A method for convolution parallel computation based on MZI space division multiplexing

Abstract

The application discloses a kind of MZI-based space division multiplexing convolution parallel computing method, including, by in situ training algorithm configuration one MZI array, make it in single control state, to N kind of preset wavelength non-coherent light respectively present with N corresponding target transmission characteristics of preset convolution kernel.For calculation, the input vector to be processed is encoded as the incident power of N wavelength light and injected into the array, and the superposition value of the power of each wavelength light is directly detected at the output port, which is the N parallel convolution result.The system of the application correspondingly includes an incident layer, a multi-wavelength working MZI array and a summation layer.The application can improve the calculation parallelism by N times without increasing hardware overhead, and breaks through the unitary matrix limitation of traditional MZI array, significantly improves the calculation density and hardware utilization efficiency, and has wide application prospect in the field of photonic computing.

Description

Technical Field

[0001] This invention relates to the field of photonic computing technology, and in particular to a spatial division multiplexing convolution parallel computing method based on MZI. Background Technology

[0002] With the booming development of computationally intensive applications such as artificial intelligence and data centers, traditional electronic computing architectures are facing increasing challenges in terms of performance improvement and power consumption control. Among these, photonic computing, with its advantages of ultra-high speed, ultra-large bandwidth, low power consumption, and inherent parallelism, is considered a potential technological path to break through the von Neumann bottleneck and continue the growth of computing power. Specifically, on-chip reconfigurable photonic processors based on Mach-Zehnder interferometer (MZI) arrays, which construct optical cores capable of arbitrary unitary matrix transformations through cascaded MZI units, have become the mainstream solution for performing basic linear algebra operations such as matrix-vector multiplication, and have shown broad application prospects in the hardware acceleration of neural networks, especially convolutional neural networks.

[0003] However, existing computing schemes based on on-chip MZI arrays still face several technical bottlenecks. First, when scaling up computation to handle higher-dimensional vectors, the number of MZI units typically increases quadratically (O(n²)), resulting in huge chip area overhead, significantly increasing manufacturing costs and limiting device integration density. Second, while existing technologies involve multi-wavelength operating modes of MZI arrays, their research focuses primarily on suppressing device dispersion effects to ensure consistent transmission characteristics in broadband communications, rather than actively utilizing response differences between different wavelengths to create independent parallel computing channels. Furthermore, for specific applications such as convolution operations, the transmission matrix of a single MZI array configured at a single wavelength is essentially a unitary matrix. When multiple different convolution kernels are directly combined into a large computation matrix, this matrix typically does not satisfy the definition of a unitary matrix, making it extremely difficult to process multiple different convolution kernels in parallel within a single computation using a single MZI array. Therefore, how to effectively utilize the multiplexing dimension of light to improve the computational density and parallelism of MZI arrays without significantly increasing hardware overhead is an urgent problem to be solved in the field of photonic computing. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a spatially divided multiplexing convolutional parallel computation method based on MZI to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a spatial multiplexing convolution parallel computing method based on MZI, comprising:

[0007] An MZI array is obtained, and the MZI array is configured by adjusting the control signals applied to each phase-shifting arm in the MZI array, so that for any wavelength of incoherent light of N preset wavelengths, the transmission characteristics of the MZI array correspond to a target transmission characteristic set for that wavelength.

[0008] The input vector to be processed is encoded into the incident light power of the N wavelengths of incoherent light, with each value corresponding to a value.

[0009] The encoded incoherent light of N wavelengths is incident on the MZI array, and the superposition value of the optical power is detected at each output port of the MZI array. The superposition value is the result of the parallel convolution operation.

[0010] As a preferred embodiment of the MZI-based spatial multiplexing convolution parallel computing method described in this invention, the configuration includes:

[0011] Select multiple convolution kernels and flatten each kernel into a one-dimensional vector;

[0012] Arrange the multiple one-dimensional vectors into a matrix, and select the column vectors of the matrix as the target output vectors corresponding to each preset wavelength;

[0013] By using a feedback algorithm, the control signals applied to each phase-shifting arm of the MZI array are adjusted so that when light of a single wavelength is incident, the optical power distribution of each output port of the MZI array matches the target output vector corresponding to that wavelength, thereby making the transmission characteristics of the MZI array approximate the target transmission characteristics.

[0014] As a preferred embodiment of the spatial multiplexing convolution parallel computing method based on MZI described in this invention, the feedback algorithm is a genetic algorithm.

[0015] As a preferred embodiment of the MZI-based spatial multiplexing convolution parallel computing method described in this invention, after selecting multiple convolution kernels and flattening each convolution kernel into a one-dimensional vector, the method further includes:

[0016] The numerical values ​​in the one-dimensional vector are linearly mapped from a preset interval containing positive and negative values ​​to a target interval containing only non-negative values.

[0017] As a preferred embodiment of the spatial division multiplexing convolution parallel computing method based on MZI described in this invention, the input vector is a one-dimensional vector formed by flattening pixel block data extracted from the image to be processed.

[0018] As a preferred embodiment of the MZI-based spatial multiplexing convolution parallel computing method described in this invention, before encoding the values ​​of each item of the input vector to be processed into the incident light power of the N wavelengths of incoherent light, the method further includes:

[0019] An adjustable optical attenuator ensures that the optical power of the N wavelengths of incoherent light remains consistent before encoding.

[0020] As a preferred embodiment of the spatial division multiplexing convolution parallel computing method based on MZI described in this invention, wherein: the MZI in the MZI array includes a phase-shifting arm that is thermally modulated, and the two sides of the phase-shifting arm are provided with deep etching structures to improve thermal modulation efficiency.

[0021] As a preferred embodiment of the spatial division multiplexing convolution parallel computing method based on MZI described in this invention, the phase shift range generated by the phase shift arm is made to exceed 2π by adjusting the control signal applied to the phase shift arm, so as to achieve separate control at multiple wavelengths.

[0022] As a preferred embodiment of the spatial division multiplexing convolution parallel computing method based on MZI described in this invention, the multi-wavelength light source is split by an arrayed waveguide grating, and the optical power of each wavelength after splitting is independently modulated using the structure of MZI to complete the encoding of the input vector.

[0023] Furthermore, the present invention also provides the following technical solution: a system for multi-wavelength parallel convolution operations, comprising:

[0024] An incident layer is used to encode the input vector into incident light power of incoherent light of N preset wavelengths;

[0025] A multi-wavelength MZI array is connected to the incident layer optical path. The MZI array is configured to have a specific, preset target transmission characteristic for each of the N wavelengths, so as to realize different matrix transformations for light of different wavelengths.

[0026] A summation layer, connected to the output optical path of the MZI array, is used to detect the superposition value of N wavelength optical power at each output port of the MZI array to obtain the parallel convolution operation result.

[0027] Compared with existing technologies, the beneficial effects of this solution are:

[0028] 1. This invention utilizes the dispersion response of an MZI array to different wavelengths to achieve N-way parallel computing on a single physical hardware device through wavelength division multiplexing (WDM). Compared to traditional methods that increase computing power by expanding the physical size of the MZI array, this invention increases computing power by N times without significantly increasing hardware overhead and chip area, greatly improving computing density and hardware utilization efficiency.

[0029] 2. Existing MZI arrays can only perform a single unitary matrix transformation at a single wavelength, making it difficult to process multiple convolutional kernels, which are typically non-unitary, in parallel. This invention, by independently configuring a target transmission characteristic for each wavelength channel, enables a single MZI array to be simultaneously equivalent to N different, non-unitary convolutional kernel matrices. This successfully solves the technical challenge of performing multiple convolutional operations in parallel, significantly expanding the application scope of MZI arrays in fields such as convolutional neural networks.

[0030] 3. This invention inherits the low-power advantage of optical computing's "propagation is computation" principle. Simultaneously, by configuring the MZI array through an in-situ training algorithm, any N sets of convolutional kernels can be flexibly mapped to N wavelength channels, achieving high software definability and hardware reconfigurability, easily adapting to different computing tasks, and minimizing system overhead. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0032] Figure 1 This is a schematic diagram of in-situ training of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0033] Figure 2 This is a diagram showing the relationship between the convolution kernel and the corresponding vector of the corresponding wavelength in the spatial division multiplexing convolution parallel computing method based on MZI according to one embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram of the image processing flow of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0035] Figure 4 This is a comparison chart of the four-way parallel output results of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention and the benchmark results of traditional computer computing.

[0036] Figure 5This is an example diagram of an MZI array application of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0037] Figure 6 This is a scan data diagram of each port of the AWG in the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of a single MZI structure in the spatial multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0039] Figure 8 This is a phase modulation efficiency diagram of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0040] Figure 9 This is a simulation test diagram of the spatial division multiplexing convolution parallel computing method based on MZI according to an embodiment of the present invention;

[0041] Figure 10 This is a flowchart of the test platform for the MZI-based spatial multiplexing convolution parallel computing method according to an embodiment of the present invention.

[0042] Figure 11 This is a schematic diagram of an on-chip integrated photonic processor for a spatially divided multiplexing convolutional parallel computing method based on MZI, according to an embodiment of the present invention. Detailed Implementation

[0043] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0047] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0048] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0049] Example 1

[0050] Reference Figures 1 to 5 This is the first embodiment of the present invention, which details a specific method for implementing parallel computation of four 2×2 convolutional kernels based on a 4×4 MZI array. This embodiment mainly includes the configuration stage of the MZI array and the parallel computation stage.

[0051] S1: MZI Array Configuration Phase

[0052] It should be noted that the goal of this configuration phase is to achieve in-situ training, such as... Figure 1 As shown, the control signals (such as voltage) applied to each phase-shifting arm in the MZI array are adjusted so that the MZI array exhibits N different target transmission characteristics for N different wavelengths of incoherent light under the same set of control signals. In this embodiment, N is taken as 4.

[0053] Furthermore, four 2×2 convolution kernels that need to be computed in parallel are selected. For example... Figure 2 As shown on the right, these four convolutional kernels can be of any value, such as kernels used for edge detection or feature extraction. Each 2×2 convolutional kernel is unfolded (flattened) by row or column to form a 4×1 one-dimensional vector.

[0054] Furthermore, since the optical power detector can only detect non-negative values, and the element values ​​in the convolution kernel are usually distributed in the range of [-1,1], a linear mapping is required.

[0055] Specifically, the values ​​in the flattened one-dimensional vectors of the convolution kernels are linearly mapped from a preset interval containing positive and negative values ​​(e.g., [-1, 1]) to a target interval containing only non-negative values ​​(e.g., [0, 1]). For example, -1 is mapped to 0, 0 to 0.5, and 1 to 1. After this processing, each convolution kernel corresponds to a one-dimensional vector whose element values ​​are all within the interval [0, 1]. Then, the four preprocessed one-dimensional vectors (each representing a convolution kernel) are arranged column-wise to form a 4×4 target matrix. Then, each column of the target matrix is ​​selected as a target output vector. (i=1, 2, 3, 4), where, Corresponding to the i-th preset wavelength The target output optical power distribution.

[0056] Furthermore, a test platform was built. A tunable laser was used as the light source, and four preset wavelengths were selected one by one. One of them. For each wavelength The power is injected sequentially from one input port (e.g., port j) of the MZI array, and the normalized power distribution is detected at all four output ports using an optical power meter or an on-chip photodiode array.

[0057] Furthermore, this embodiment employs a genetic algorithm as the feedback algorithm. The execution process of this feedback algorithm is as follows:

[0058] The control voltage values ​​of all adjustable phase-shifting arms in the MZI array are combined into a long one-dimensional solution vector. Then, a loss function is defined to quantize the unitary matrix of the i-th iteration. With the target matrix The gap between them.

[0059] Specifically, the defined loss function can be expressed as:

[0060]

[0061] Where L is the total loss function value, representing the nth wavelength among the four preset wavelengths. Under the control voltage of the current iteration, the MZI array for wavelength is The transmission matrix presented by the light. This is the wavelength The set target transmission matrix. It is a physical quantity that is actually measured, and its meaning is: when the wavelength is The vector consisting of the normalized optical power detected at all (4 in this case) output ports when light is incident from the nth input port. It is a predefined target output optical power vector corresponding to the above input conditions. This represents the square of the L2 norm (Euclidean distance) of the difference between two vectors.

[0062] It needs to be explained that the physical meaning of the defined loss function is that, for each wavelength, light is injected from the corresponding input port, then the mean square error between the measured output light power distribution vector and the target vector is calculated, and finally all these errors are summed up to obtain the total loss value.

[0063] In each generation, the genetic algorithm generates a set of voltage solution vectors and applies them one by one to the MZI array, measuring the corresponding loss function. Based on the magnitude of the loss function, each solution vector is assigned a fitness (reproduction probability); solutions with low loss have high fitness. The next generation of solution vectors is generated through operations such as selection, crossover, and mutation. This process iterates until the loss function converges to a minimum (e.g., no better solution appears for several consecutive generations). The final voltage solution vector obtained at this point is the desired result.

[0064] It should be noted that the MZI array was successfully configured after training. Under this fixed voltage, when the wavelength is... When light passes through the array, its behavior approximates that of the first target matrix (by...). (Definition); when the wavelength is When light passes through, its behavior approximates that of the second target matrix (composed of...). (Definition), and so on, thus realizing the computational function of encoding different wavelengths on a single physical structure.

[0065] Furthermore, a 2×2 pixel block (receptive field, RF) is extracted from the image to be processed using a sliding window method. The gray values ​​of this pixel block (normalized to [0, 1]) are flattened to form a 4×1 input vector. Then, the input vector The four element values ​​are respectively encoded into the optical power of the four input ports of the MZI array.

[0066] Specifically, the total incident optical power of the j-th input port (j=1, 2, 3, 4) is related to... The value of the j-th element is proportional to the value of the input. At each input port, a value containing... Four incoherent light sources of different wavelengths.

[0067] Furthermore, the encoded optical signal enters the configured MZI array for transmission. Since light of different wavelengths is incoherent, they interact independently with the MZI array. At the nth output port, the optical power meter detects the direct superposition of the optical power output from the four wavelength channels at that port. This total optical power value is one of the four results obtained after the input pixel block is convolved with the four convolution kernels in parallel. Simultaneous detection at all four output ports allows for the acquisition of four convolution results at once. (Refer to...) Figure 3 .

[0068] It should be noted that the reference Figure 4 The above method was applied to image processing, and the four-way parallel output results were then compared with the benchmark results (Ground Truth) calculated by traditional computers. The signal-to-noise ratio (SNR) reached 35.57dB and the structural similarity (SSIM) reached 0.6679, which proved the high fidelity and effectiveness of the proposed method.

[0069] Example 2

[0070] Reference Figures 5 to 11 This is a second embodiment of the present invention, which provides a system for multi-wavelength parallel convolution operations. The system mainly includes an incident layer, a multi-wavelength working MZI array, and a summation layer, such as... Figure 5 As shown.

[0071] Specifically, the incident layer is responsible for generating multi-wavelength light sources and encoding the input vector into optical power. The system uses multi-wavelength light sources and separates four incoherent optical channels (such as...) through an on-chip integrated arrayed waveguide grating (AWG). Figure 6 (As shown in the spectrum). After AWG splitting, an MZI structure can be connected after each optical channel as an adjustable optical attenuator to precisely modulate the optical power to encode the input vector.

[0072] Specifically, for multi-wavelength MZI arrays, and in order to fully describe the present invention, this embodiment will first explain the transmission characteristics of the basic unit constituting the array, namely a single MZI.

[0073] Furthermore, an ideal MZI consists of two 3dB beam splitters (i.e., a splitting ratio k=0.5) and two phase-shifting arms, such as... Figure 7 As shown. Its transmission characteristics can be represented by a 2×2 transmission matrix. describe.

[0074] Specifically, the MZI transmission characteristics can be given by the following formula:

[0075]

[0076] in, This indicates the phase shift of the first phase-shifting arm. This indicates the phase shift of the second phase-shifting arm, where k is the splitting ratio. For an ideal 3dB beam splitter, k = 0.5. Phase shift or The relationship with the actual applied voltage can be written in the following form:

[0077]

[0078] in, The initial phase difference between the two phase-shifting arms when no power is applied. Where E is the electro-optic efficiency and E is the applied voltage.

[0079] Furthermore, this 2×2 transmission matrix is ​​a complete second-order unitary matrix, meaning it can achieve optical beam splitting with arbitrary splitting ratios and phases for two beams. Moreover, arranging the MZIs in a triangular or square array can cover any N×N unitary matrix. Based on this, any unitary matrix can be decomposed and reduced in order as follows: it can be written as the product of multiple 2×2 complete unitary matrices, allowing it to be decomposed into the phase shift corresponding to each MZI in the array, resulting in:

[0080]

[0081] in, To represent the extended transmission matrix of the MZI between waveguides m and m+1, it can be written as a 2×2 matrix with m and m+1 rows. The remaining rows are expanded into an identity matrix.

[0082] Furthermore, to achieve efficient multi-wavelength parallel computing, this embodiment optimizes the aforementioned single MZI. The phase-shifting arm employs a thermo-optical effect for phase modulation, and is covered with a micro-heater made of TiN material. Crucially, deeply etched isolation trenches are designed on both sides of the phase-shifting arm waveguide. For example... Figure 8 Simulation results show that this structure effectively prevents heat diffusion to the substrate and adjacent waveguides, concentrating heat in the target waveguide region, thereby significantly improving thermal tuning efficiency. Experimental data indicate that compared to structures without deep etching, thermal tuning efficiency is improved by 36.8%, enabling high performance at wavelengths of 1548.8 nm and 1552.0 nm. The voltage required for phase shift is reduced to approximately 1.8V. Simultaneously, high thermal efficiency allows for achieving higher voltages at lower control voltages. The large phase shift is crucial for multi-wavelength configurations. For example... Figure 9 This illustrates that since the transmission characteristics of MZI (such as splitting ratio, initial phase, and electro-optic efficiency) are all wavelength-dependent, different phase shifts are required to achieve the respective target phase for different wavelengths. (Exceeding...) The phase shift range provides additional degrees of freedom, making it more possible to find a "common solution" that can take into account the target characteristics of all wavelengths by adjusting a single voltage value to simultaneously meet the different phase requirements of multiple wavelengths in a two-dimensional (or multi-dimensional) wavelength-phase space.

[0083] Specifically, for the summation layer, at the output of the MZI array, light is led out of the chip via a grating coupler and detected by an external optical power meter array. Since the light from each wavelength channel is incoherent, the readings of the optical power meter naturally sum the output power of each wavelength; this summation value is the final calculation result. Alternatively, a photodiode array or an external optical power meter can be integrated on-chip as the summation layer, as described in the reference. Figure 10 In addition, you can refer to Figure 11 The PC runs the genetic algorithm and calculates the loss function. The FPGA drives the digital-to-analog converter (DAC) based on the solution vector generated by the algorithm, providing precise control voltage to all the microheaters in the MZI array; the optical power meter readings are fed back to the PC for the next round of iterative optimization. A CCD camera is used to assist in aligning the fiber optic cable with the grating coupler.

[0084] It should be noted that, through the above structural design and optimization, the system of this embodiment can stably and efficiently execute the multi-wavelength parallel convolution method described in Embodiment 1.

[0085] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0090] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for MZI-based space division multiplexing convolution parallel computation, characterized in that, The method comprises: obtaining an MZI array, and configuring the MZI array by adjusting control signals applied to each phase-shifting arm in the MZI array, so that for any one of N preset wavelengths of non-coherent light, the transmission characteristics of the MZI array correspond to a target transmission characteristic set for the any one wavelength; the configuration comprises: selecting a plurality of convolution kernels and flattening each convolution kernel into a one-dimensional vector; arranging a plurality of the one-dimensional vectors into a matrix, and selecting column vectors of the matrix as target output vectors corresponding to each preset wavelength; adjusting the control signals applied to each phase-shifting arm of the MZI array through a feedback algorithm, so that when single-wavelength light is incident, the light power distribution of each output port of the MZI array matches the target output vector corresponding to the wavelength, thereby making the transmission characteristics of the MZI array approximate the target transmission characteristics; an input vector to be processed, each numerical value of which corresponds to the incident light power of non-coherent light encoded into the N wavelengths; the N wavelengths of non-coherent light after encoding are incident to the MZI array, and the superposition value of the light power at each output port of the MZI array is detected, which is the result of the parallel convolution operation.

2. The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, The feedback algorithm is a genetic algorithm. 3.The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, After the step of selecting a plurality of convolution kernels and flattening each convolution kernel into a one-dimensional vector, the method further comprises: linearly mapping the numerical values in the one-dimensional vector from a preset interval containing positive and negative values to a target interval containing only non-negative values.

4. The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, The input vector is a one-dimensional vector formed by flattening pixel block data extracted from an image to be processed.

5. The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, Before the step of obtaining an input vector to be processed, each numerical value of which corresponds to the incident light power of non-coherent light encoded into the N wavelengths, the method further comprises: an adjustable optical attenuator, which keeps the light power of the N wavelengths of non-coherent light consistent before encoding.

6. The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, The MZI in the MZI array comprises a phase-shifting arm that is phase-shifted by heat, and deep etching structures are arranged on both sides of the phase-shifting arm to improve the heat adjustment efficiency.

7. The MZI-based space division multiplexing convolution parallel computing method according to claim 6, wherein, By adjusting the control signals applied to the phase-shifting arm, the phase shift range of the phase-shifting arm exceeds 2π, thereby achieving independent control of multiple wavelengths.

8. The MZI-based space division multiplexing convolution parallel computing method of claim 1, wherein, The arrayed waveguide grating is used to split the multi-wavelength light source, and the structure of the MZI is used to independently modulate the light power of each wavelength after splitting, thereby completing the encoding of the input vector.

9. A system for multi-wavelength parallel convolution operation, characterized in that, The method comprises: an incident layer for encoding an input vector into the incident light power of N preset wavelengths of non-coherent light; an MZI array operating at multiple wavelengths, which is connected in optical path with the incident layer, and is configured to have a specific and preset target transmission characteristic for each of the N wavelengths, so as to achieve different matrix transformations of light of different wavelengths; the configuration comprises: selecting a plurality of convolution kernels and flattening each convolution kernel into a one-dimensional vector; arranging a plurality of the one-dimensional vectors into a matrix, and selecting column vectors of the matrix as target output vectors corresponding to each preset wavelength; The control signals applied to each phase-shifting arm of the MZI array are adjusted by a feedback algorithm so that when a single wavelength light is incident, the light power distribution of each output port of the MZI array matches the target output vector corresponding to the wavelength, thereby making the transmission characteristics of the MZI array approximate the target transmission characteristics; a summing layer connected to the output light path of the MZI array, for detecting the superimposed value of the light power of N wavelengths at each output port of the MZI array to obtain the parallel convolution operation result.

Images