A fast and high-performance online training method for optical diffraction neural network devices

Abstract

The present invention discloses a fast, high-performance online training method for an optical diffraction neural network device. The method comprises the following steps: first, using a knowledge distillation module to pre-train a simulated diffraction neural network model and assigning the pre-trained parameters to a physical diffraction neural network device; second, using a domain transfer module to estimate the error in the pre-trained parameters; the domain transfer module compares the output domain of the physical diffraction neural network device with the output domain of the simulated diffraction neural network model to estimate the output error of the physical diffraction neural network device caused by pre-trained parameter mismatch; finally, using the domain transfer module to generate a new learning domain, adding a re-distillation module to fine-tune the parameters of the adaptive simulated diffraction neural network model, and re-assigning the trained parameters to the physical diffraction neural network device. This method ensures that optical diffraction neural networks can more stably and efficiently implement artificial intelligence computing tasks in real-world scenarios, such as image recognition.

Description

Technical Field

[0001] The present invention relates to a method for rapidly training an actual physical model system in the field of optical computing, and in particular to a rapid and high-performance online training method for an optical diffraction neural network device. Background Art

[0002] Deep learning is one of the fastest-growing areas in artificial intelligence. Models derived from gradient descent and backpropagation algorithms can accomplish a wide range of visual tasks, including image classification, speech recognition, object detection and segmentation. However, as deep learning models grow in size, parameters, and structure, their inference processes place increasing demands on hardware computing power. Under limited conditions, these demands are limited to tasks requiring low latency, low power consumption, and high bandwidth. Furthermore, the hardware computing power required for neural network models is constrained by Moore's Law and is expected to reach a bottleneck in the near future.

[0003] Thanks to the proliferation of optoelectronic components on integrated optoelectronic platforms and the continuous advancement of optoelectronic technology in recent years, the concept of building optical machine learning architectures—optical neural networks—with optical components has become a reality. Compared to traditional electrical machine learning models, the propagation-as-computation nature of optical neural networks enables computational speeds approaching the speed of light. Furthermore, the structure of optical networks makes them well-suited for large-scale parallel computing, significantly improving their energy efficiency. Currently, optical networks have found applications in image recognition, imaging, and matrix calculations, becoming a key approach to solving electrical machine learning model problems.

[0004] Diffractive neural networks are a type of optical network that uses the diffraction propagation of light as a mechanism for information propagation in deep learning. Diffractive neural networks are composed of multiple layers of diffractive surfaces that modulate the amplitude and phase of light waves, thereby extracting features from the optical signal and implementing intelligent functions such as image classification and object detection at the output layer. Model training is a crucial component of deep learning. For diffractive neural networks, training begins by simulating the physical structure of a network composed of various optical components on a device. This simulated network is then trained on the device using a deep learning algorithm, continuously updating its parameters. Finally, the trained simulated network parameters are deployed to various physical optical components, completing the training of the diffractive neural network. However, the optical network model simulated on the device is not entirely consistent with the actual physical structure. This is because the process of building an optical neural network using optical components inevitably introduces various errors, such as optical component displacement, noise, and cascade errors between components. Consequently, when the simulated network parameters are applied to the physical system, these deviations in model construction can cause the performance of the physical optical network to degrade compared to the ideal case during task inference, resulting in performance differences from the simulated network on the device. Summary of the Invention

[0005] In order to overcome the problem of significant performance degradation caused by various systematic errors in actual optical diffraction neural network devices, the present invention proposes a fast and high-performance online training method for optical diffraction neural network devices, so that optical diffraction neural network devices with systematic errors can quickly achieve high-accuracy calculations.

[0006] The technical solutions adopted in the present invention are as follows:

[0007] A fast and high-performance online training method for a light diffraction neural network device comprises the following steps:

[0008] 1) Pre-training a simulated diffraction neural network model using a knowledge distillation training module to obtain an output domain of the simulated diffraction neural network model; in the knowledge distillation training module, a convolutional neural network in the electrical domain serves as a teacher network, and the simulated diffraction neural network model serves as a student network; the simulated diffraction neural network model is a simulation model that simulates a physical diffraction neural network device;

[0009] 2) assigning the pre-trained model parameters to the physical diffraction neural network device to obtain the output domain of the physical diffraction neural network device; using the domain transfer module to compare the output domain of the physical diffraction neural network device with the output domain of the simulated diffraction neural network model, estimate the system error of the physical diffraction neural network device, and update the model parameters in the domain transfer module;

[0010] 3) Use the updated domain transfer module to generate a new learning domain, establish an adaptive simulated diffraction neural network model with the same structure and parameters as the simulated diffraction neural network model, use the re-distillation module to fine-tune the parameters of the adaptive simulated diffraction neural network model, and re-assign the trained parameters to the physical diffraction neural network device.

[0011] Furthermore, in the knowledge distillation training module, first, the image is input into the teacher network and the student network respectively, and then the outputs of the teacher network and the student network are respectively imported into the softmax function with temperature to obtain the teacher soft label and the student soft prediction; the softmax function with temperature obtains richer information differences between categories by adjusting the temperature parameters, and calculates the difference between the student soft prediction and the teacher soft label to obtain the distillation loss; the difference between the soft prediction of the student network and the true label of the data set is calculated to obtain the category loss; finally, the distillation loss and the category loss are weighted and summed to obtain the final loss, and the parameters of the student network are continuously updated to obtain the pre-trained simulated diffraction neural network model.

[0012] Furthermore, the teacher network adopts the trained ResNet18 residual convolutional neural network, which includes a convolutional layer, two pooling layers, eight residual modules and a fully connected layer, wherein each residual module includes two convolutional layers, and there is a connection across modules in each residual module; the fully connected layer has the number of output nodes equal to the number of data set categories, which is used to finally output the results of the teacher network.

[0013] Furthermore, the physical diffraction neural network device includes an input layer, a diffraction processing layer and an output layer; the light wave is incident perpendicular to the input plane, and after being encoded by the input layer, it diffracts backward with the input information and enters the diffraction processing layer; the diffraction processing layer is composed of a plurality of diffraction planes composed of a plurality of physical structures, and there are a plurality of processing units on each diffraction plane to process the information carried by the light wave, each of which can be regarded as a neuron node on the diffraction processing layer, and the light wave is modulated in amplitude and phase by the processing unit on the surface of the diffraction processing layer, and then propagates to the next plane through free space, passing through layer by layer, and finally reaches the output layer after being processed by multiple diffraction layers; a detector is arranged above the output layer to detect the intensity of the light field reaching the output layer.

[0014] Furthermore, the optical signal transmitted in the simulated diffraction neural network model is represented by a tensor, and the diffraction process of light in free space is calculated using the angular spectrum method.

[0015] Furthermore, the domain transfer module includes a multi-layer perceptron module and a U-shaped network module, which divides the different output labels and learning labels of the simulated diffraction neural network model and the physical diffraction neural network device into four domains: actual output domain, actual learning domain, ideal output domain and ideal learning domain, which respectively represent the output and learning distribution under different situations; among them, the direct output of the physical diffraction neural network device is called the actual output domain. Due to the existence of systematic errors, the data distribution of this domain is quite different from the ideal distribution; the output domain of the teacher network is called the actual learning domain, which is the object learned by the simulated diffraction neural network model during initialization training; the output of the simulated diffraction neural network model is the ideal output domain, which represents the output of the network model under ideal conditions without systematic errors; the ideal learning domain is the domain obtained by estimating the systematic error, that is, the domain to be solved; the input of the domain transfer module is the output of the simulated diffraction neural network model, that is, the ideal output domain. The output result of the domain transfer module is compared with the output of the teacher network, that is, the actual learning domain, and the loss is calculated, and the model parameters in the domain transfer module are updated.

[0016] Furthermore, the multilayer perceptron module includes an input layer, an output layer, and a hidden layer in the middle. A batch normalization layer and an activation function layer are provided after the hidden layer. Similarly, an activation function layer is also provided after the output layer. The multilayer perceptron module is used to process the output value of the output layer of the physical diffraction neural network device. Therefore, the number of nodes in the input layer is equal to the number of categories in the data set.

[0017] Furthermore, the U-shaped network module includes a convolution module, a downsampling module, a deconvolution module and a splicing module, wherein the convolution module includes three convolution operations, and each convolution operation is followed by a ReLU activation function; the downsampling module includes a convolution operation with a step size of 2 and a ReLU activation function operation; the deconvolution module includes a deconvolution operation and a ReLU activation function operation; the splicing module splices the two tensors in the channel direction.

[0018] Furthermore, in the step 3), the output domain of the simulated diffraction neural network model in step 2) is input into the updated domain migration module to obtain a new learning domain; the output of the adaptive simulated diffraction neural network model is compared with the new learning domain, the difference is calculated, and the loss is obtained, thereby updating the parameters of the adaptive simulated diffraction neural network model.

[0019] This method utilizes an electrical domain neural network as an aid. Its domain migration module estimates the errors of the physically deployed network using the output domains of the physical diffraction neural network device and the simulated diffraction neural network model. Based on these errors, it regenerates a new adaptive learning domain. By learning from this new learning domain, the simulated diffraction neural network model adjusts its parameters in real time and redeploys these adjusted parameters to the physical system. Through an efficient training process, the model can rapidly adjust, overcoming the significant accuracy drop caused by system errors.

[0020] The beneficial effects of the present invention are mainly manifested in:

[0021] (1) Compared with the current mainstream diffraction neural network training method, the present invention designs a new type of fast and high-performance online training method for optical diffraction neural network devices. The error of the diffraction neural network device is estimated by using the methods of knowledge distillation and domain transfer, so as to update the parameters of the network, ensuring that the device can more stably and perform high-performance artificial intelligence computing tasks such as image recognition in real scenarios.

[0022] (2) The training speed of this method is very fast. The error estimation and model parameter adjustment process are fast and efficient. The training process based on the method of the present invention can be completed in 5 to 10 minutes. At the same time, the ideas of this method can also be extended to other physical neural networks, with good portability and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is an overall flow chart of the method of the present invention;

[0024] Figure 2 It is a flowchart of implementation details in the present invention;

[0025] Figure 3 is a flow chart of the domain migration process in the present invention;

[0026] Figure 4 is a flow chart of updating the domain migration module in the present invention;

[0027] Figure 5 is a flow chart of the redistillation module of the present invention;

[0028] Figure 6 is a structural diagram of the multi-layer perceptron in the domain migration module of the present invention;

[0029] Figure 7 It is a structural diagram of the U-shaped network in the domain migration module of the present invention. DETAILED DESCRIPTION

[0030] In order to more clearly illustrate the purpose and technical solutions of the present invention, the present invention is further described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 and Figure 2 As shown, Figure 1 This is a rough flow chart. Figure 2 The following is an overview of the implementation details of the entire invention method. The present invention includes the following steps:

[0032] 1) Pre-train the simulated diffraction neural network model through the knowledge distillation training module, such as Figure 3 As shown, the convolutional neural network in the electric domain serves as the teacher network, and the simulated diffraction neural network model serves as the student network. Through this training process, the parameters of the pre-trained simulated diffraction neural network model are obtained. The simulated diffraction neural network model is a simulation model that simulates a physical diffraction neural network device. The light signal transmitted in this model is represented by a tensor, and the diffraction process of light in free space is calculated using the angular spectrum method. The simulated physical diffraction neural network device includes an input layer, a diffraction processing layer, and an output layer. The input layer can be composed of a laser and a digital micromirror, the diffraction processing layer can be composed of a spatial light modulator and a lens, and the output layer can be composed of an optical precision camera.

[0033] A light wave is incident perpendicular to the input plane. After being encoded by the input layer, it carries the input information and diffracts backward, entering the diffraction processing layer. The diffraction processing layer is composed of several diffraction planes with physical structures. Each diffraction plane has several processing units that process the information carried by the light wave. Each processing unit can be considered a neuron node in the diffraction processing layer. The light wave undergoes amplitude and phase modulation in the processing units on the surface of the diffraction processing layer. It then propagates through free space to the next plane, passing through layers. Finally, after being processed by multiple diffraction layers, the light wave reaches the output layer. For deep learning tasks such as image classification, the input of the physical diffraction neural network device is the light information of the image encoded by the input plane, and the final inference result corresponds to the category number of the image. A detector is placed above the output layer to detect the intensity of the light field reaching the output layer. The detector is divided into several equal-sized regions. The detector calculates the light intensity of each region. The number of the detection region with the maximum intensity light signal is the final image classification inference result.

[0034] In the knowledge distillation training module, the selected electric domain teacher neural network is a type of residual convolutional neural network, ResNet18. The network is trained on a specific data set (MNIST handwritten digit data set) to obtain a teacher network with high accuracy. In this embodiment, the ResNet18 residual convolutional neural network includes a convolutional layer, two pooling layers, eight residual modules and a fully connected layer, wherein each residual module includes two convolutional layers, and there is a connection across modules in each residual module. For the fully connected layer, the number of its output nodes should be equal to the number of data set categories, which is used to finally output the results of the teacher network. After the input image is scaled to the appropriate size, it is input into the teacher network and the student network respectively, and the outputs of the teacher network and the student network are respectively imported into the softmax function with temperature to obtain the teacher soft label p and the student soft prediction q. The calculation formula corresponding to this step is:

[0035]

[0036] where z i Represents the output of the i-th dataset sample in the same batch in the teacher network or student network. T represents the temperature value. A higher temperature value can make the semantic information difference between different categories richer. The difference between the student soft prediction p and the teacher soft label q is calculated to obtain the distillation loss. The loss used here is the KL divergence loss. At the same time, the difference between the soft prediction p of the student network and the true label g of the dataset is calculated to obtain the category loss. The cross entropy loss CE is used here. The calculation formulas for the two losses are as follows:

[0037]

[0038]

[0039] Among them, m represents the number of categories in the dataset, and i represents the subscript value corresponding to the element in the vector output by the student network and the teacher network or the label value vector.

[0040] The distillation loss and the category loss are weighted and summed to obtain the final loss result, and the parameters of the student network are continuously updated to obtain the pre-trained simulated diffraction neural network model. 2) The parameters of the pre-trained simulated diffraction neural network model are assigned to the physical diffraction neural network device to obtain the output domain of the physical diffraction neural network device. The domain transfer module estimates the system errors of the physical diffraction neural network device, such as assembly error, alignment error, and manufacturing error, by comparing the output domain of the physical diffraction neural network device with the output domain of the simulated diffraction neural network model. Figure 4As shown in Figure 1, domain transfer is the most important part of the entire fast and high-performance online training method. The pre-trained parameters of the simulated diffraction neural network model are deployed on the physical diffraction neural network device, which will bring about large errors. For the different output labels and learning labels of the simulated and physical diffraction neural networks, these label distributions are divided into four domains, representing the output and learning distributions in different situations, namely the actual output domain, the actual learning domain, the ideal output domain and the ideal learning domain; among them, the direct output of the physical diffraction neural network device is called the actual output domain Z e , the data distribution of this domain is quite different from the ideal distribution. The output of the teacher network is called the actual learning domain Z t , which is the object learned by the optical network during initialization training. The output of the simulated diffraction neural network model is the ideal output domain. In an ideal situation, the output of the model should be consistent with the learned value, that is, the actual learning domain Z t Therefore, the ideal output domain here can be regarded as roughly equivalent to the actual learning domain output by the teacher network. l , is the domain obtained by estimating the system error, which is the domain to be solved in this training method. Ideally, the output of the student network should be the same as that of the teacher network:

[0041] M[θ m , x i ]=T(x i )=Z t

[0042] Where M represents the trained student simulated diffraction neural network model, T represents the teacher network, θ m represents the parameters learned by the student network after distillation, x i is the i-th image of the input. When the parameters are deployed in the physical diffraction neural network device, we have:

[0043] E{M[θ m , x i ]}=Z e

[0044] Where E represents the various errors in the system. Because of the error E in the physical diffraction neural network device, the output of the model is different from the actual learning domain. The role of the domain transfer module is to estimate the error E and change the learning parameter θ in the model. m , the final result should be:

[0045] E{M[θ′ m , x i ]}=Z t

[0046] The error E can be regarded as a mapping process. By fitting this mapping, the error is estimated. Assuming that the error E exists, the mapping from the output domain to the learning domain is f, then:

[0047] Z t =f(Z e )

[0048] After completing the fitting process for the mapping f, the desired ideal learning domain should be:

[0049] Z l =f(Z t )

[0050] The domain transfer module is responsible for fitting the above mapping f, specifically, fitting the output of the physical diffraction neural network device, that is, the mapping process to the actual learning domain. Figure 4 The domain transfer module consists of two parts. One part is the multi-layer perceptron MLP, which includes an input layer, an output layer and a hidden layer, as shown in Figure 6 As shown, the input layer and the output layer have a total of 10 nodes, and the middle hidden layer uses 40 nodes in this embodiment. In order to speed up the training of the model and make the model converge faster, the BatchNorm layer is used after the hidden layer. In addition, the sigmoid activation function layer is used after the hidden layer and the output layer. The input data processed by the multi-layer perceptron module is the electrical signal data of the output layer detector of the physical diffraction neural network device. The other part of the domain migration module is a U-shaped network, including a convolution module, a downsampling module, a deconvolution module and a splicing module, as shown in FIG. Figure 7 As shown. Among them, the convolution module contains three convolution operations, and each convolution operation will pass through the ReLU activation function. For the downsampling module, it includes a convolution operation with a step size of 2 and a ReLU activation function operation. For the deconvolution module, it includes a deconvolution operation and a ReLU activation function operation. For the splicing module, it splices two tensors in the direction of the channel. The data processed by the U-shaped network model is the light field plane information of the output layer of the physical diffraction neural network device. Taking a certain test data as an example, the input and output of the network are both 200×200 tensors. Since the label values ​​in the output domain and the learning domain are one-to-one corresponding, these label pairs constitute the training set of the domain transfer module. The loss function used by the training domain transfer module is L1 loss, and the calculation formula is as follows:

[0051] L=|MLP(Z elogit )-Z tlogit |+|UNet(Z egraph )-Z tgraph |

[0052] where Z elogit and Z tlogitRepresent the label value vectors of the actual output domain and the ideal output domain respectively, and Z egraph and Z tgraph The light field plane value tensors represent the actual output domain and the ideal output domain, respectively. By calculating the loss and updating the parameters of the domain transfer module, the trained domain transfer module can reversely map the known ideal output domain to the desired ideal learning domain.

[0053] 3) Use the domain transfer module to generate a new learning domain, add the re-distillation module to fine-tune the parameters of the simulated diffraction neural network model, and re-assign the trained parameters to the diffraction neural network device in the physical domain, such as Figure 5 As shown in the figure, the output of the simulated diffraction neural network model is input into the domain transfer module to obtain a new learning domain. An adaptive simulated diffraction neural network model is established, whose structure and parameters are the same as the simulated diffraction neural network model pre-trained in the first step. The adaptability of the adaptive simulated diffraction neural network model is reflected in its ability to quickly adjust to the actual error environment. The output of the adaptive simulated diffraction neural network model is compared with the new learning domain, and the difference is calculated to obtain the loss, which is used to update the parameters of the adaptive simulated diffraction neural network model. In the previous step, the output of the trained domain transfer module is set as the new label. The domain transfer module, which carries error estimation information, plays a key role in the re-distillation process. The re-distillation process is the process of readjusting the adaptive network parameters to adapt to the system error and can be considered as fine-tuning the adaptive student simulated diffraction neural network model. In this process, the output of the simulated diffraction neural network model is the ideal output domain u. Therefore, to obtain the ideal learning domain, the ideal output domain u is processed by the domain transfer module to obtain the re-distilled learning domain. Subsequently, the output v of the adaptive student simulated diffraction neural network model is compared with the output of the domain transfer module, the loss is calculated, and its own parameters are adjusted. The loss function of the re-distillation process also consists of two parts. The distillation loss comes from the difference between the output of the adaptive student simulated diffraction neural network model and the output of the domain transfer module. The loss type used is KL loss. The category loss comes from the difference between the output of the adaptive student simulated diffraction neural network model and the true label g. The loss type used is cross entropy loss. The calculation formula is as follows:

[0054]

[0055] Among them, α and β are balance parameters used to determine the ratio of the two losses, where L kd , L clsDenote the distillation loss and classification loss, respectively. m represents the number of samples in the dataset, i represents the label value corresponding to the sample, and dtm represents the processing performed by the domain transfer module. The parameters of the adaptive student diffraction network, adjusted through re-distillation, were deployed in a physical device, resulting in a diffraction neural network system capable of stably and efficiently performing real-world AI computing tasks such as image recognition.

Claims

1. A fast and high-performance online training method for a light diffraction neural network device, characterized in that: The following steps are involved: 1) Pre-training a simulated diffraction neural network model using a knowledge distillation training module to obtain an output domain of the simulated diffraction neural network model; in the knowledge distillation training module, a convolutional neural network in the electrical domain serves as a teacher network, and the simulated diffraction neural network model serves as a student network, wherein the inputs of the teacher network and the student network are both images; and the simulated diffraction neural network model is a simulation model that simulates a physical diffraction neural network device; 2) assigning the model parameters obtained by pre-training to the physical diffraction neural network device to obtain the output domain of the physical diffraction neural network device; comparing the output domain of the physical diffraction neural network device with the output domain of the simulated diffraction neural network model through the domain migration module, estimating the system error of the physical diffraction neural network device, and updating the model parameters in the domain migration module; the domain migration module includes a multi-layer perceptron module and a U-shaped network module, the multi-layer perceptron module includes an input layer, an output layer, and an intermediate hidden layer, a batch normalization layer and an activation function layer are provided after the hidden layer, and an activation function layer is also provided after the output layer; the U-shaped network module includes a convolution module, a downsampling module, a deconvolution module, and a splicing module; 3) Use the updated domain transfer module to generate a new learning domain, establish an adaptive simulated diffraction neural network model with the same structure and parameters as the simulated diffraction neural network model, use the re-distillation module to fine-tune the parameters of the adaptive simulated diffraction neural network model, and re-assign the trained parameters to the physical diffraction neural network device.

2. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: In the knowledge distillation training module, first, the image is input into the teacher network and the student network respectively, and then the outputs of the teacher network and the student network are respectively imported into the softmax function with temperature to obtain the teacher soft label and the student soft prediction; the softmax function with temperature obtains the information difference between categories by adjusting the temperature parameter, and calculates the difference between the student soft prediction and the teacher soft label to obtain the distillation loss; the difference between the soft prediction of the student network and the true label of the dataset is calculated to obtain the category loss; finally, the distillation loss and the category loss are weighted and summed to obtain the final loss, and the parameters of the student network are continuously updated based on this to obtain the pre-trained simulated diffraction neural network model.

3. The fast and high-performance online training method for a light diffraction neural network device according to claim 1 is characterized in that: The teacher network adopts the trained ResNet18 residual convolutional neural network, which includes a convolutional layer, two pooling layers, eight residual modules and a fully connected layer, wherein each residual module includes two convolutional layers, and there is a connection across modules in each residual module; the fully connected layer has the number of output nodes equal to the number of data set categories, which is used to finally output the results of the teacher network.

4. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: The physical diffraction neural network device includes an input layer, a diffraction processing layer, and an output layer: the light wave is incident perpendicular to the input plane, and after being encoded by the input layer, it diffracts backward with the input information and enters the diffraction processing layer; the diffraction processing layer is composed of a plurality of diffraction planes composed of a plurality of physical structures, and each diffraction plane has a plurality of processing units that process the information carried by the light wave. Each of these processing units can be regarded as a neuron node on the diffraction processing layer. The light wave is modulated in amplitude and phase by the processing unit on the surface of the diffraction processing layer, and then propagates through free space to the next plane, passing through layer by layer. Finally, after being processed by multiple diffraction layers, the light wave reaches the output layer; a detector is provided above the output layer to detect the intensity of the light field reaching the output layer.

5. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: The optical signal transmitted in the simulated diffraction neural network model is represented by a tensor, and the diffraction process of light in free space is calculated using the angular spectrum method.

6. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: The domain transfer module divides the different output labels and learning labels of the simulated diffraction neural network model and the physical diffraction neural network device into four domains: actual output domain, actual learning domain, ideal output domain and ideal learning domain, which represent the output and learning distribution under different circumstances respectively; among them, the direct output of the physical diffraction neural network device is called the actual output domain. Due to the existence of systematic errors, the data distribution of this domain is different from the ideal distribution; the output domain of the teacher network is called the actual learning domain, which is the object learned by the simulated diffraction neural network model during initialization training; the output of the simulated diffraction neural network model is the ideal output domain, which represents the output of the network model under ideal conditions without systematic errors; the ideal learning domain is the domain obtained after estimating the systematic error, that is, the domain to be solved; the input of the domain transfer module is the output of the simulated diffraction neural network model, that is, the ideal output domain. The output result of the domain transfer module is compared with the output of the teacher network, that is, the actual learning domain, and the loss is calculated, and the model parameters in the domain transfer module are updated.

7. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: The number of nodes in the input layer of the multilayer perceptron module is equal to the number of data set categories.

8. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: In the U-shaped network module, the convolution module includes three convolution operations, and each convolution operation is followed by a ReLU activation function; the downsampling module includes a convolution operation with a step size of 2 and a ReLU activation function operation; the deconvolution module includes a deconvolution operation and a ReLU activation function operation; the splicing module splices the two tensors in the channel direction.

9. The fast and high-performance online training method for a light diffraction neural network device according to claim 1, characterized in that: In the step 3), the output domain of the simulated diffraction neural network model in step 2) is input into the updated domain migration module to obtain a new learning domain; the output of the adaptive simulated diffraction neural network model is compared with the new learning domain, the difference is calculated, and the loss is obtained, thereby updating the parameters of the adaptive simulated diffraction neural network model.

Images