelligent computing system.

[0095] Therefore, the present application can input data into the system in the form of optical analog signals, and can quickly and intelligently complete the corresponding machine learning tasks.

[0096] In order to enable those skilled in the art to further understand the task learning method of the optoelectronic fusion reconfigurable analog intelligent computing system of the present application, the following examples are listed in conjunction with the accompanying drawings to schematically illustrate the implementation steps of the method.

[0097] Specifically, FIG6 is a flow chart of a method for completing a machine learning task using an optoelectronic fusion reconfigurable analog intelligent computing system according to a specific embodiment of the present application. As shown in FIG6 , the method includes the following steps:

[0098] Step S601, establishing an optoelectronic fusion simulation computing system model for implementing machine learning tasks;

[0099] Step S602, establishing numerical simulation models of the all-optical phase modulation layer and the electrical simulation calculation layer;

[0100] Step S603: Select appropriate training set, validation set, and test set according to the machine learning task;

[0101] Step S604, optimizing model parameters using a back propagation algorithm;

[0102] Step S605: manufacture and build an optoelectronic fusion simulation computing hardware system, and fine-tune system parameters through adaptive algorithms to achieve the machine learning task to be completed.

[0103] Therefore, this application can establish a photoelectric fusion reconfigurable analog intelligent computing system model, determine the training set according to the current learning task, use machine learning and other algorithms to perform numerical simulation training on the photoelectric fusion analog computing system, and continuously optimize parameters during the training process to ultimately complete the current learning task.

[0104] According to the task learning method of the optoelectronic fusion reconfigurable analog intelligent computing system proposed in the embodiment of the present application, by establishing a simulation model of the optical diffraction phase modulation layer and the electrical simulation matrix operation layer in the optoelectronic fusion reconfigurable analog intelligent computing system, according to the current task to be learned, the training set and the test set are determined, and the optoelectronic fusion reconfigurable analog intelligent computing system is numerically simulated and trained using the backpropagation algorithm until the training results meet the preset end conditions. The parameters of the simulation model are optimized according to the training results to obtain an optimized optoelectronic fusion reconfigurable analog intelligent computing system. According to the optimized optoelectronic fusion reconfigurable analog intelligent computing system, data is input to complete the current task to be learned. In this way, the limitations of all-optical artificial neural networks and digital computing processors in completing machine learning tasks are solved, and the advantages of high speed, low energy consumption, high robustness, and reconfigurability are achieved.

[0105] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0106] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this application, "N" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0107] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing a custom logical function or process step, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed in a different order than shown or discussed, including performing functions in a substantially simultaneous manner or in a reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application pertain.

[0108] It should be understood that various parts of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, it can be implemented using any one or a combination of the following technologies known in the art: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.

[0109] Those skilled in the art will appreciate that all or part of the steps in the method of the above-described embodiment may be accomplished by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0110] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. An optoelectronic fusion reconfigurable analog intelligent computing system, characterized in that, Including: An optical analog computing module and an electrical analog computing module, where the optical analog computing module is used to perform feature extraction and dimensionality reduction processing on the input optical analog signal to obtain an optical feature signal that meets the preset dimension; the electrical analog computing module is used to convert the optical feature signal into an electrical analog signal and perform calculations on the electrical analog signal to obtain an analog calculation result.

2. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 1, characterized in that The optical analog computing module includes: an optical analog signal input component for inputting the optical analog signal; an optical diffraction phase modulation component for performing feature extraction and dimensionality reduction on the optical analog signal to obtain an optical feature signal that meets the preset dimension; an optical feature signal output component for outputting the optical feature signal.

3. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 2, wherein The electrical analog computing module includes: an optoelectronic signal conversion component for converting the optical feature signal into the electrical analog signal by using a photodetector; an electrical analog signal matrix operation component for performing matrix operations on the electrical analog signal through a preset analog circuit and obtaining the analog calculation result based on Kirchhoff's law and the matrix operation result; an electrical analog signal output component for outputting the analog calculation result.

4. The optoelectronic fusion reconfigurable analog intelligent computing system according to any one of claims 1-3, characterized in that, It further includes: an electrical digital computing module for converting the electrical analog signal into a digital signal and performing processing in a preset digital domain to obtain a digital calculation result.

5. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 4, wherein The electrical digital computing module includes: an analog-to-digital signal conversion component for converting the electrical analog signal into a digital electrical signal; an electrical digital signal operation component for performing operations and processing on the digital electrical signal to obtain the digital calculation result; an electrical digital signal output component for outputting the digital calculation result.

6. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 5, wherein The optical analog signal is a coherent light field of input data information or a light field of a natural scene.

7. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 6, wherein The relationship satisfied by the optical analog signal and the optical feature signal is: l = Wx; where l is the output low-dimensional optical feature signal light field, W is the light field transformation matrix corresponding to the light passing through the free propagation process and the optical diffraction phase modulation component, and x is the input optical analog signal light field.

8. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 7, wherein The relationship satisfied by the optical feature signal and the analog calculation result is: i = g(l); y = Bi; where i is the output current of the optoelectronic signal conversion module, the function g is the response characteristic of the photodetector in the optoelectronic signal conversion component to the input light field, l is the low-dimensional optical feature signal light field input to the electrical analog computing module, y is the output voltage signal of the electrical analog computing module, B is the matrix parameter corresponding to the electrical analog signal matrix operation component, and i is the output current of the optoelectronic signal conversion module.

9. The optoelectronic fusion reconfigurable analog intelligent computing system according to claim 2, wherein The optical diffraction phase modulation component is physically manufactured by 3D printing or lithography technology.

10. A task learning method for an optoelectronic fusion reconfigurable analog intelligent computing system, which uses the optoelectronic fusion reconfigurable analog intelligent computing system described in any one of claims 1-9, wherein, The method includes the following steps: establishing a first numerical simulation model of the optical diffraction phase modulation layer and a second numerical simulation model of the electrical analog matrix operation layer in the optoelectronic fusion reconfigurable analog intelligent computing system; Obtain the current task to be learned, and determine the training set according to the current task to be learned; based on the training set, use the preset backpropagation algorithm to optimize the numerical simulation training of the optoelectronic fusion reconfigurable analog intelligent computing system until the training result meets the preset end condition, and optimize the parameters of the first numerical simulation model and the parameters of the second numerical simulation model according to the training result to obtain an optimized optoelectronic fusion reconfigurable analog intelligent computing system; and Complete the current task to be learned according to the optimized optoelectronic fusion reconfigurable analog intelligent computing system.

Images