Photon neural network device, photon neural network training method and chip
Through the design of photon neural network devices, the photoelectric hardware circuit is used to use the polarization and processing of optical signals to solve the problem of time -consuming and resource consumption of the existing neural network training process, and achieve efficient calculation and signal processing.
Patent Information
- Application Number
- CN202411472060.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-21
- Publication Date
- 2025-01-28
AI Technical Summary
The training process of existing neural networks requires a lot of time and computing resources, and computer-based methods have problems of inefficiency.
By using the polarization status of light to indicate the electrical signal, design a photon neural network device, and use the input layer, hidden layer and output layer of the optoelectronic hardware circuit structure to achieve polarizing and processing of optical signals.
It significantly improves the calculation speed of the neural network, saves computing resources and energy, and realizes low -power and high -speed computing and signal processing.
Smart Images
Figure CN119378627A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of photonic neural network technology, and in particular to a photonic neural network device, a photonic neural network training method and a chip. Background Art
[0002] With the development of artificial intelligence technology, the application of neural networks has entered all walks of life. However, the current implementation of neural networks is based on computer methods, which requires converting pictures or text into computer input signals, and then performing a large number of matrix multiplication and addition operations through multi-layer neural networks to realize the forward propagation of signals in different neural networks. The output layer is compared with the target to obtain the error, and then the weight network is modified layer by layer through the back propagation of the error. After multiple cycles, the result closest to the final target is achieved. This process requires a lot of time and computing resources. Summary of the invention
[0003] Based on this, it is necessary to provide a photonic neural network device, a photonic neural network training method and a chip that can use the polarization state of light to represent electrical signals and realize a neural network implemented by optoelectronic hardware circuits to address the above technical problems.
[0004] In a first aspect, the present application provides a photonic neural network device, comprising:
[0005] An input layer, comprising at least one input unit, for receiving multiple input electrical signals, performing optical polarization modulation, thereby converting the multiple input electrical signals into optical polarization codes, and outputting corresponding optical signals;
[0006] A hidden layer, connected to the input layer, comprising a plurality of hidden modules connected in sequence, wherein the hidden module comprises at least one hidden unit, connected to the corresponding input unit, and used to obtain different polarization states of each optical signal, perform polarization adjustment on the optical signal of each path, and output a corresponding optical polarization adjustment signal;
[0007] The output layer includes at least one output unit connected to the corresponding implicit unit and used to separate the light intensity of each optical polarization adjustment signal in the polarization direction to obtain multiple output electrical signals.
[0008] In one embodiment, the implicit unit includes:
[0009] A polarization detection module, connected to the corresponding input unit, for performing polarization splitting on the optical signal and determining an initial polarization state light intensity signal corresponding to the optical signal;
[0010] an amplification and addition module, connected to the polarization detection modules of all the implicit units, for amplifying and adding the initial polarization state light intensity signals outputted from the polarization detection modules to obtain an amplification control signal;
[0011] The polarization control module is connected to the amplification and addition module and the polarization detection module, and is used to adjust the polarization of the optical signal according to the amplification control signal and output a corresponding optical polarization adjustment signal.
[0012] In one embodiment, the polarization detection module includes:
[0013] A first beam splitter prism, connected to the corresponding input unit, and used to perform energy beam splitting on the optical signal to obtain an optical signal after energy beam splitting;
[0014] A first polarization beam splitter prism, connected to the first beam splitter, for performing polarization beam splitting on the optical signal after the energy beam splitting to obtain polarization beam splitting signals of light in mutually perpendicular directions;
[0015] The first detector is connected to the first polarization beam splitter prism and is used to perform effective beam splitting detection on the light polarization beam splitting signals in the mutually perpendicular directions and output an initial polarization state light intensity signal corresponding to the light signal, wherein the initial polarization state light intensity signal is an electrical signal.
[0016] In one embodiment, the amplification and addition module includes:
[0017] An amplifier, whose input end is connected to the polarization detection modules of all the implicit units, and is used to amplify the initial polarization state light intensity signal of each optical signal to obtain a corresponding amplified signal;
[0018] An adder is connected to the output end of the amplifier and is used to add the amplified signals to obtain the amplified control signal.
[0019] In one embodiment, the polarization control module comprises:
[0020] A polarization rotation controller, wherein a first input end is connected to the output end of the polarization detection module, and a second input end is connected to the output end of the amplification and addition module, and is used to control the amplification control signal to perform linear or nonlinear conversion to obtain a polarization rotation angle of the optical path; according to the polarization rotation angle of the optical path, the phase shift of the optical signal is adjusted to output the corresponding optical polarization adjustment signal.
[0021] In one embodiment, the input unit comprises:
[0022] The polarization encoder is used to receive the input electrical signal, perform optical polarization modulation according to the level state of the input electrical signal, and output the optical signal.
[0023] In one embodiment, the output unit includes:
[0024] A second polarization beam splitter prism is connected to the output end of the corresponding implicit unit and is used to perform polarization beam splitting on the optical polarization adjustment signal to obtain optical adjustment beam splitting signals in mutually perpendicular directions;
[0025] The second detector is connected to the second polarization splitter prism, and is used to perform effective beam splitting detection on the light adjustment beam splitting signals in the mutually perpendicular directions, and output the latest polarization state light intensity signal corresponding to the light polarization adjustment signal, wherein the latest polarization state light intensity signal is the output electrical signal.
[0026] In a second aspect, the present application further provides a photonic neural network training method, which is applied to the photonic neural network device as described in the first aspect above, and the method comprises:
[0027] Inputting multiple input electrical signals into the input layer for optical polarization encoding, and outputting corresponding optical signals;
[0028] Setting initial parameters of each hidden unit in the hidden layer, controlling the hidden layer to perform polarization adjustment on each optical signal according to different polarization states of the optical signals, and outputting a corresponding optical polarization adjustment signal;
[0029] The output layer is controlled to separate the light intensity of each of the optical polarization adjustment signals in the polarization direction to obtain multi-channel output electrical signals.
[0030] In one embodiment, the method further comprises:
[0031] Comparing the multiple output electrical signals with a preset electrical signal to obtain a comparison result;
[0032] According to the comparison result, the initial parameters of each hidden layer are corrected to obtain the updated parameters of each hidden layer.
[0033] In a third aspect, the present application further provides a chip, comprising the photonic neural network device as described in the first aspect above.
[0034] The above-mentioned photonic neural network device, photonic neural network training method and chip, through the input layer, hidden layer and output layer of the optoelectronic hardware circuit structure, convert the original training electrical signal into an optical signal, adjust the polarization of the optical signal, and realize the photonic neural network's low-power and high-parallel high-speed computing and signal processing, significantly improve the computing speed of the neural network, and save computing resources and energy. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0036] Figure 1 is a schematic diagram of the structure of a photonic neural network device in one embodiment;
[0037] Figure 2 is a schematic diagram of the structure of a hidden unit in an embodiment;
[0038] Figure 3 is a schematic structural diagram of a polarization detection module in one embodiment;
[0039] Figure 4 It is a structural schematic diagram of an amplified addition module in one embodiment;
[0040] Figure 5 is a schematic structural diagram of a polarization control module in one embodiment;
[0041] Figure 6 is a schematic diagram of the structure of an input unit in one embodiment;
[0042] Figure 7 FIG. 4 is a schematic diagram of the structure of an output unit in an embodiment. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0044] Unless otherwise defined, the technical terms or scientific terms involved in this application should be understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the" and the like involved in this application do not indicate a quantitative limitation, and may represent the singular or plural. The terms "include", "comprise", "have" and any of their variations involved in this application are intended to cover non-exclusive inclusions; for example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units that are not listed, or may also include other steps or units inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects before and after are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.
[0045] In an exemplary embodiment, Figure 1 As shown, a photonic neural network device is provided, comprising an input layer, a hidden layer and an output layer connected in sequence.
[0046] The input layer includes at least one input unit 1 for receiving multiple input electrical signals, performing optical polarization modulation on the multiple input electrical signals, thereby converting the multiple input electrical signals into optical polarization codes, and outputting corresponding optical signals.
[0047] The hidden layer includes a plurality of implicit modules connected in sequence. The number of implicit modules can be set according to the training requirements of the neural network. The number of optical paths between each implicit module can be different. Part of the optical path of each implicit module can be terminated and not propagated to the next implicit module. The implicit module includes at least one implicit unit 2, which is connected to the corresponding input unit 1, and is used to adjust the polarization of the optical signal of each path according to the different polarization states of each optical signal, and output the corresponding optical polarization adjustment signal.
[0048] The output layer includes at least one output unit 3, which is connected to the corresponding implicit unit 2 and is used to separate the light intensity of each optical polarization adjustment signal in the polarization direction to obtain multiple output electrical signals.
[0049] In the above-mentioned photonic neural network device, the input layer, hidden layer, and output layer constructed by the optoelectronic hardware circuit are used to convert the original electrical signal used for training into an optical signal, and the polarization of the optical signal is adjusted to achieve low-power, high-parallelism, high-speed computing and signal processing of the neural network, significantly improve the computing speed of the neural network, save computing resources and energy, and can also be applied to artificial intelligence computing to achieve ultra-low energy consumption and ultra-high-speed AI computing and training.
[0050] In an exemplary embodiment, Figure 2 As shown, the implicit unit 2 includes: a polarization detection module 21, an amplification and addition module 22 and a polarization control module 23. Among them:
[0051] The polarization detection module 21 is connected to the corresponding input unit 1 and is used to perform polarization splitting on the optical signal and determine an initial polarization state light intensity signal corresponding to the optical signal.
[0052] The amplification and addition module 22 is connected to the polarization detection modules 21 of all the implicit units, and is used to amplify and add the initial polarization state light intensity signals output from each polarization detection module 21 to obtain an amplification control signal.
[0053] The polarization control module 23 is connected to the amplification and addition module 22 and the polarization detection module 21, and is used to adjust the polarization of the optical signal according to the amplification control signal and output a corresponding optical polarization adjustment signal.
[0054] Specifically, when the polarization detection module 21 receives the optical signal output by the corresponding input unit 1, it performs energy splitting and polarization splitting on the optical signal, extracts the effective splitting of the optical signal, and thereby determines the initial polarization state light intensity signal of the optical signal. The initial polarization state light intensity signal output by each polarization detection module 21 is input as an electrical signal into each amplification and addition module 22, and the amplification and addition module 22 amplifies and sums all the initial polarization state light intensity signals passing through the same amplification and addition module 22 according to the set amplification factor and bias parameters, and outputs the amplification control signal to the polarization control module 23. At the same time, the polarization detection module 21 transmits the optical signal after energy splitting to the polarization control module 23, and the polarization control module 23 uses the amplification control signal to adjust the polarization of the optical signal after energy splitting, and outputs the corresponding optical polarization adjustment signal.
[0055] In this embodiment, the polarization state of the optical signal is adjusted through the polarization detection module 21, the amplification and addition module 22 and the polarization control module 23, thereby realizing low-latency and flexible processing of large-scale data signals, and its calculation speed far exceeds that of traditional computer neural networks. Compared with traditional computer neural networks, it can be miniaturized and is suitable for various embedded systems.
[0056] In one embodiment, if Figure 3 As shown, the polarization detection module 21 includes a first beam splitter prism 211 , a first polarization beam splitter prism 212 and a first detector 213 .
[0057] The first beam splitter prism 211 is connected to the corresponding input unit 1 and is used to perform energy splitting on the optical signal to obtain an optical signal after energy splitting.
[0058] The first polarization beam splitter prism 212 is connected to the first beam splitter 211 and is used to perform polarization beam splitting on the energy-split optical signal to obtain polarization beam splitting signals in mutually perpendicular directions.
[0059] The first detector 213 is connected to the first polarization beam splitting prism 212, and is used to perform effective beam splitting detection on the optical polarization beam splitting signals in the mutually perpendicular directions, and output an initial polarization state light intensity signal corresponding to the optical signal, wherein the initial polarization state light intensity signal is an electrical signal.
[0060] In detail, after the optical signal enters the first beam splitter 211, it is split according to the energy of the light, and the energy-split optical signal enters the first polarization beam splitter prism 212, and then the first polarization beam splitter prism 212 performs polarization beam splitting on the energy-split optical signal to obtain polarization beam splitting signals in mutually perpendicular directions. Assuming that after polarization beam splitting, the polarization beam splitting signal representing binary 0 is in the horizontal direction, and the polarization beam splitting signal representing binary 1 is in the vertical direction, then the vertical beam splitting signal is the effective part of the optical signal, and after the effective beam splitting in the polarization beam splitting signal is detected by the first detector 213, it is converted into an initial polarization state light intensity signal corresponding to the optical signal, wherein the initial polarization state light intensity signal is an electrical signal.
[0061] In this embodiment, high-precision beam splitting detection of effective light intensity signals is achieved through the first beam splitter prism 211, the first polarization beam splitter prism 212 and the first detector 213, thereby improving the accuracy of polarization state detection of the photonic neural network and providing a reliable basis for subsequent signal processing and control.
[0062] In one embodiment, if Figure 4 As shown, the amplifying and adding module 22 includes an amplifier 221 and an adder 222 .
[0063] The input end of the amplifier 221 is connected to the polarization detection modules 21 of all the implicit units, and is used to amplify the initial polarization state light intensity signal of each optical signal to obtain a corresponding amplified signal.
[0064] The adder 222 is connected to the output end of the amplifier 221 and is used to add the amplified signals to obtain the amplified control signal.
[0065] In detail, the input end of the amplifier 221 in each amplification and addition module 22 is connected to the first detectors 213 of all polarization detection modules 21, receives the initial polarization state light intensity signals output by all the first detectors 213, and amplifies each initial polarization state light intensity signal according to the amplification factor and bias parameters set by the amplifier 221, and inputs the corresponding generated amplified signals into the adder 222 for addition and summation, so as to obtain the amplification control signal for adjusting the optical signal of the optical path input unit 1.
[0066] In this embodiment, through the amplifier 221 and the adder 222 in each amplification and addition module, efficient and parallel amplification control signal calculation is achieved, thereby improving the calculation rate of the photonic neural network.
[0067] In one embodiment, if Figure 5 As shown, the polarization control module 23 includes a polarization rotation controller 231 .
[0068] The first input end of the polarization rotation controller 231 is connected to the output end of the polarization detection module 21, and the second input end is connected to the output end of the amplification and addition module 22, and is used to control the amplification control signal to perform linear, semi-linear or nonlinear conversion to obtain the polarization rotation angle of the optical path; according to the polarization rotation angle of the optical path, adjust the phase shift of the optical signal and output the corresponding optical polarization adjustment signal.
[0069] In detail, the polarization rotation controller 231 is composed of an adjustable optical phase shifter and a quarter wave plate. The adjustable optical phase shifter controls the phase shift of the optical signal through different voltages, and can modulate the light to any angle to achieve the rotation of the polarization angle of the light. The quarter wave plate is used to convert the polarization state of the light into the corresponding optical path polarization rotation angle.
[0070] Specifically, the linear, semi-linear or non-linear conversion formula is:
[0071] θ=f(∑Wx+b);
[0072] Among them, θ represents the polarization rotation angle of the optical path; f can be a linear, semi-linear or nonlinear function, such as Sigmoid, tanh, Relu, maxout, etc.; W represents the amplification factor of the amplifier; b represents the bias parameter; x represents the light intensity signal.
[0073] In this embodiment, the polarization angle of light is controlled with high precision through the polarization rotation controller 231, thereby achieving efficient neural network calculation.
[0074] In one embodiment, Figure 6 As shown, the input unit 1 includes a polarization encoder 101, which is used to receive the input electrical signal, and perform polarization modulation of light according to the level state of the input electrical signal to obtain the optical signal.
[0075] In detail, the polarization encoder 101 performs binary optical polarization encoding according to the high and low level states of the input electrical signal waveform to obtain an optical signal represented in binary form.
[0076] In this embodiment, polarization encoding of electrical signals into optical signals is implemented by the polarization encoder 101, so that other circuit modules of the subsequent photonic neural network can use the optical signal to perform efficient signal processing.
[0077] In one embodiment, Figure 7 As shown, the output unit 3 includes a second polarization beam splitter prism 301 and a second detector 302 .
[0078] The second polarization beam splitter prism 301 is connected to the output end of the corresponding implicit unit 2 and is used to perform polarization beam splitting on the optical polarization adjustment signal to obtain optical adjustment beam splitting signals in mutually perpendicular directions.
[0079] The second detector 302 is connected to the second polarization splitter prism 301, and is used to perform effective beam splitting detection on the light adjustment beam splitting signals in the mutually perpendicular directions, and output the latest polarization state light intensity signal corresponding to the light polarization adjustment signal, wherein the latest polarization state light intensity signal is the output electrical signal.
[0080] In this embodiment, the effective splitting of the optical polarization adjustment signal is extracted through the second polarization beam splitter prism 301 and the second detector 302, thereby improving the accuracy and speed of the final output electrical signal.
[0081] In one embodiment, a photonic neural network training method is provided, which is applied to the photonic neural network device as described in 1, wherein each input unit of the input layer includes a polarization encoder 101, each hidden unit of the hidden layer includes a first beam splitter prism 211, a first polarization beam splitter prism 212, a first detector 213, an amplifier 221, an adder 222, and a polarization rotation controller 231, and each output unit of the output layer includes a second polarization beam splitter prism 301 and a second detector 302. The functions and connection relationships of each module are as described in the above embodiments, and the method specifically includes the following steps 801 to 805.
[0082] Step 801: Input multiple input electrical signals into an input layer for optical polarization encoding, and output corresponding optical signals.
[0083] Specifically, taking the image data set as a training set as an example, the image data is split according to each row or column of pixels to obtain input electrical signals corresponding to multiple channels of image pixel data. Each input electrical signal is input into the corresponding input unit of the input layer, and the polarization encoder 101 is used to perform optical polarization encoding on the input electrical signal to output the corresponding optical signal.
[0084] Step 802, setting the initial parameters of each hidden unit in the hidden layer, controlling the hidden layer to perform polarization adjustment on each optical signal according to different polarization states of each optical signal, and outputting a corresponding optical polarization adjustment signal.
[0085] Specifically, according to the accuracy requirements for processing image data, the number of implicit modules connected in sequence in the hidden layer and the initial parameters of each implicit unit in each implicit module are set. When the optical signals output by the input layer enter the hidden layer, the initial parameters of each implicit unit are used to perform polarization adjustment according to the different polarization states of the input optical signals, and finally the optical polarization adjustment signal corresponding to each optical signal is output.
[0086] Step 803: Control the output layer to perform light intensity separation on each of the optical polarization adjustment signals in a polarization direction to obtain multi-channel output electrical signals.
[0087] Specifically, the optical polarization adjustment signal corresponding to each optical signal passes through the corresponding second polarization beam splitter prism 301 and the second detector 302 in the output layer to separate the light intensity in the polarization direction to obtain multiple output electrical signals.
[0088] Step 804: compare the multi-channel output electrical signal with a preset electrical signal to obtain a comparison result.
[0089] Specifically, the multi-channel output electrical signals are compared with the preset electrical signals of the image data, and the difference between the two is calculated.
[0090] Step 805: According to the comparison result, the initial parameters of the hidden layer are modified to obtain updated parameters of the hidden layer.
[0091] Specifically, the change amount required for back propagation is calculated based on the difference between the two, and the amplification factor and bias parameters of the amplifier 221 in the hidden unit are modified layer by layer.
[0092] Repeat the above steps 801 to 805 until the convergence result is correct. And train all data sets. After all data sets are trained, all the final magnifications and bias parameters are the final model parameters of the photonic neural network.
[0093] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.
[0094] In one embodiment, a chip is provided, the chip comprising the photonic neural network device as described in the above embodiments. The chip can implement the functions described in the photonic neural network device as described in the above embodiments.
[0095] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.
[0096] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0097] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present application. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.
Claims
1. A photonic neural network device, characterized in that: include: An input layer, comprising at least one input unit, for receiving multiple input electrical signals, performing optical polarization modulation, thereby converting the multiple input electrical signals into optical polarization codes, and outputting corresponding optical signals; A hidden layer, connected to the input layer, comprising a plurality of hidden modules connected in sequence, wherein the hidden module comprises at least one hidden unit, connected to the corresponding input unit, and used to obtain different polarization states of each optical signal, perform polarization adjustment on the optical signal of each path, and output a corresponding optical polarization adjustment signal; The output layer includes at least one output unit connected to the corresponding implicit unit and used to separate the light intensity of each optical polarization adjustment signal in the polarization direction to obtain multiple output electrical signals.
2. The photonic neural network device according to claim 1, characterized in that: The hidden unit comprises: A polarization detection module, connected to the corresponding input unit, for performing polarization splitting on the optical signal and determining an initial polarization state light intensity signal corresponding to the optical signal; an amplification and addition module, connected to the polarization detection modules of all the implicit units, and used to amplify and add the initial polarization state light intensity signals outputted from the polarization detection modules to obtain an amplification control signal; The polarization control module is connected to the amplification and addition module and the polarization detection module, and is used to adjust the polarization of the optical signal according to the amplification control signal and output a corresponding optical polarization adjustment signal.
3. The photonic neural network device according to claim 2, characterized in that: The polarization detection module comprises: A first beam splitter prism, connected to the corresponding input unit, and used to perform energy beam splitting on the optical signal to obtain an optical signal after energy beam splitting; A first polarization beam splitter prism, connected to the first beam splitter, for performing polarization beam splitting on the optical signal after the energy beam splitting to obtain polarization beam splitting signals of light in mutually perpendicular directions; The first detector is connected to the first polarization beam splitter prism and is used to perform effective beam splitting detection on the light polarization beam splitting signals in the mutually perpendicular directions and output an initial polarization state light intensity signal corresponding to the light signal, wherein the initial polarization state light intensity signal is an electrical signal.
4. The photonic neural network device according to claim 3, characterized in that: The amplification and addition module comprises: An amplifier, whose input end is connected to the polarization detection modules of all the implicit units, and is used to amplify the initial polarization state light intensity signal of each optical signal to obtain a corresponding amplified signal; An adder is connected to the output end of the amplifier and is used to add the amplified signals to obtain the amplified control signal.
5. The photonic neural network device according to claim 4, characterized in that: The polarization control module comprises: A polarization rotation controller, wherein a first input end is connected to the output end of the polarization detection module, and a second input end is connected to the output end of the amplification and addition module, and is used to control the amplification control signal to perform linear or nonlinear conversion to obtain a polarization rotation angle of the optical path; according to the polarization rotation angle of the optical path, the phase shift of the optical signal is adjusted to output the corresponding optical polarization adjustment signal.
6. The photonic neural network device according to claim 1, characterized in that: The input unit comprises: The polarization encoder is used to receive the input electrical signal, perform optical polarization modulation according to the level state of the input electrical signal, and output the optical signal.
7. The photonic neural network device according to claim 1, characterized in that: The output unit comprises: A second polarization beam splitter prism is connected to the output end of the corresponding implicit unit and is used to perform polarization beam splitting on the optical polarization adjustment signal to obtain optical adjustment beam splitting signals in mutually perpendicular directions; The second detector is connected to the second polarization splitter prism, and is used to perform effective beam splitting detection on the light adjustment beam splitting signals in the mutually perpendicular directions, and output the latest polarization state light intensity signal corresponding to the light polarization adjustment signal, wherein the latest polarization state light intensity signal is the output electrical signal.
8. A photonic neural network training method, applied to the photonic neural network device according to any one of claims 1 to 5, characterized in that: The method comprises: Inputting multiple input electrical signals into the input layer for optical polarization encoding, and outputting corresponding optical signals; Setting initial parameters of each hidden unit in the hidden layer, controlling the hidden layer to perform polarization adjustment on each optical signal according to different polarization states of the optical signals, and outputting a corresponding optical polarization adjustment signal; The output layer is controlled to separate the light intensity of each of the optical polarization adjustment signals in the polarization direction to obtain multi-channel output electrical signals.
9. The photonic neural network training method according to claim 8, characterized in that: The method further comprises: Comparing the multiple output electrical signals with a preset electrical signal to obtain a comparison result; According to the comparison result, the initial parameters of each hidden layer are corrected to obtain the updated parameters of each hidden layer.
10. A chip, characterized in that: A photonic neural network device comprising any one of claims 1 to 7.