Photoelectric signal processing device, optical neural network model training method and optical computing equipment

Abstract

The invention discloses a photoelectric signal processing device, an optical neural network model training method and optical computing equipment, and relates to the technical field of optical signal processing. The optical signal processing device comprises at least two optical signal processing layers, at least one electric signal transmission layer and an electric signal control layer; the optical signal processing layers are arranged on the device substrate, the optical signal processing layers and the electric signal transmission layers are arranged at intervals, and the electric signal control layer is arranged on the optical signal processing layer located on the top layer. By performing three-dimensional stacking on the optical signal processing layers, the expandability of the optical calculation device is improved. Through three-dimensional stacking of the optical signal processing layers, the overall number of the optical processing units in the optical calculation device is increased, the optical calculation device has the capacity of processing a large number of data tasks, when tasks with large data volumes are processed, the intensity of a laser light source received by each optical signal processing layer is weak, and the light processing efficiency is improved. Therefore, the silicon optical material is prevented from being damaged due to strong light irradiation.

Description

Technical Field

[0001] The present invention relates to the technical field of optical signal processing, and particularly relates to an optoelectronic signal processing device, a method for training an optical neural network model, and an optical computing device. Background Art

[0002] An optical computing device is a new type of computing device based on the principle of photonics, using optical signals as information carriers. Since optical signals are hardly affected by electromagnetic interference, they have high stability and anti-interference ability; since optical signals propagate at high speed, optical computing devices have high-speed data transmission and processing capabilities. They are suitable for processing parallel computing tasks with high complexity and large data volumes.

[0003] However, when processing tasks with a large amount of data, a high-intensity laser light source is required. Irradiating a silicon photonic material with a high-intensity laser light source will cause damage to the silicon photonic material. Therefore, it is difficult for optical computing devices to be extended into a large-scale computing system, and the scalability is poor. Summary of the Invention

[0004] The present application provides an optoelectronic signal processing device, a method for training an optical neural network model, and an optical computing device, which at least solve the problem that it is difficult for optical computing devices to be scaled up on a large scale.

[0005] In a first aspect, the present application provides an optical signal processing device, including: At least two layers of optical signal processing layers, at least one layer of electrical signal transmission layers, and an electrical signal control layer; The optical signal processing layers are disposed on a device substrate, the optical signal processing layers and the electrical signal transmission layers are disposed at intervals, and the electrical signal control layer is disposed on the optical signal processing layer at the top layer; The electrical signal control layer is configured to generate an electrical signal for modulating the parameters of the optical signal processing units in the first layer of optical signal processing layer; The first layer of optical signal processing layer is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signal generated by the electrical signal control layer, determine the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generate an electrical signal for adjusting the second layer of optical signal processing layer; The Nth layer of optical signal processing layer is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signal generated by the (N - 1)th layer of optical signal processing layer, determine the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generate an electrical signal for modulating the (N + 1)th layer of optical signal processing layer, where N is an integer greater than or equal to 2, representing the number of layers of the optical signal processing layers; The electrical signal transmission layer is disposed between the Nth optical signal processing layer and the (N + 1)th optical signal processing layer, and is used to transmit the electrical signal generated by the Nth optical signal processing layer to the (N + 1)th optical signal processing layer through the through-silicon vias disposed in the electrical signal transmission layer.

[0006] In a second aspect, a method for training an optical neural network model is provided, which is applied to the optical signal processing device described in the first aspect, and includes: Setting the initial values of the parameters of at least one optical signal processing unit corresponding to the model to be trained; Creating a loss function for the model to be trained; Adjusting the values of the parameters of at least one optical signal processing unit corresponding to the model to be trained by using the finite difference method until the convergence condition is reached, so as to determine the final values of the optical signal processing unit parameters.

[0007] In a third aspect, an optical computing device is provided, which includes the optical signal processing device described in the first aspect and a memory.

[0008] The beneficial effects brought by the technical solutions provided in the embodiments of the present application are as follows: By three-dimensionally stacking the optical signal processing layers, the scalability of the optical computing device is improved. Through the three-dimensionally stacking of the optical signal processing layers, the overall number of optical processing units in the optical computing device is increased, enabling the optical computing device to have the ability to process a large amount of data tasks. When processing tasks with a large amount of data, the intensity of the laser light source received by each optical signal processing layer is weak, so as to avoid damage to the silicon photonics material due to strong light irradiation. Description of the Drawings

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0010] Figure 1 It is a schematic structural diagram of an optical signal processing device provided by an embodiment of the present application; Figure 2 It is a schematic diagram of an optical signal processing layer provided by an embodiment of the present application; Figure 3 It is a schematic diagram of a photoelectric conversion module provided by an embodiment of the present application; Figure 4 It is a schematic diagram of an interferometer network provided by an embodiment of the present application; Figure 5 It is a schematic diagram of a Mach-Zehnder interferometer provided by an embodiment of the present application; Figure 6It is a schematic diagram of the optical signal processing layer composed of an interferometer module provided by an embodiment of the present application; Figure 7 It is a schematic diagram of a microring resonator provided by an embodiment of the present application; Figure 8 It is a schematic diagram of the optical signal processing layer processing an optical signal provided by an embodiment of the present application; Figure 9 It is a schematic diagram of the electrical signal transmission provided by an embodiment of the present application; Figure 10 It is a schematic diagram of the three-dimensional stacked structure of the multi-layer optical signal processing layer provided by an embodiment of the present application; Figure 11 It is a schematic diagram of a method for training an optical neural network model provided by an embodiment of the present application; Figure 12 It is a schematic diagram of an optical computing device provided by an embodiment of the present application. Detailed implementation manners

[0011] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present application.

[0012] Unless otherwise defined, the technical terms or scientific terms used in this disclosure shall have the ordinary meaning understood by those of ordinary skill in the art to which this disclosure belongs. The "first", "second", and similar terms used in this disclosure do not denote any order, quantity, or importance, but are only used to distinguish different components. Similarly, the terms such as "a", "an", or "the" do not denote a quantity limitation, but mean that there is at least one. The numbers in the accompanying drawings of the specification only represent the distinction between each functional component or module, and do not represent the logical relationship between the components or modules. The term "including" or "comprising" and similar terms mean that the elements or objects appearing before this term cover the elements or objects listed after this term and their equivalents, without excluding other elements or objects. The term "connected" or "coupled" and similar terms are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0013] Next, various embodiments according to the present disclosure will be described in detail with reference to the accompanying drawings. It should be noted that in the drawings, the same reference numerals are assigned to components having substantially the same or similar structures and functions, and repeated descriptions thereof will be omitted.

[0014] Regarding the problem that optical computing devices are difficult to scale up in the prior art, the present application provides the following embodiments.

[0015] In some embodiments, as Figure 1 shown, an optical signal processing device includes: At least two layers of optical signal processing layers C a , where a = 1, 2,..., N, at least one layer of electrical signal transmission layer T, and an electrical signal control layer SC; The optical signal processing layer is disposed on the device substrate, the optical signal processing layer and the electrical signal transmission layer are spaced apart, and the electrical signal control layer is disposed on the topmost optical signal processing layer.

[0016] Among them, the device substrate includes a silicon (Si) substrate and a silicon dioxide substrate (SiO2 substrate), the silicon dioxide substrate is disposed on the silicon substrate, and the Nth layer of optical signal processing layer is disposed on the silicon dioxide substrate.

[0017] The electrical signal control layer is provided with through-silicon vias (TSVs) connecting the upper and lower surfaces of the electrical signal transmission layer in this layer, and is used to generate an electrical signal for modulating the parameters of the optical signal processing units in the first layer of optical signal processing layer; The first layer of optical signal processing layer is used to receive and modulate the parameters of the optical signal processing units in this layer according to the electrical signal generated by the electrical signal control layer, determine the optical information of this layer according to the light source and the parameters of the optical signal processing units in this layer, and generate an electrical signal for adjusting the second layer of optical signal processing layer; The Nth layer of optical signal processing layer is used to receive and modulate the parameters of the optical signal processing units in this layer according to the electrical signal generated by the (N - 1)th layer of optical signal processing layer, determine the optical information of this layer according to the light source and the parameters of the optical signal processing units in this layer, and generate an electrical signal for modulating the (N + 1)th layer of optical signal processing layer, where N is an integer greater than or equal to 2, representing the number of layers of the optical signal processing layer; The electrical signal transmission layer is disposed between the Nth layer of optical signal processing layer and the (N + 1)th layer of optical signal processing layer, and is used to transmit the electrical signal generated by the Nth layer of optical signal processing layer to the (N + 1)th layer of optical signal processing layer through the through-silicon vias disposed in the electrical signal transmission layer.

[0018] For example: An optical signal processing device generally includes two optical signal processing layers C1 and C2. The first optical signal processing layer C1 is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the electrical signal control layer, determine the optical information of this optical signal processing layer based on the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generate electrical signals for adjusting the second optical signal processing layer. These electrical signals are transmitted to the second optical signal processing layer through the through-silicon vias in the optical signal transmission layer between the first optical signal processing layer and the second optical signal processing layer. The second optical signal processing layer C2 is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the first optical signal processing layer, and determine the optical information of this optical signal processing layer based on the light source and the parameters of the optical signal processing units in this optical signal processing layer. Since there are only two optical signal processing layers, although the second optical signal processing layer can generate electrical signals, these electrical signals are no longer transmitted.

[0019] Another example: An optical signal processing device generally includes three optical signal processing layers C1, C2, and C3. The first optical signal processing layer is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the electrical signal control layer, determine the optical information of this optical signal processing layer based on the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generate electrical signals for adjusting the second optical signal processing layer. The electrical signals generated by the first optical signal processing layer are transmitted to the second optical signal processing layer through the through-silicon vias in the optical signal transmission layer between the first optical signal processing layer and the second optical signal processing layer. The second optical signal processing layer C2 is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the first optical signal processing layer, determine the optical information of this optical signal processing layer based on the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generate electrical signals for modulating the third optical signal processing layer. The electrical signals generated by the second optical signal processing layer are transmitted to the third optical signal processing layer through the through-silicon vias in the optical signal transmission layer between the second optical signal processing layer and the third optical signal processing layer. The third optical signal processing layer C3 is configured to receive and modulate the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the second optical signal processing layer, and determine the optical information of this optical signal processing layer based on the light source and the parameters of the optical signal processing units in this optical signal processing layer. Since there are only three optical signal processing layers, although the third optical signal processing layer can generate electrical signals, these electrical signals are no longer transmitted.

[0020] Specifically, as Figure 2As shown, the optical signal processing layer is provided with: a light source LS, an electro-optic modulation module EOM, a wavelength division multiplexer MUX, at least one optical processing branch LP, a demultiplexer DMUX, a photoelectric conversion module PD, and a photodetector PED; The electro-optic modulation module is connected to the wavelength division multiplexer, the wavelength division multiplexer is connected to at least one optical processing branch, each of the at least one optical processing branch is correspondingly connected to a demultiplexer, the demultiplexer is correspondingly connected to the photoelectric conversion module, and each of the at least one optical processing branch is also correspondingly connected to the photodetector; The light source is used to generate a laser with a preset wavelength; The electro-optic modulation module is used to modulate the light intensity of the laser; The wavelength division multiplexer is used to simultaneously transmit optical signals of at least one wavelength; The optical processing branch is used to determine the optical signal corresponding to this optical processing branch according to the light source; The demultiplexer is used to decompose and obtain an optical signal of a single wavelength; The photoelectric conversion module is used to generate and send an electrical signal for modulating the adjacent lower-layer optical signal processing layer; The photodetector is used to detect the optical signal generated by the corresponding optical processing branch.

[0021] Preferably, the modulator is: a lithium niobate electro-optic modulator, a polymer electro-optic modulator, or a silicon-based integrated electro-optic modulator.

[0022] Specifically, as Figure 3 shown, the photoelectric conversion module includes: a photosensitive element, an operational amplifier, a resistor, and a capacitor; The electrical signal output terminal of the photosensitive element is connected to the inverting input terminal of the operational amplifier, both ends of the resistor R are connected in parallel to the inverting input terminal and the output terminal of the operational amplifier, the capacitor C is connected in parallel to the inverting input terminal "-" and the output terminal OP of the operational amplifier o , the non-inverting input terminal "+" of the operational amplifier is grounded. The electrical signal input terminal of the photosensitive element is connected to the voltage V p . The photosensitive element receives the input light source.

[0023] Preferably, the wavelength division multiplexer is an arrayed waveguide grating.

[0024] Preferably, the demultiplexer is a nonlinear Bragg grating.

[0025] Preferably, the light source is a laser.

[0026] Preferably, the laser is an InGaAsP laser or an InP laser.

[0027] Optionally, the wavelength range of the laser generated by the laser is 1300 nm to 2200 nm.

[0028] InGaAsP (Indium Gallium Arsenide Phosphide) lasers and InP (Indium Phosphide) lasers are commonly used lasers in the field of optical fiber communication. The bandgap energy of the material can be adjusted by changing the alloy composition, thereby achieving laser output at different wavelengths. Typically, the typical wavelength of INGaAsP lasers is from 1300 nm to 1550 nm. The wavelength generated by InP lasers can be extended to 2200 nm.

[0029] As Figure 2 shown, any optical processing branch has: an optical branch input end LP1 and an optical branch output end LP2; The optical branch input end is connected to the output end of the wavelength division multiplexer, and the optical branch output end is connected to the input end of the corresponding demultiplexer; Any optical processing branch includes at least one optical processing unit LU; The optical processing unit has: an optical input port LU1 and an optical output port LU2; In any optical processing branch, the optical output end of the previous-stage optical processing unit is cascaded with the optical input end of the next-stage optical processing unit. The optical input end of the first optical processing unit serves as the optical branch input end, and the optical output end of the last optical processing unit serves as the optical branch output end.

[0030] Optionally, the optical processing unit is an interferometer module; The interferometer module has an interferometer module input end and an interferometer module output end; The interferometer module input end serves as the optical input port, and the interferometer module output end serves as the optical output port.

[0031] In this embodiment, the optical processing unit in any optical signal processing layer of the optical signal processing device is an interferometer module.

[0032] In some embodiments, as Figure 4 shown, the interferometer module is an interferometer network; The interferometer network has: a first network input end IW11, a second network input end IW12, a third network input end IW13, a first network output end IW21, a second network output end IW22, and a third network output end IW23; The interferometer network includes: a first interferometer I1, a second interferometer I2, and a third interferometer I3; The first interferometer has: a first input end of the first interferometer, a second input end of the first interferometer, a first output end of the first interferometer, and a second output end of the first interferometer; The second interferometer has: a first input end of the second interferometer, a second input end of the second interferometer, a first output end of the second interferometer, and a second output end of the second interferometer; The third interferometer has: a first input end of the third interferometer, a second input end of the third interferometer, a first output end of the third interferometer, and a second output end of the third interferometer; The first input end of the first interferometer, the second input end of the first interferometer, and the first input end of the second interferometer serve as the input ends of the interferometer module. The first output end of the first interferometer is connected to the second input end of the second interferometer. The second output end of the first interferometer is connected to the second input end of the third interferometer. The second output end of the second interferometer is connected to the first output end of the third interferometer. The second output end of the second interferometer, the first output end of the third interferometer, and the second output end of the third interferometer serve as the output ends of the interferometer module.

[0033] Specifically, the first interferometer, the second interferometer, and the third interferometer are Mach-Zehnder interferometers.

[0034] In some other embodiments, as Figure 5 shown, the interferometer module is a single Mach-Zehnder interferometer; The Mach-Zehnder interferometer has: a first input port In1, a second input port In2, a first output port Out1, and a second output port Out2; The first input port and the second input port serve as the input ends of the interferometer module, and the first output port and the second output port serve as the output ends of the interferometer module; Further, the Mach-Zehnder interferometer includes: a first coupler, a second coupler, a first phase shifter, and a second phase shifter; One end of the first coupler serves as the first input port and the second input port, or as the first input end of the first interferometer and the second input end of the first interferometer, or as the first input end of the second interferometer and the second input end of the second interferometer, or as the first input end of the third interferometer and the second input end of the third interferometer; The other end of the first coupler is connected to one end of the first phase shifter through a waveguide. The other end of the first phase shifter is connected to one end of the second coupler through a waveguide. The other end of the first coupler is also connected to one end of the second coupler through a waveguide. The other end of the second coupler is connected to the second phase shifter through a waveguide; The other end of the second phase shifter serves as the first output port; The other end of the second coupler serves as the second output port.

[0035] The Mach-Zehnder interferometer (MZI) is a commonly used silicon photonics device and is also a natural minimum matrix core, as Figure 5As shown, the matrix network constructed by cascading it according to specific rules can perform arbitrary matrix multiplication. The Mach-Zehnder interferometer consists of two couplers and two sets of interference arms, and the interference arms are waveguides through which light propagates. One of each set of interference arms is equipped with a tunable phase shifter, which provides phase shifts of θ and Φ respectively. The black rectangles are couplers, and the white rectangles are tunable phase shifters. A single Mach-Zehnder interferometer has two inputs and two outputs, and the input-output relationship can be represented by a unitary rotation matrix R ( θ, Φ ): ; ; By modulating the phases θ and Φ of the tunable phase shifters, R ( θ, Φ ) can be changed to achieve an arbitrary unitary rotation matrix. This phase modulation is achieved through an external electrical signal. Currently, the commonly used tunable phase shifters are electro-optic phase shifters and thermo-optic phase shifters, that is, by adjusting the input voltage or the temperature of the device through an electrical signal to achieve the adjustment of the phases θ and Φ . The input optical signal outputs from two ports with the power allocated by this matrix, and the measurement can obtain the calculation result.

[0036] According to the triangular decomposition algorithm, any n×n unitary matrix can theoretically be decomposed into the continuous multiplication of n(n - 1) / 2 unitary rotation matrices. For example, for a 3×3 Unitary matrix, the decomposition follows U (3×3) = U 3 U 2 U 1 , where U 1 , U 2 , U 3 has the following form: ; Here R i = R ( θ 𝑖 , Φ 𝑖 ), and each θ 𝑖 and Φ 𝑖It can be modulated separately. The maximum executable unitary matrix is judged according to the number of input and output of the triangular network. A 3×3 unitary matrix requires 3 pairs of input and output to form the interferometer network described above. Based on the optical interference calculation of the Mach-Zehnder interferometer, it has strong adjustability and can realize any calculation process.

[0037] The Mach-Zehnder interferometer includes: a first coupler, a second coupler, a first phase shifter, and a second phase shifter; One end of the first coupler serves as the first input end and the second input end of the first interferometer, or as the first input end and the second input end of the second interferometer, or as the first input end and the second input end of the third interferometer; The other end of the first coupler is connected to one end of the first phase shifter through a waveguide. The other end of the first phase shifter is connected to one end of the second coupler through a waveguide. The other end of the first coupler is also connected to one end of the second coupler through a waveguide. The other end of the second coupler is connected to the second phase shifter through a waveguide; The other end of the second phase shifter serves as the first output end of the first interferometer, or as the first output end of the second interferometer, or as the first output end of the third interferometer; The other end of the second coupler serves as the second output end of the first interferometer, or as the second output end of the second interferometer, or as the second output end of the third interferometer.

[0038] Figure 6 It shows a layer of optical signal processing layer in which the optical processing units are all composed of interferometer modules. The optical signal processing layer composed of interferometer modules is suitable for processing the normalization matrix.

[0039] In some other embodiments, as Figure 7 shown, the optical processing unit is a microring resonator; The microring resonator has: a microring resonator input port and a first output port of the microring resonator; The microring resonator input port serves as the optical input port, and the first output port of the microring resonator serves as the optical output port.

[0040] In this embodiment, the optical processing unit in any optical signal processing layer in the optical signal processor is a microring resonator.

[0041] The microring resonator includes: a first straight waveguide SWG1, a second straight waveguide SWG2, a ring waveguide RWG, and a modulator (not shown in the figure); The first straight waveguide has: a first input end of the straight waveguide SWG11 and a first output end of the straight waveguide SWG12; The first input end of the straight waveguide serves as the microring resonator input port, and the first output end of the straight waveguide serves as the first output port of the microring resonator; The first straight waveguide is coupled to the ring waveguide, the ring waveguide is coupled to the second straight waveguide, and the modulator is attached to the ring waveguide.

[0042] The ring waveguide is a resonant cavity structure formed by connecting the head and tail of a bent waveguide. Its resonance condition is that for light of a specific wavelength, after passing through the micro-ring once, it exactly satisfies constructive interference. Then the light of this wavelength forms resonance in the ring and the light intensity increases; while light of other wavelengths cannot form resonance and is output from the waveguide, as Figure 7 shown. Figure 7 It is a schematic diagram of a common micro-ring resonator structure. Among them, k1 and k2 are coupling coefficients. k1 determines the proportion of light from the first straight waveguide to the ring waveguide, and k2 determines the proportion of light from the ring waveguide to the second straight waveguide.

[0043] According to the characteristics of the micro-ring, the relationship between the transmission spectrum of the upload-download micro-ring and the round-trip phase shift can be calculated. By using this relationship, some mathematical operations can be realized. Through wavelength-division multiplexing technology, controlling different micro-rings to resonate at different wavelengths can achieve parallel matrix operations. Wavelength-division multiplexing means that waves of different wavelengths carry different information and are processed differently, as Figure 8 shown, where MUX is a wavelength-division multiplexer.

[0044] The optical signal operation process is to input an n×1 vector: I = [i1, i2,..., i n T , which is composed of a group of laser light sources including different wavelengths λ1, λ2,..., λ n . Different wavelengths are represented by different colors in the figure. X is a parameter matrix.

[0045] Here, n intensity modulators are used to modulate these n laser light sources respectively to form the input vector. On the chip, the light source is divided into n equal parts and then input into each row respectively. Each row on the chip has n micro-rings, and the resonance peaks of the micro-rings exactly align with the wavelengths of the n input light sources respectively, forming a one-to-one correspondence operation relationship. Each micro-ring only corresponds to one of the input wavelengths and will not affect the input of other wavelengths. The common through-end and common download-end of each row of micro-rings are detected by a balanced detector at the same time, and their differential result constitutes one element o a of the output vector. The outputs of n rows of micro-rings form the output vector O = [o1, o2,..., o n T . The matrix operation formula implemented is as follows: ; The schematic diagram of the three-dimensional stacked structure is as Figure 9 shown, C a refers to the a-th layer optical signal processing layer, PD represents a photodiode, C a-2 ​​Electrical signal E a―2 Transmitted to C a―1 The modulation module M of the layer a―1 , M a―1 For C a―1 The optical computing module O a―1 Has an impact, and then adjusts O a―1 The optical signal transmitted to the photodiode, and the photodiode converts the optical signal into an electrical signal, and this signal is transmitted to the next layer of optical signal processing layer C through the through-silicon via a , and so on.

[0046] The optical signal processing layer composed of micro-ring resonators is suitable for processing non-normalized matrices.

[0047] In some other embodiments, the optical processing units in at least one of the two or more optical signal processing layers are different from the optical processing units in another optical signal processing layer.

[0048] Taking the optical signal processing device with 2 layers of optical signal processing layers as an example: The optical processing unit in one of the optical signal processing layers is an interferometer network; the optical processing unit in the other optical signal processing layer is a micro-ring resonator. This application does not limit which layer of optical processing unit is set as the interferometer network and which layer of optical processing unit is set as the micro-ring resonator.

[0049] The optical signal processing layer composed of both an interferometer network and a micro-ring resonator is suitable for processing scenarios where both a normalized matrix and a non-normalized matrix exist.

[0050] The electro-optic modulator mainly modulates by adjusting the refractive index of the material through an electric field, thereby changing the phase of light, and finally achieving modulation.

[0051] The change of the refractive index of the material with the electric field can be expressed as: ; Among them, n0 represents the refractive index when no external electric field is applied. When the refractive index changes linearly with the applied external electric field (a≠0, and the coefficients of the other orders such as b and c are all 0), it is called the first-order electro-optic effect, that is, the Pockels effect.

[0052] Generally, the first-order electro-optic effect is defaulted to be used, that is, the change law of the refractive index and the electric field satisfies: ; For the modulation of the Mach-Zehnder interferometer, the relationship between the refractive index of the material and the phase: When light travels in a waveguide with a length of L, the phase change caused by the applied external electric field can be expressed as: ; where λ0 is the operating wavelength, n e is the extraordinary refractive index of lithium niobate, r 33 is the electro-optic coefficient of lithium niobate, V A represents the applied voltage, G represents the electrode spacing, Γ mo is the electro-optic overlap integral.

[0053] Δθ and Δ Φ can both be calculated using the above equations.

[0054] For microring resonator modulation, attention is paid to how to implement the calculation of O = X * I.

[0055] Here, X also essentially modulates the refractive index through an electro-optic modulator, thereby achieving modulation.

[0056] Therefore, there is also: ; X can be determined by Δ Φ : ; where I in and I out are the input and output optical intensities respectively. t is the self-coupling coefficient of the microring resonator, α is the internal loss factor, which can generally be simplified to 1, e is the natural constant (2.71828...), and i represents the imaginary unit.

[0057] A schematic diagram of the three-dimensional stacked structure of the multi-layer optical signal processing layer is as Figure 10 shown, Figure 10 in which a complete three-layer optical signal processing layer is shown. Among them, C1 represents the schematic diagram of the structure of the first-layer optical signal processing layer, C2 represents the schematic diagram of the structure of the second-layer optical signal processing layer, and C3 represents the schematic diagram of the structure of the third-layer optical signal processing layer.

[0058] In the optical signal processing device, an optical signal processing layer in which the optical processing units are all composed of interferometer modules can be adopted, or an optical signal processing layer in which the optical processing units are all composed of microring resonators can be adopted. It is also possible to adopt a structure in which the optical processing units of some optical signal processing layers are composed of interferometer modules and the optical processing units of other optical signal processing layers are composed of microring resonators. This application does not make any limitations.

[0059] In some other embodiments, as Figure 11 shown, an optical neural network model training method, applied to the optical signal processing device described above, includes: S100: Set the initial values of the parameters of at least one optical signal processing unit corresponding to the model to be trained; S200: Create the loss function of the model to be trained; S300: Adjust the values of at least one optical signal processing unit parameter corresponding to the model to be trained using the finite difference method until the convergence condition is reached to determine the final values of the optical signal processing unit parameters.

[0060] The model to be trained is usually an optical neural network model. For a structure with a microring resonator as the optical signal unit, the matrix elements of the parameter matrix X described above are updated. For a structure with an interferometer module as the optical signal unit, the first phase θ and the second phase are updated. Φ .

[0061] The convergence condition includes: reaching the preset training statement or the loss function reaching the preset value.

[0062] The loss function can be selected as cross - entropy or the MSE method, which is not limited in this application.

[0063] Optionally, for an optical signal processing device, the on - chip training of the optical neural network is performed using the stochastic error gradient algorithm.

[0064] Preferably, for an optical signal processing device, the optical neural network is trained using the finite difference method.

[0065] In some embodiments, the optical signal processing unit is an interferometer module, and at least one optical signal processing unit parameter is the first phase θ and the second phase corresponding to the optical signal processing unit. Φ , and adjusting the values of at least one optical signal processing unit parameter corresponding to the model to be trained using the finite difference method includes: S310a: Fine - tune the value of the first phase θ and the value of the second phase Φ . S320a: Calculate the loss function of the model to be trained based on the fine - tuned value of the first phase and the fine - tuned value of the second phase; S330aa: In response to the decrease in the value of the loss function, update the value of the first phase with the fine - tuned first phase and update the value of the second phase with the fine - tuned second phase; S330ab: In response to the increase in the value of the loss function, keep the values of the first phase and the second phase unchanged.

[0066] In other embodiments, the optical signal processing unit is a microring resonator, and at least one optical signal processing unit parameter is the parameter matrix corresponding to the optical signal processing unit. Adjusting the values of at least one optical signal processing unit parameter corresponding to the model to be trained using the finite difference method includes: S310b: Fine - tune any parameter matrix element in the parameter matrix to obtain a fine - tuned parameter matrix; S320b: Calculate the loss function of the model to be trained based on the fine-tuned parameter matrix; S330ba: In response to the decrease in the value of the loss function, update the parameter matrix elements of the parameter matrix with the fine-tuned parameter matrix; S330bb: In response to the increase in the value of the loss function, keep the parameter matrix elements in the parameter matrix unchanged.

[0067] In some other embodiments, an optical computing device, such as Figure 12 shown, includes the optical signal processing device and the memory described above. The optical signal processing device therein includes: At least two layers of optical signal processing layers, at least one layer of electrical signal transmission layer, and an electrical signal control layer; The optical signal processing layers are disposed on the device substrate. The optical signal processing layers and the electrical signal transmission layer are spaced apart. The electrical signal control layer is disposed on the topmost optical signal processing layer.

[0068] Among them, the device substrate includes a silicon (Si) substrate and a silicon dioxide substrate (SiO2 substrate). The silicon dioxide substrate is disposed on the silicon substrate. The Nth layer of optical signal processing layer is disposed on the silicon dioxide substrate.

[0069] The electrical signal control layer is provided with through-silicon vias (TSVs) connecting the upper and lower surfaces of the electrical signal transmission layer in this layer, for generating electrical signals for modulating the parameters of the optical signal processing units in the first layer of optical signal processing layer; The first layer of optical signal processing layer is for receiving and modulating the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the electrical signal control layer, determining the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generating electrical signals for adjusting the second layer of optical signal processing layer; The Nth layer of optical signal processing layer is for receiving and modulating the parameters of the optical signal processing units in this optical signal processing layer according to the electrical signals generated by the (N - 1)th layer of optical signal processing layer, determining the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing units in this optical signal processing layer, and generating electrical signals for modulating the (N + 1)th layer of optical signal processing layer, where N is an integer greater than or equal to 2, representing the number of layers of the optical signal processing layer; The electrical signal transmission layer is disposed between the Nth layer of optical signal processing layer and the (N + 1)th layer of optical signal processing layer, for transmitting the electrical signals generated by the Nth layer of optical signal processing layer to the (N + 1)th layer of optical signal processing layer through the through-silicon vias disposed in the electrical signal transmission layer.

[0070] Specifically, as Figure 2As shown in the figure, the optical signal processing layer is provided with: a light source LS, an electro-optic modulation module EOM, a wavelength division multiplexer MUX, at least one optical processing branch LP, a demultiplexer DMUX, a photoelectric conversion module PD, and a photodetector PED; The electro-optic modulation module is connected to the wavelength division multiplexer, the wavelength division multiplexer is connected to at least one optical processing branch, at least one optical processing branch is correspondingly connected to a demultiplexer, the demultiplexer is correspondingly connected to the photoelectric conversion module, and at least one optical processing branch is also correspondingly connected to the photodetector; The light source is used to generate a laser with a preset wavelength; The electro-optic modulation module is used to modulate the light intensity of the laser; The wavelength division multiplexer is used to simultaneously transmit optical signals of at least one wavelength; The optical processing branch is used to determine the optical signal corresponding to this optical processing branch according to the light source; The demultiplexer is used to decompose and obtain an optical signal of a single wavelength; The photoelectric conversion module is used to generate and send an electrical signal for modulating the next adjacent optical signal processing layer; The photodetector is used to detect the optical signal generated by the corresponding optical processing branch.

[0071] Preferably, the modulator is: a lithium niobate electro-optic modulator, a polymer electro-optic modulator, or a silicon-based integrated electro-optic modulator.

[0072] Specifically, as Figure 3 shown, the photoelectric conversion module includes: a photosensitive element, an operational amplifier, a resistor, and a capacitor; The electrical signal output terminal of the photosensitive element is connected to the inverting input terminal of the operational amplifier, both ends of the resistor R are connected in parallel to the inverting input terminal and the operational amplifier output terminal, and the capacitor C is connected in parallel to the inverting input terminal "-" and the operational amplifier output terminal OP o , the non-inverting input terminal "+" of the operational amplifier is grounded. The electrical signal input terminal of the photosensitive element is connected to the voltage V p . The photosensitive element receives the input light source.

[0073] Preferably, the wavelength division multiplexer is an arrayed waveguide grating.

[0074] Preferably, the demultiplexer is a nonlinear Bragg grating.

[0075] Preferably, the light source is a laser.

[0076] Preferably, the laser is an InGaAsP laser or an InP laser.

[0077] Optionally, the laser wavelength range generated by the laser is 1300nm to 2200nm.

[0078] AsFigure 2 As shown, any optical processing branch has: an optical branch input end LP1 and an optical branch output end LP2; The optical branch input end is connected to the output end of the wavelength division multiplexer, and the optical branch output end is connected to the input end of the corresponding demultiplexer; Any optical processing branch includes at least one optical processing unit LU; The optical processing unit has: an optical input port LU1 and an optical output port LU2; In any optical processing branch, the optical output end of the previous-stage optical processing unit is cascaded with the optical input end of the subsequent-stage optical processing unit. The optical input end of the first optical processing unit serves as the optical branch input end, and the optical output end of the last optical processing unit serves as the optical branch output end.

[0079] Optionally, the optical processing unit is an interferometer module; The interferometer module has an interferometer module input end and an interferometer module output end; The interferometer module input end serves as the optical input port, and the interferometer module output end serves as the optical output port.

[0080] In some embodiments, as Figure 4 shown, the interferometer module is an interferometer network; The interferometer network has: a first network input end IW11, a second network input end IW12, a third network input end IW13, a first network output end IW21, a second network output end IW22, and a third network output end IW23; The interferometer network includes: a first interferometer I1, a second interferometer I2, and a third interferometer I3; The first interferometer has: a first input end of the first interferometer, a second input end of the first interferometer, a first output end of the first interferometer, and a second output end of the first interferometer; The second interferometer has: a first input end of the second interferometer, a second input end of the second interferometer, a first output end of the second interferometer, and a second output end of the second interferometer; The third interferometer has: a first input end of the third interferometer, a second input end of the third interferometer, a first output end of the third interferometer, and a second output end of the third interferometer; The first input end of the first interferometer, the second input end of the first interferometer, and the first input end of the second interferometer serve as the interferometer module input end. The first output end of the first interferometer is connected to the second input end of the second interferometer. The second output end of the first interferometer is connected to the second input end of the third interferometer. The second output end of the second interferometer is connected to the first output end of the third interferometer. The second output end of the second interferometer, the first output end of the third interferometer, and the second output end of the third interferometer serve as the interferometer module output end.

[0081] Specifically, the first interferometer, the second interferometer, and the third interferometer are Mach-Zehnder interferometers.

[0082] In some other embodiments, as Figure 5 shown, the interferometer module is a single Mach-Zehnder interferometer; The Mach-Zehnder interferometer has: a first input port In1, a second input port In2, a first output port Out1, and a second output port Out2; The first input port and the second input port serve as the input ends of the interferometer module, and the first output port and the second output port serve as the output ends of the interferometer module; Furthermore, the Mach-Zehnder interferometer includes: a first coupler, a second coupler, a first phase shifter, and a second phase shifter; One end of the first coupler serves as the first input port and the second input port, or as the first input end and the second input end of the first interferometer, or as the first input end and the second input end of the second interferometer, or as the first input end and the second input end of the third interferometer; The other end of the first coupler is connected to one end of the first phase shifter through a waveguide, the other end of the first phase shifter is connected to one end of the second coupler through a waveguide, the other end of the first coupler is also connected to one end of the second coupler through a waveguide, and the other end of the second coupler is connected to the second phase shifter through a waveguide; The other end of the second phase shifter serves as the first output port; The other end of the second coupler serves as the second output port.

[0083] The Mach-Zehnder interferometer includes: a first coupler, a second coupler, a first phase shifter, and a second phase shifter; One end of the first coupler serves as the first input end and the second input end of the first interferometer, or as the first input end and the second input end of the second interferometer, or as the first input end and the second input end of the third interferometer; The other end of the first coupler is connected to one end of the first phase shifter through a waveguide, the other end of the first phase shifter is connected to one end of the second coupler through a waveguide, the other end of the first coupler is also connected to one end of the second coupler through a waveguide, and the other end of the second coupler is connected to the second phase shifter through a waveguide; The other end of the second phase shifter serves as the first output end of the first interferometer, or as the first output end of the second interferometer, or as the first output end of the third interferometer; The other end of the second coupler serves as the second output end of the first interferometer, or as the second output end of the second interferometer, or as the second output end of the third interferometer.

[0084] Figure 6 It shows an optical signal processing layer in which the optical processing units are all composed of interferometer modules.

[0085] In some other embodiments, as Figure 7 shown, the optical processing unit is a microring resonator; The microring resonator has: a microring resonator input port and a first microring resonator output port; The microring resonator input port serves as the optical input port, and the first microring resonator output port serves as the optical output port.

[0086] The microring resonator includes: a first straight waveguide SWG1, a second straight waveguide SWG2, a ring waveguide RWG, and a modulator (not shown in the figure); The first straight waveguide has: a first straight waveguide input end SWG11 and a first straight waveguide output end SWG12; The first straight waveguide input end serves as the microring resonator input port, and the first straight waveguide output end serves as the first microring resonator output port; The first straight waveguide is coupled to the ring waveguide, the ring waveguide is coupled to the second straight waveguide, and the modulator is attached to the ring waveguide.

[0087] The ring waveguide is a resonant cavity structure formed by connecting the head and tail of a bent waveguide. Its resonance condition is that for light of a specific wavelength, after passing through the microring once, it exactly satisfies constructive interference, then the light of this wavelength forms resonance in the ring and the light intensity increases; while light of other wavelengths cannot form resonance and is output from the waveguide, as Figure 7 shown. Figure 7 It is a schematic diagram of a common microring resonator structure, where k1 and k2 are coupling coefficients. k1 determines the proportion of light from the first straight waveguide to the ring waveguide, and k2 determines the proportion of light from the ring waveguide to the second straight waveguide.

[0088] According to the characteristics of the microring, the relationship between the transmission spectrum of the upload - download microring and the round - trip phase shift can be calculated. By using this relationship, some mathematical operations can be realized. Through wavelength - division multiplexing technology, by controlling different microrings to resonate at different wavelengths, parallel matrix operations can be achieved. Wavelength - division multiplexing means that waves of different wavelengths carry different information and are processed differently, as Figure 8 shown, where MUX is a wavelength - division multiplexer.

[0089] The schematic diagram of the three - dimensional stack structure is as Figure 9 shown, C a refers to the a - th layer of the optical signal processing layer, PD represents a photodiode, C a-2 The electrical signal E a―2 is transmitted to the modulation module M a―1 of the C a―1 layer, and M a―1 modulates Ca―1 optical computing module O a―1 has an impact and then adjusts O a―1 The optical signal transmitted to the photodiode is converted into an electrical signal by the photodiode and transmitted to the next-layer optical signal processing layer C through the through-silicon via a , and so on.

[0090] In some other embodiments, the optical processing units in one of the at least two optical signal processing layers are different from the optical processing units in the other optical signal processing layer.

[0091] The electro-optic modulator mainly adjusts the refractive index of the material through an electric field, thereby changing the phase of light, and finally realizes modulation.

[0092] The schematic diagram of the three-dimensional stacked structure of the multi-layer optical signal processing layer is as Figure 10 shown Figure 10 shows the complete three-layer optical signal processing layer. Where C1 represents the schematic diagram of the structure of the first-layer optical signal processing layer, C2 represents the schematic diagram of the structure of the second-layer optical signal processing layer, and C3 represents the schematic diagram of the structure of the third-layer optical signal processing layer.

[0093] The beneficial effects brought by the technical solutions provided in the embodiments of the present application are as follows: By three-dimensionally stacking the optical signal processing layers, the scalability of the optical computing device is improved. By three-dimensionally stacking the optical signal processing layers, the overall number of optical processing units in the optical computing device is increased, enabling the optical computing device to have the ability to process a large amount of data tasks. When processing tasks with a large amount of data, the intensity of the laser light source received by each optical signal processing layer is weak to avoid damage to the silicon optical material due to strong light irradiation.

[0094] Those skilled in the art can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0095] The above has introduced the technical solution provided by this application in detail. Specific examples are used in this text to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, based on the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

[0096] The above has introduced in detail an optical signal processing device provided by this application. Specific examples are used in this text to elaborate on the principle and implementation manner of this application. The above embodiments are only the preferred embodiments of this application, which are used to help understand the method and its core idea of this application, and are not intended to limit this application. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle of this application, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this application also fall within the protection scope of the claims of this application.

Claims

1. An optical signal processing device, characterized in that, Comprising: At least two optical signal processing layers, at least one electrical signal transmission layer, and an electrical signal control layer; The optical signal processing layer is disposed on the device substrate, the optical signal processing layer and the electrical signal transmission layer are spaced apart, and the electrical signal control layer is disposed on the optical signal processing layer at the top layer; The electrical signal control layer is configured to generate an electrical signal for modulating the parameters of the optical signal processing unit in the first optical signal processing layer; The first optical signal processing layer is configured to receive and modulate the parameters of the optical signal processing unit in this optical signal processing layer according to the electrical signal generated by the electrical signal control layer, determine the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing unit in this optical signal processing layer, and generate an electrical signal for adjusting the second optical signal processing layer; The Nth optical signal processing layer is configured to receive and modulate the parameters of the optical signal processing unit in this optical signal processing layer according to the electrical signal generated by the (N - 1)th optical signal processing layer, determine the optical information of this optical signal processing layer according to the light source and the parameters of the optical signal processing unit in this optical signal processing layer, and generate an electrical signal for modulating the (N + 1)th optical signal processing layer, where N is an integer greater than or equal to 2, representing the number of optical signal processing layers; The electrical signal transmission layer is disposed between the Nth optical signal processing layer and the (N + 1)th optical signal processing layer, and is configured to transmit the electrical signal generated by the Nth optical signal processing layer to the (N + 1)th optical signal processing layer through the through-silicon vias disposed in the electrical signal transmission layer.

2. The optical signal processing device according to claim 1, characterized in that, The optical signal processing layer is provided with: a light source, a light intensity modulation module, a wavelength division multiplexer, at least one optical processing branch, a demultiplexer, an optoelectronic conversion module, and a photodetector; The light intensity modulation module is connected to the wavelength division multiplexer, the wavelength division multiplexer is connected to the at least one optical processing branch, the at least one optical processing branch is correspondingly connected to a demultiplexer, the demultiplexer is correspondingly connected to the optoelectronic conversion module, and the at least one optical processing branch is also correspondingly connected to the photodetector; The light source is configured to generate a laser with a preset wavelength; The light intensity modulation module is configured to modulate the light intensity of the laser; The wavelength division multiplexer is configured to simultaneously transmit optical signals of at least one wavelength; The optical processing branch is configured to determine the optical signal corresponding to this optical processing branch according to the light source; The demultiplexer is configured to decompose to obtain an optical signal of a single wavelength; The optoelectronic conversion module is configured to generate an electrical signal for modulating the adjacent lower optical signal processing layer, and transmit it to the adjacent lower optical signal processing layer through the through-silicon vias disposed corresponding to the optoelectronic conversion module in the corresponding optical signal transmission layer; The photodetector is configured to detect the optical signal generated by the corresponding optical processing branch.

3. The optical signal processing device according to claim 2, wherein, The optoelectronic conversion module includes: a photosensitive element, an operational amplifier, a resistor, and a capacitor; The electrical signal output terminal of the photosensitive element is connected to the inverting input terminal of the operational amplifier, both ends of the resistor are connected in parallel between the inverting input terminal and the operational amplifier output terminal of the operational amplifier, and the capacitor is connected in parallel between the inverting input terminal and the operational amplifier output terminal; The electrical signal generated at the output end of the operational amplifier is transmitted through the through-silicon via provided corresponding to the output end of the operational amplifier in the corresponding optical signal transmission layer to the adjacent lower-layer optical signal processing layer.

4. The optical signal processing device according to claim 2, characterized in that, Any optical processing branch has: an optical branch input end and an optical branch output end; The input end of the optical branch is connected to the output end of the wavelength division multiplexer, and the output end of the optical branch is connected to the input end of the corresponding demultiplexer; Any optical processing branch includes at least one optical processing unit; The optical processing unit has: an optical input port and an optical output port; In any optical processing branch, the optical output end of the previous-stage optical processing unit is cascaded with the optical input end of the subsequent-stage optical processing unit. The optical input end of the first optical processing unit serves as the optical branch input end, and the optical output end of the last optical processing unit serves as the optical branch output end.

5. The optical signal processing device according to claim 4, characterized in that, The optical processing unit is an interferometer module; The interferometer module has an interferometer module input end and an interferometer module output end; The input end of the interferometer module serves as the optical input port, and the output end of the interferometer module serves as the optical output port.

6. The optical signal processing device according to claim 5, wherein The interferometer module is an interferometer unit or an interferometer network; The interferometer module is based on: ； Perform modulation, where Δ Φ represents the phase shift angle, λ0 represents the operating wavelength, n e represents the extraordinary optical refractive index of lithium niobate, r 33 represents the electro-optic coefficient of lithium niobate, V A represents the applied voltage, G represents the electrode spacing, L represents the waveguide length, Γ mo represents the electro-optic overlap integral.

7. The optical signal processing device according to claim 4, characterized in that The optical processing unit is a micro-ring resonator; The micro-ring resonator has: a micro-ring resonator input port and a first output port of the micro-ring resonator; The input port of the micro-ring resonator serves as the optical input port, and the first output port of the micro-ring resonator serves as the optical output port.

8. The optical signal processing device according to claim 7, wherein, The micro-ring resonator includes a modulator; The modulator is based on: ； Perform modulation, where I out represents the output optical intensity, I in represents the input optical intensity, t is the self-coupling coefficient of the microring resonator, α is the internal loss factor, e is the natural constant, Φ is the phase angle.

9. The optical signal processing device according to claim 4, characterized in that The optical processing units in one of the at least two optical signal processing layers are different from the optical processing units in another optical signal processing layer.

10. The optical signal processing device according to claim 2, wherein The wavelength division multiplexer and the demultiplexer are arrayed waveguide gratings; Or, The wavelength division multiplexer and the demultiplexer are nonlinear Bragg gratings.

11. The optical signal processing device according to claim 2, characterized in that, The light source is an InGaAsP laser or an InP laser, and the laser wavelength range generated by the laser is 1300 nm to 2200 nm.

12. A method for training an optical neural network model, characterized in that, The method is applied to the optical signal processing device according to any one of claims 1-11, and includes: Setting an initial value of at least one optical signal processing unit parameter corresponding to the to-be-trained model; Creating a loss function of the to-be-trained model; Adjusting the values of at least one optical signal processing unit parameter corresponding to the to-be-trained model by using the finite difference method until a convergence condition is reached to determine the final values of the optical signal processing unit parameters.

13. The optical neural network model training method according to claim 12, wherein The optical signal processing unit is an interferometer module, and the at least one optical signal processing unit parameter is the first phase and the second phase corresponding to the optical signal processing unit. The adjusting the values of at least one optical signal processing unit parameter corresponding to the to-be-trained model by using the finite difference method includes: Fine-tuning the value of the first phase and the value of the second phase; Calculating the loss function of the to-be-trained model based on the fine-tuned value of the first phase and the fine-tuned value of the second phase; In response to the decrease in the value of the loss function, updating the value of the first phase with the fine-tuned first phase and updating the value of the second phase with the fine-tuned second phase; In response to an increase in the value of the loss function, the values of the first phase and the second phase are kept unchanged.

14. The optical neural network model training method according to claim 12, characterized in that The optical signal processing unit is a microring resonator, and the at least one optical signal processing unit parameter corresponds to a parameter matrix for the optical signal processing unit. Adjusting the value of at least one optical signal processing unit parameter corresponding to the to-be-trained model by using the finite difference method includes: Fine-tuning any parameter matrix element in the parameter matrix to obtain a fine-tuned parameter matrix; Calculating the loss function of the to-be-trained model based on the fine-tuned parameter matrix; In response to a decrease in the value of the loss function, updating the parameter matrix element of the parameter matrix with the fine-tuned parameter matrix; In response to an increase in the value of the loss function, keeping the parameter matrix elements in the parameter matrix unchanged.

15. An optical computing device, characterized in that, It includes the optical signal processing device according to any one of claims 1-11 and a memory.

Images