Large-scale distributed photon convolutional neural network architecture design method based on noise perception search

Abstract

The invention discloses a large-scale distributed photon convolutional neural network architecture design method based on noise perception search, and the method comprises the steps: flexibly decomposing a complex task corresponding to a large-scale convolutional neural network into a plurality of independent simple subtasks through coding, and forming a distributed architecture; an error network modeling method based on a Mach-Zehnder interference module establishes an offset weight matrix in a real hardware environment, constructs a noise-aware microloss function to update architecture parameters, and designs a photon convolutional network structure suitable for subtasks; therefore, optical loss and noise accumulation are effectively suppressed, and the expandability of the architecture is improved. According to the method, the scale and the computing power of the photon convolutional neural network based on the Mach-Zehnder interference module are improved, and a complex artificial intelligence application scene is further supported.

Description

TECHNICAL FIELD

[0001] The patent relates to the technical field of artificial intelligence and photonic computing, and particularly relates to a large-scale distributed photonic convolutional neural network architecture design method based on noise perception search. BACKGROUND

[0002] The rapid development of artificial intelligence technology represented by deep neural networks (DNNs) has led to an increase in computing power that far exceeds Moore's Law. In the DNN-intensive computing environment, the traditional electronic computing architecture based on the von Neumann architecture design faces the problem of huge energy consumption and time delay caused by data exchange between memory units and processing units. Photonic convolutional neural networks, as a new computing paradigm, have physical characteristics of large bandwidth, high speed and low power consumption, and an efficient computing mode based on local weight sharing mechanism. Using photons as the basic information carrier, photonic convolutional neural networks can perform core operations of neural networks in the optical domain, thus fundamentally reducing data exchange and being considered as one of the key technologies to break through the existing computing power bottleneck.

[0003] The photonic computing architecture based on an array of Mach Zehnder interferometers (MZIs) is highly compatible with mature CMOS technology, and its technical frontier has developed from principle verification to high-performance new architecture design. However, due to manufacturing process defects, device coupling loss and crosstalk, the simulation of photonic computing based on MZI is still limited to simplified benchmark tasks such as handwritten digit recognition and vowel recognition. As the scale expands, the exponential accumulation of optical loss and noise will cause a sharp deterioration of the signal-to-noise ratio, making it difficult for existing architectures to expand into large-scale optoelectronic computing systems while maintaining high accuracy, and unable to meet the needs of complex artificial intelligence tasks.

[0004] To address this challenge, the original computing task is divided into several subtasks using a "divide and conquer" strategy, and a shallow and wide network architecture is constructed to effectively suppress the accumulation of optical loss and noise and improve the scalability of the architecture. The distributed optoelectronic intelligent computing method proposed in Chinese Patent No. CN117349225 B relies on a fixed diffraction weight matrix to compress high-dimensional input, which is too ideal for the assumption that all two-dimensional features are applicable, and cannot dynamically adjust the network structure to adapt to different subtask requirements.

[0005] In summary, how to design a distributed architecture that supports noise suppression, high flexibility and large scale is a problem that needs to be solved for photonic convolutional neural networks based on MZI. SUMMARY

[0006] This patent addresses the problem that existing MZI-based photonic convolutional neural network architectures are difficult to scale to large scale due to light loss and noise. It proposes a method for designing a large-scale distributed photonic convolutional neural network architecture based on noise-aware search. The method includes the following steps:

[0007] S1. By analyzing the large-scale convolutional neural network, the number of classifications is N. C For each category of complex tasks, perform N L The N-valued encoding of the bits is used as row vectors to form an encoding matrix M, which forms a decomposition module used to decompose complex tasks into N... L The system consists of N independent, simple subtasks for each category, forming a distributed architecture; the value of N is in the range of 2 ≤ N < N. C The N L The range of values ​​is

[0008] S2. Using an MZI-based error network modeling method, an offset weight matrix W is established under real hardware conditions. A noise-aware differentiable loss function is constructed to update the architecture parameters, forming a noise-aware search algorithm. This algorithm is used to design the photonic convolutional network structure and mapped to an information loading module, a linear processing module, and a nonlinear processing module, forming N. L A base learner module based on noise-aware search is used to perform the N described in S1. L There are N independent simple subtasks with N categories; where:

[0009] The information loading module is used to load one-dimensional light field information of a sample through a single MZI;

[0010] The linear processing module is used to perform convolution operations through an on-chip MZI array composed of multiple directional couplers (DCs), multiple thermo-optical phase shifters (PSs), and an on-chip semiconductor optical amplifier array;

[0011] Furthermore, the on-chip MZI array consists of two unitary matrix modules and one diagonal matrix module, which are used to convert the convolution operation into a general matrix-vector multiplication through image transposition, and to decompose the matrix into the product of a unitary matrix U, a diagonal matrix S, and a unitary matrix V through singular value decomposition to achieve arbitrary convolution operations; furthermore, the Clements-type matrix decomposition method is used to map the elements of the unitary matrix to the phase change parameters of the PSs of the MZI array;

[0012] The nonlinear processing module is used to convert the optical calculation results into electrical signals by using a photodetector, output the modified linear unit activation function response, and convert it into a one-dimensional optical signal by an electro-optic intensity modulator and PSs.

[0013] S3. For each input sample, based on N described in S2 L The N outputs of the base learner modules based on noise-aware search L Bit-N value encoding, calculating N for each category in the encoding matrix M described in S1. L The Hamming distance is encoded using a standard N-valued encoding. The standard class corresponding to the minimum distance is output to form a synthesis module, which is used to reconstruct the final large-scale convolutional neural network with N classes. C The classification results of complex tasks.

[0014] The noise-aware search algorithm includes the following steps:

[0015] S1. Construct a hypernetwork, which is defined as a directed acyclic graph containing n directed edges. Each directed edge is defined as a hybrid operation containing K parallel candidate operations.

[0016] S2. Define the architecture parameters as the probability of selecting each candidate operation, and define the weight parameters as the weights within each candidate operation;

[0017] S3. To reduce memory usage, a binary gate g is used to define the output of the mixed operation:

[0018]

[0019] In the formula Indicates the output of the mixed operation. Let p be the set of candidate operations. i Indicates candidate operation o i The probability of choosing;

[0020] S4. Divide the dataset into training set, validation set and test set according to the proportion, and reallocate the N-valued labels according to the encoding matrix M;

[0021] S5. By modeling an error network based on MZI, a noise-aware differentiable loss function for updating architecture parameters is constructed, modeled as follows:

[0022]

[0023] In the formula, Loss represents the differentiable loss function for noise perception. CE Represents the standard cross-entropy loss function. This represents a weight decay term to prevent overfitting. λ1 and λ2 represent the expected value of the network's relative variation distance (RVD), where λ1 and λ2 are both scaling factors. This represents the expected value of RVD in the i-th network layer. This represents the probability of selecting the j-th candidate operation in the i-th network layer. Let W represent the RVD of the j-th candidate operation in the i-th network layer, W represent the offset weight matrix, and W0 represent the ideal weight matrix. m,n , These represent the elements in the m-th row and n-th column of W and W0, respectively.

[0024] S6. Initialize the architecture parameters using a uniform distribution and the weight parameters using a Gaussian distribution.

[0025] S7. Use gradient descent to train the architecture parameters and weight parameters alternately, and stop training after the iterative constraints are met; wherein, the weight parameters are updated using the training set based on the standard cross-entropy loss function, and the architecture parameters are updated using the validation set based on the noise-aware differentiable loss function.

[0026] S8. Determine the photonic convolutional network structure based on the architecture parameters obtained after training.

[0027] Furthermore, the MZI-based error network modeling method establishes a 2×2 unit-level transfer matrix T in a real hardware environment by introducing optical loss, crosstalk noise, and quantization noise error terms. MZ The offset weight matrix W under network-level error is reconstructed based on Clements-type matrix decomposition, singular value decomposition, and image transposition.

[0028] Among them, the 2×2 unit-level transition matrix T MZ It can be modeled as:

[0029]

[0030] In the formula T PL Let T be the waveguide transfer matrix. DC Let T be the transition matrix of DC. PS Let α be the transition matrix of PS. p l represents the propagation loss per unit distance. MZI α represents the waveguide length of MZI. DC σ represents the insertion loss of DC, κ represents the cross-coupling coefficient of DC, and σ represents the insertion loss of DC. dc α represents the standard deviation of the DC cross-coupling coefficient. m γ represents the metal absorption loss of PS, γ represents the actual phase setting value considering quantization error and manufacturing defects, and σ represents the metal absorption loss of PS. ps κ represents the standard deviation of the phase setting of PS, γ0 represents the nominal phase setting value determined by the finite bit resolution, and both κ and γ follow a Gaussian distribution.

[0031] In summary, this patent provides a design method for a large-scale distributed photonic convolutional neural network architecture based on noise-aware search. It flexibly decomposes complex tasks into multiple independent, simple sub-tasks through encoding, forming a distributed architecture. Based on the MZI-based error network modeling method, it establishes an offset weight matrix under real hardware conditions, constructs a noise-aware differentiable loss function to update architecture parameters, forms a noise-aware search algorithm, and designs a photonic convolutional network structure suitable for sub-tasks. This effectively suppresses optical loss and noise accumulation, improves network scale and computing power, and thus supports complex artificial intelligence task scenarios. Attached Figure Description

[0032] Figure 1 Flowchart of a method for designing a large-scale distributed photonic convolutional neural network architecture based on noise-aware search;

[0033] Figure 2 This is a schematic diagram of a large-scale distributed photonic convolutional neural network architecture based on noise-aware search.

[0034] Figure 3 This is a schematic diagram of the base learner module structure based on noise-aware search.

[0035] Figure 4 A schematic diagram of the MZI configuration for the linear processing module;

[0036] Figure 5 A flowchart of the noise-aware search algorithm;

[0037] Figure 6 This is a schematic diagram of hypernetwork parameter updates;

[0038] Figure 7 This is a schematic diagram of the error network modeling method based on MZI. Detailed Implementation

[0039] To make the objectives, technical methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] Figure 1 and Figure 2 The diagrams and illustrations show the design method for a large-scale distributed photonic convolutional neural network architecture based on noise-aware search. The design method includes the following steps:

[0041] S1. By analyzing the large-scale convolutional neural network and the number of classifications corresponding to it, N... C For each category of complex tasks, perform N LThe N-valued encoding of the bits is used as row vectors to form the encoding matrix M, which forms the decomposition module 10, used to decompose complex tasks into N... L The system consists of N independent, simple subtasks for each category, forming a distributed architecture; the value of N is in the range of 2 ≤ N < N. C The N L The range of values ​​is

[0042] For example, the Cifar10 dataset is used as a 10-class classification task, which is broken down into 50 independent, simple subtasks of 3-class classification. Specifically, each of the 10 classes is encoded using a 50-bit ternary encoding, forming an encoding matrix M as row vectors. Correspondingly, the encoding matrix M contains 50 column vectors, each corresponding to a 10-bit ternary encoding. Based on the encoding, the original 10 classes are reassigned to 3 classes, and the subtasks to be performed by each base learner module are reconstructed.

[0043] It is important to note that the following principles should be followed during the encoding process: (1) The row vectors of the encoding matrix M should be as different as possible from each other, so that even if one base learner makes a mistake, the accumulation of all the errors is not enough to transform the codeword of one class into the codeword of another class. (2) The column vectors of the encoding matrix M should be as uncorrelated as possible to ensure the diversity of the base learner set. (3) The number of each class of the subtask corresponding to the column vectors of the encoding matrix M should be roughly balanced to avoid overly simple or difficult splits and maintain a moderate subtask complexity.

[0044] S2. Using the MZI-based error network modeling method, an offset weight matrix W is established under real hardware conditions. A noise-aware differentiable loss function is constructed to update the architecture parameters, forming a noise-aware search algorithm. This algorithm is used to design the photonic convolutional network structure and is mapped to an information loading module 201, multiple linear processing modules 202, and multiple nonlinear processing modules 203, forming N... L A base learner module 20 based on noise-aware search is shown in the schematic diagram below. Figure 2 As shown, N is used to perform S1. L There are N independent simple subtasks with N categories; where:

[0045] The information loading module 201 is used to load one-dimensional light field information of a sample through a single MZI.

[0046] The linear processing module 202 is used to perform convolution operations through an on-chip MZI array composed of multiple DCs and multiple PSs and an on-chip semiconductor optical amplifier array.

[0047] The on-chip MZI array consists of two unitary matrix modules and one diagonal matrix module. It is used to convert the convolution operation into a general matrix-vector multiplication through image transposition and to decompose the matrix into the product of a unitary matrix U, a diagonal matrix S, and a unitary matrix V through singular value decomposition to achieve arbitrary convolution operation. Furthermore, the Clements-type matrix decomposition method is used to map the elements of the unitary matrix to the phase change parameters of the PSs of the MZI array.

[0048] Figure 3 This is a schematic diagram of the MZI configuration for the linear processing module. Specifically, a 5×4 matrix can be decomposed into a 5×5 unitary matrix, with the main diagonal containing a diagonal matrix of 4 elements and a 4×4 unitary matrix, and physically mapped to 19 MZIs arranged in a rectangular pattern. A semiconductor optical amplifier is used at each output port to compensate for optical loss.

[0049] The nonlinear processing module 203 is used to convert the optical calculation results into electrical signals by using a photodetector, output the modified linear unit activation function response, and convert it into a one-dimensional optical signal by using an electro-optic intensity modulator and PSs.

[0050] It should be noted that the linear processing module 202 and the nonlinear processing module 203 are collective terms for modules with the same processing function. Figure 2 The multiple linear processing modules 202 and multiple nonlinear processing modules 203 are identical in functional logic, but differ in hardware implementation.

[0051] For the decomposed subtasks, a larger scale can enhance the representational ability of the neural network, but at the same time, the accumulation of optical loss and noise will reduce computational performance. To balance the contradiction between network size and system error, this invention proposes a noise-aware search algorithm. This algorithm embeds optical loss and noise impact as core optimization constraints into the architecture search loop, dynamically balancing model capacity and noise resistance, thereby automatically converging to find the neural network structure with optimal performance in a real hardware environment.

[0052] This example constructs a hypernetwork containing 5 directed edges, each with 3 candidate operations: 3×3 convolution, 5×5 convolution, and skip connections. The architecture parameters are defined as the probability of selecting each candidate operation, and the weight parameters are defined as the weights within each candidate operation. The input information is processed through multiple candidate operations simultaneously, and the feature maps obtained from each operation are summed according to their corresponding probabilities based on the architecture parameters to generate a new mixed feature map. However, to reduce memory usage, this example uses a binarized path to save memory, sampling and activating a single candidate operation according to the architecture parameters.

[0053]

[0054] In the formula Indicates the output of the mixed operation. Let p be the set of candidate operations. i Indicates candidate operation o i The probability of selection.

[0055] Take a column vector from the random sampling encoding matrix M, redistribute the dataset labels, and divide the training set, validation set, and test set in a 9:1:2 ratio. Figure 4 This is a flowchart of the noise-aware search algorithm. At the beginning of the algorithm, the network architecture parameters are initialized using a uniform distribution, and the weight parameters are initialized using a Gaussian distribution. Furthermore, an 80-round warm-up training strategy is employed, updating the weight parameters with a learning rate of 0.05 to avoid drastic gradient fluctuations interfering with the architecture parameters and ensuring efficient convergence in later stages.

[0056] Figure 5 This is a schematic diagram of the hypernetwork parameter update. During the weight parameter training phase, the architecture parameters are kept frozen. Binary gating is sampled according to formula (12), and the weight parameters of the activation path are updated on the training set using gradient descent based on the standard cross-entropy loss function. Conversely, during the architecture parameter training phase, the weight parameters are kept frozen. After resetting the binary gating, the architecture parameters are updated on the validation set using gradient descent based on the noise-aware differentiable loss function.

[0057] Specifically, to incorporate the impact of optical loss and noise in the hardware on computational accuracy as differentiable constraints into the architecture search process, such as... Figure 6 As shown, an error network modeling method based on MZI is further proposed.

[0058] For example, taking into account the propagation loss in the waveguide, the waveguide transfer matrix is ​​obtained:

[0059]

[0060] In the formula α p l represents the propagation loss per unit distance. MZI α represents the waveguide length of MZI. p The value is 2dB / cm, l MZI The value is 300μm.

[0061] For example, considering the insertion loss of the DC and the crosstalk noise caused by manufacturing errors in the splitting ratio, the transition matrix of the DC is modeled as follows:

[0062]

[0063] In the formula α DC Let κ represent the insertion loss of DC, σ represent the cross-coupling coefficient of DC and follow a Gaussian distribution, and σ represent the insertion loss of DC. dc α represents the standard deviation of the cross-coupling coefficient of DC.DC The value is 0.13dB, σ dc The value is 0.01.

[0064] For example, considering the metal absorption loss of the microheater in the PS, the phase setting deviation caused by finite bit precision, and the quadratic relationship between the phase setting and the applied voltage, the transition matrix of the PS is modeled as follows:

[0065]

[0066] In the formula α m σ represents the metal absorption loss of the PS, γ represents the actual phase setting value considering quantization error and manufacturing defects, γ follows a Gaussian distribution, and σ ps α represents the standard deviation of the phase setting of the PS, γ0 represents the nominal phase setting value determined by a finite bit-bit resolution, and α m The value is 0.23 dB, σ ps The value of B is 0.01, and the value of B is 8.

[0067] Furthermore, we can obtain the 2×2 unit-level MZI transfer matrix T under the influence of optical loss and noise. MZ :

[0068]

[0069] The ideal weight matrix W0 is obtained by training using the machine learning backpropagation algorithm. The corresponding unitary matrix elements are then obtained using singular value decomposition (SVD), and the Clements-type matrix decomposition method is used to map these unitary matrix elements to the phase change parameters of the PSs of the MZI array. Based on the aforementioned MZI-based error network modeling method, the offset weight matrix W under the influence of optical loss and noise can be obtained. In this example, RVD is used as a performance metric to measure the difference between W and the ideal weight matrix W0.

[0070]

[0071] In the formula W m,n , Let W and W0 represent the elements in the m-th row and n-th column, respectively.

[0072] Furthermore, RVD is used to measure the impact of optical loss and noise on computational accuracy, and is embedded as a differentiable constraint term in the standard loss function to obtain a noise-aware differentiable loss function:

[0073]

[0074]

[0075] Loss in the formula CE Represents the standard cross-entropy loss function. This represents a weight decay term to prevent overfitting. Let λ1 and λ2 represent the expected value of the network's RVD, and both λ1 and λ2 represent scaling factors. This represents the expected value of RVD in the i-th network layer. This represents the probability of selecting the j-th candidate operation in the i-th network layer. Let represent the RVD of the j-th candidate operation in the i-th network layer.

[0076] After completing the preset training cycle, the final photonic convolutional network structure is determined based on the current architecture parameters. It is understandable that, following the balance principle in constructing the encoding matrix M, N... L Each base learner module based on noise-aware search can use the same photonic convolutional network structure.

[0077] S3. For each input sample, based on N described in S2 L The N outputs of the base learner modules based on noise-aware search L Bit-N value encoding, calculating N for each category in the encoding matrix M described in S1. L The Hamming distance is encoded using a standard N-valued encoding, and the standard class corresponding to the minimum distance is output to form a synthesis module 30, which is used to reconstruct the final large-scale convolutional neural network with N classes. C The classification results of complex tasks.

[0078] For example, for each sample in the Cifar10 dataset, each base learner module outputs the three-class classification results of the subtask in parallel, synthesizing the final 50-bit ternary code. The Hamming distance between the synthesized code of the sample and the standard ternary code of each class in the encoding matrix M is calculated, and the standard class corresponding to the minimum distance is taken as the final classification result, completing the overall computation task.

[0079] It should be noted that: (1) The range of N is 2 ≤ N < N C When N is smaller, the complexity of the corresponding subtask is lower, the size of the photonic convolutional network structure obtained based on noise-aware search is smaller, and the light loss and noise impact is smaller, but the number of base learners required is higher, resulting in higher hardware resource consumption, and vice versa. (2) N L The range of values ​​is When N L The larger the value of , the greater the distance between each category code, thus improving classification accuracy at the cost of hardware resource consumption.

[0080] As can be seen from the above embodiments, the large-scale distributed photonic convolutional neural network architecture design method based on noise-aware search proposed in this patent can be used to construct flexible and variable network structures, effectively suppress light loss and noise accumulation, improve network scale and computing power, and thus support complex artificial intelligence task scenarios.

Claims

1. A method for designing a large-scale distributed photonic convolutional neural network architecture based on noise-aware search, characterized in that, Includes the following steps: S1. By analyzing the large-scale convolutional neural network, the number of classifications is N. C For each category of complex tasks, perform N L The N-valued encoding of the bits is used as row vectors to form an encoding matrix M, which forms a decomposition module used to decompose complex tasks into N... L The system consists of N independent, simple subtasks for each category, forming a distributed architecture; the value of N is in the range of 2 ≤ N < N. C The N L The range of values ​​is S2. Using an error network modeling method based on Mach-Zehnder interferometry modules, an offset weight matrix W is established under real hardware conditions. A noise-aware differentiable loss function is constructed to update the architecture parameters, forming a noise-aware search algorithm. This algorithm is used to design the photonic convolutional network structure and mapped to an information loading module, a linear processing module, and a nonlinear processing module, forming N. L A base learner module based on noise-aware search is used to perform the N described in S1. L There are N independent simple subtasks with N categories; where: The information loading module is used to load one-dimensional light field information of a sample through a single Mach-Zehnder interferometer module. The linear processing module is used to perform convolution operations through an on-chip Mach-Zehnder interferometer module array composed of multiple directional couplers and multiple thermo-optical phase shifters, and an on-chip semiconductor optical amplifier array. The on-chip Mach-Zehnder interferometer module array consists of two unitary matrix modules and one diagonal matrix module. It is used to convert the convolution operation into a general matrix-vector multiplication through image transposition and to decompose the matrix into the product of a unitary matrix U, a diagonal matrix S, and a unitary matrix V through singular value decomposition to achieve arbitrary convolution operation. Furthermore, the Clements-type matrix decomposition method is used to map the elements of the unitary matrix to the phase change parameters of the thermo-optical phase shifter of the Mach-Zehnder interferometer module array. The nonlinear processing module is used to convert optical calculation results into electrical signals by using a photodetector, output the response of the modified linear unit activation function, and convert it into a one-dimensional optical signal by an electro-optic intensity modulator and a thermo-optic phase shifter. S3. For each input sample, based on N described in S2 L The N outputs of the base learner modules based on noise-aware search L Bit-N value encoding, calculating N for each category in the encoding matrix M described in S1. L The Hamming distance is encoded using a standard N-valued encoding. The standard class corresponding to the minimum distance is output to form a synthesis module, which is used to reconstruct the final large-scale convolutional neural network with N classes. C The classification results of complex tasks; The decomposition module breaks down the global task into multiple independent subtasks based on the encoding. The base learner module based on noise-aware search searches and trains convolutional network structures for each subtask. The synthesis module decodes the outputs of each base learner module based on the encoding of the decomposition module.

2. The architecture design method according to claim 1, characterized in that, The noise-aware search algorithm includes the following steps: S1. Construct a hypernetwork, which is defined as a directed acyclic graph containing n directed edges. Each directed edge is defined as a hybrid operation containing K parallel candidate operations. S2. Define the architecture parameters as the probability of selecting each candidate operation, and define the weight parameters as the weights within each candidate operation; S3. To reduce memory usage, a binary gate g is used to define the output of the mixed operation: In the formula Indicates the output of the mixed operation. Let p be the set of candidate operations. i Indicates candidate operation o i The probability of choosing; S4. Divide the dataset into training set, validation set and test set according to the proportion, and reallocate the N-valued labels according to the encoding matrix M; S5. Using an error network modeling method based on Mach-Zehnder interferometry modules, a noise-aware differentiable loss function for updating architecture parameters is constructed and modeled as follows: In the formula, Loss represents the differentiable loss function for noise perception. CE Represents the standard cross-entropy loss function. This represents a weight decay term to prevent overfitting. λ1 and λ2 represent the expected value of the network's relative offset distance, respectively, and both represent scaling factors. This represents the expected value of the relative offset distance of the i-th network layer. This represents the probability of selecting the j-th candidate operation in the i-th network layer. W represents the relative offset distance of the j-th candidate operation in the i-th network layer, W represents the offset weight matrix, and W0 represents the ideal weight matrix. m,n , These represent the elements in the m-th row and n-th column of W and W0, respectively. S6. Initialize the architecture parameters using a uniform distribution and the weight parameters using a Gaussian distribution. S7. Use gradient descent to train the architecture parameters and weight parameters alternately, and stop training after the iterative constraints are met; wherein, the weight parameters are updated using the training set based on the standard cross-entropy loss function, and the architecture parameters are updated using the validation set based on the noise-aware differentiable loss function. S8. Determine the photonic convolutional network structure based on the architecture parameters obtained after training.

3. The algorithm steps according to claim 2, characterized in that, The error network modeling method based on the Mach-Zehnder interferometer module establishes a 2×2 unit-level transfer matrix T in a real hardware environment by introducing optical loss, crosstalk noise, and quantization noise error terms. MZ The offset weight matrix W under network-level error is reconstructed based on Clements-type matrix decomposition, singular value decomposition, and image transpose method, where the 2×2 unit-level transfer matrix T MZ It can be modeled as: In the formula T PL Let T be the waveguide transfer matrix. DC Let T be the transition matrix of the directional coupler. PS Let α be the transfer matrix of the thermo-optical phase shifter. p l represents the propagation loss per unit distance. MZI α represents the waveguide length of the Mach-Zehnder interferometer module. DC σ represents the insertion loss of the directional coupler, κ represents the cross-coupling coefficient of the directional coupler, and σ represents the insertion loss of the directional coupler. dc α represents the standard deviation of the cross-coupling coefficient of the directional coupler. m γ represents the metal absorption loss of the thermo-optical phase shifter, γ represents the actual phase setting value considering quantization errors and manufacturing defects, and σ represents the actual phase setting value. ps κ represents the standard deviation of the phase setting of the thermo-optical phase shifter, γ0 represents the nominal phase setting value determined by the finite bit resolution, and both κ and γ follow a Gaussian distribution.

Images