An antenna cover agent model-oriented neural network architecture search method

Abstract

The application discloses a neural network architecture search method for a radome agent model, and comprises the following steps: inputting a training set, a verification set, a prediction set and search space parameters into a neural network model; training the neural network model; calculating a multi-task conflict matrix; calculating an indication matrix according to the multi-task conflict matrix; generating a modulation module according to the calculated indication matrix; judging whether the current neural network model meets a maximum model limit; adding the modulation module to the neural network model; training the newly-added modulation module using the training set to complete pre-training; outputting and saving the current neural network model architecture and parameters thereof; inputting a radome sample to be detected into the trained neural network model; generating a prediction value according to the read-in radome sample to be detected and outputting a prediction result. The application effectively alleviates gradient conflicts, flexibly realizes model architecture search and can be used in a subsequent detailed optimization process.

Description

Technical Field

[0001] This invention belongs to the field of radome technology, specifically relating to a neural network architecture search method for radome proxy models. Background Technology

[0002] The radome surrogate model technique is a method used to replace the complex electrical performance analysis of radomes, and it is crucial in radome design and optimization. To ensure rapid design decisions based on given requirements within the engineering cycle, multi-task learning (MTL) provides an effective approach by simultaneously solving multiple related tasks through knowledge sharing. However, a key obstacle to its application is the task interference problem, where different tasks compete with each other, leading to gradient conflicts and negative transfer during the optimization process.

[0003] In their paper "Gradient surgery for multi-task learning (Advances in Neural Information Processing Systems, 2020)," Yu et al. proposed a method to achieve a balance between competing tasks by manipulating gradients / loss. Gradient manipulation techniques involve adjusting the gradient direction, deleting parts of the gradient vector, and rotating shared features to ensure that the set gradient direction does not conflict with the gradient of any task. However, this ignores the Pareto stationary point property.

[0004] Yue et al. proposed a CoNAL method in their paper "Learning conflict-noticed architecture for multi-task learning (2023)" to alleviate gradient conflicts during architecture learning. They introduced specific (task-exclusive) modules as candidate search spaces. When gradient conflicts are detected during architecture learning, CoNAL adaptively switches to the task-exclusive module. However, this ignores the Pareto stationary point property and therefore may not be the optimal solution.

[0005] The main shortcomings of existing technologies lie in gradient conflict and task interference, especially in multi-task learning, where competition between different tasks leads to gradient conflicts and negative transfer during the optimization process. Although some methods attempt to address these issues, such as gradient operations and task-exclusive modules, they often neglect the Pareto stationary point property, thus potentially failing to reach the optimal solution. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, the present invention aims to provide a neural network architecture search method for radome proxy models. This method effectively alleviates gradient conflicts by considering the conflicts of the trained model and generating corresponding modulation modules using an indicator matrix, thereby flexibly realizing model architecture search and can be used in subsequent detailed optimization processes.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A neural network architecture search method for radome proxy models includes the following steps;

[0009] Step (1): Input the training set, validation set and prediction set into the neural network model in sequence to search for spatial parameters;

[0010] Step (2): Train the neural network model using the training dataset;

[0011] Step (3): Calculate the multi-task conflict matrix using the trained neural network model;

[0012] Step (4): Calculate the indication matrix based on the multi-task conflict matrix;

[0013] Step (5): Generate a modulation module based on the calculated indicator matrix;

[0014] Step (6): Determine whether the current neural network model meets the maximum model limit; if the number of parameters of the current model is lower than the limit, continue to the next step; if the number of parameters of the current model is equal to the limit, the current model construction process needs to be terminated and the process should proceed to step (9).

[0015] Step (7): Add the modulation module generated in step (5) to the neural network model;

[0016] Step (8): Train the newly added modulation module using the training dataset. After completing the pre-training, proceed to step (2).

[0017] Step (9): Output and save the current neural network model architecture and its parameters;

[0018] Step (10): Input the radome sample to be tested into the trained neural network model;

[0019] Step (11): Generate a predicted value based on the read-in radome sample and output the prediction result.

[0020] In step (1), the training dataset D is input into the neural network model sequentially. train Validate dataset D val Prediction dataset Dpre The data parameters, the shared layer structure f of the neural network S Modulation module h i,j The neural network structure parameters included maximum model capacity limits and gradient conflict thresholds; training parameters included training epochs, learning step size, and number of detections; and training, validation, and prediction sets were simulation data of the S21 transmission coefficient obtained from element frequency selection surface simulation of the radome using CST simulation software. The simulation settings parameters were substrate thickness and incident angle, ranging from 1-5 mm and 0-60 degrees, respectively, with the simulation result being the S21 transmission coefficient. Simulation data with a substrate thickness of 1-2 mm was used as the training set, 3-4 mm as the validation set, and 5 mm as the prediction set, which were then input into the neural network model sequentially. These elements collectively determined the model design, training process, and final performance.

[0021] In step (2), the training set and the shared layer network are initialized, the shared layer network is trained using the training set, and an early stopping strategy is used until convergence is achieved.

[0022] The specific steps for training the neural network in step (2) are as follows:

[0023] Using training set D train The model parameters are trained, and an early stopping mechanism is used to monitor its loss. The early stopping mechanism automatically terminates training if the loss does not improve after 50 training epochs.

[0024] In step (3), the multi-task gradient conflict matrix C is calculated using the validation set on the trained neural network model obtained in step (2), and the validation set D is used. val Calculate the gradient of backpropagation on the model trained in step (2), where the gradient with respect to task i is grad. i The gradient of task j is grad. j ;

[0025] The gradient conflict between task gradients is calculated using the following formula:

[0026] conflict = cos(grad) i ,grad j )

[0027] The multi-task gradient conflict matrix C is obtained by calculating the gradient conflict between each pair of tasks. C is a symmetric matrix.

[0028] Where the conflict matrix element C ij The calculation formula is as follows:

[0029] C ij =cos(grad i,grad j )

[0030] Wherein, cos(grad i ,grad j ) represents the cosine value between gradient vector i and gradient vector j. Its value represents the degree of conflict between task i and task j. The smaller the value, the greater the conflict.

[0031] Step (4) specifically involves: calculating the indicator matrix based on the multi-task gradient conflict matrix C and the gradient conflict threshold input in step (1). The specific calculations are as follows:

[0032]

[0033] in Representing the conflict threshold, D was used in the experiment. val If it is divided into multiple batches, there will be several indicator matrices for each batch. In multi-batch testing, the full validation set conflict matrix is ​​obtained by summing the indicator matrices, and the calculation formula is as follows:

[0034]

[0035] The final indicator matrix is ​​generated by employing either an average strategy or a maximum strategy. The formula for calculating the average strategy is as follows:

[0036]

[0037] Where I average Let I represent the average value of matrix I. The formula for calculating the maximum value is as follows:

[0038]

[0039] Where I top This represents the maximum value of matrix I. The significance of the average value strategy and the maximum value strategy is whether the number of modulation modules generated after each collision detection is half of the search space or only one modulation module is generated.

[0040] The specific steps (5) are as follows: Based on the indication matrix provided in step (4), a modulation module for the corresponding layer is generated to modulate gradient conflicts;

[0041] When two tasks conflict, the modulation module mitigates the conflict between task i and task j. The modulation module can use either highly fine-grained neurons or less fine-grained modules. Formally, when task i and task j conflict at layer l of the network, cos(g) i,l ,g j,l) < 0, where gi,l and gj,l represent the gradients of task i and task j at layer l, respectively;

[0042] According to the definition of the cosine between two gradient vectors:

[0043]

[0044] Two of the vectors conflict, which means that... <g i,l ,g j,l ><0, now assume tasks i and j share Q parameters at layer l, and make <g i,l ,g j,l >=-c, where c is a positive number. The adjustment module consists of P parameters, and the gradient g after adding the adjustment module will be... i,l g j,l Divided into two parts: and and The inner product of two gradient vectors is expressed as:

[0045]

[0046] Gradient conflict occurred based on the previous assumptions, and <g i,l ,g j,l >=-c, by increasing the dimensionality of the vectors in the two two-dimensional planes, the two gradients are folded so that they have a common descent direction, which alleviates gradient conflict and thus all tasks have a common descent gradient.

[0047] Step (6) specifically involves:

[0048] Based on the maximum model parameter limit in step (1), determine the maximum allowed number of parameters and compare the total number of parameters of the current model with the set maximum model parameter limit.

[0049] The relevant calculation formula for adding the modulation module to the neural network in step (7) is as follows:

[0050] The input / output calculation formulas for a fully shared layer are as follows:

[0051]

[0052] in Represents the input of layer l. f represents the output of layer l. S (·) represents the (l+1)th shared layer;

[0053] The input and output calculation formulas for the modulation module between task i and task j are as follows:

[0054]

[0055] in This represents the input of the l-th layer modulation module. This represents the output after modulation by the modulation module.

[0056] The specific method for the pre-trained modulation module in step (8) is as follows:

[0057] After adding the modulation module, freeze the parameters of the shared layer and the previously added modulation module, initialize the newly added module using the Xavier initialization strategy, and use the training set D. train Pre-train it using the same early stopping mechanism described in step (2).

[0058] The beneficial effects of this invention are:

[0059] This method first inputs a training set, a validation set, a test set, a conflict matrix threshold, an indicator matrix threshold, and corresponding search space parameters. Second, it trains the initial neural network by sharing a backbone and applying an early stopping strategy. Then, it uses the validation set to calculate the conflict matrix of the trained neural network. Next, it calculates the corresponding indicator matrix using the conflict threshold and generates the corresponding modulation module from the indicator matrix. Finally, it determines whether the maximum number of model parameters and the maximum number of iterations are satisfied. If satisfied, it outputs the corresponding model architecture and model parameters; otherwise, it adds the corresponding modulation module and trains until convergence.

[0060] This invention, when establishing a neural network-based surrogate model for the radome, mitigates gradient conflicts by introducing an indicator matrix and generating corresponding modulation modules, taking into account the conflicts in the trained model. By using different indicator matrix generation methods, different numbers of modulation modules can be generated at different stages, flexibly realizing model architecture search.

[0061] Traditional evolutionary or gradient-based algorithms (DARTs) search and train simultaneously. This leads to current performance weights not necessarily reflecting final performance, causing gradient-based neural network architecture search methods to favor simpler models, while evolutionary algorithm-based methods are typically slower and require more iterations. Our proposed gradient conflict-based neural network architecture search method offers low search costs and does not rely on the performance of the training model as a search criterion. It can adapt to any neural network model. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the training process of the method of the present invention.

[0063] Figure 2 This is a schematic diagram of the model search space provided by the present invention.

[0064] Figure 3 This is a schematic diagram of gradient processing provided by the present invention.

[0065] Figure 4 This is a comparison chart of the simulated curve and the predicted curve of the transmission coefficient of the present invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings.

[0067] Please see Figure 1 A neural network architecture search method for radome proxy models includes the following steps:

[0068] (1) Input the training set, validation set and prediction set into the neural network model in sequence, and search for spatial parameters;

[0069] The training dataset D was input into the neural network model sequentially. train Validate dataset D val Prediction dataset D pre The data parameters, the shared layer structure f of the neural network S Modulation module h i,j The neural network structure parameters include maximum model capacity limits and gradient conflict thresholds; the neural network training parameters include training epochs, learning step size, and number of detections. The training set, validation set, and prediction set are simulation data of the S21 transmission coefficient obtained from element frequency selection surface simulation of the radome using CST simulation software. The simulation settings parameters are substrate thickness and incident angle, ranging from 1-5 mm and 0-60 degrees, respectively. The simulation result is the S21 transmission coefficient. Simulation data with a substrate thickness of 1-2 mm is used as the training set, simulation data with a substrate thickness of 3-4 mm is used as the validation set, and simulation data with a substrate thickness of 5 mm is used as the prediction set, which are then input into the neural network model sequentially. These elements collectively determine the model design, training process, and final performance.

[0070] (2) Train the neural network model using the training dataset;

[0071] Initialize the training set and the shared layer network, train the shared layer network using the training set, and use an early stopping strategy until convergence.

[0072] (3) Calculate the multi-task conflict matrix;

[0073] This step is to calculate the multi-task conflict matrix, which is to prepare for the calculation of the indicator matrix and the generation of the modulation unit in the following step (4). The calculation of the indicator matrix in step (4) requires the multi-task conflict matrix to participate in the calculation.

[0074] The multi-task gradient conflict matrix C is calculated using the validation set on the trained model obtained in step (2), where the conflict matrix elements Cij The calculation formula is as follows:

[0075] C ij =cos(grad i ,grad j )

[0076] Wherein, cos(grad i ,grad j ) represents the cosine value between gradient vector i and gradient vector j. Its value represents the degree of conflict between task i and task j. The smaller the value, the greater the conflict.

[0077] (4) Calculate the indicator matrix;

[0078] This step is to calculate the multi-task conflict matrix, which prepares for the subsequent generation of modulation units.

[0079] The indicator matrix is ​​calculated based on the multi-task gradient conflict matrix and the conflict threshold, as follows:

[0080]

[0081] in This represents the conflict threshold.

[0082] (5) Generate modulation modules. Based on the indication matrix provided in step (4), generate modulation modules for the corresponding layers to modulate gradient conflicts.

[0083] The principle of the modulation module is as follows: Figure 3 The gradient adjustment diagram shown illustrates how the inner product formula of two gradient vectors is used in step (11) to adjust the gradient. Figure 2 Represented as h, Figure 2 (b) is a schematic diagram of an adaptive fusion module consisting of many modulation modules.

[0084] The modulation module's function is to fold the two gradients by increasing the dimensionality of the vectors in the two two-dimensional planes, so that they have a common descent direction, thereby alleviating gradient conflicts and enabling all tasks to have a common gradient descent effect.

[0085] (6) Determine whether the current model satisfies the maximum model limit;

[0086] Based on the maximum number of model parameters limit in step (1), determine the maximum allowed number of parameters. Compare the total number of parameters of the current model with the set maximum number of model parameters limit. If the number of parameters of the current model is lower than the limit, continue to the next step; if the number of parameters of the current model is equal to the limit, the current model construction process needs to be terminated and the process should proceed to step (9).

[0087] (7) Add the modulation module generated in step (5) to the neural network.

[0088] (8) Use the training set to train the newly added modulation module. After completing the pre-training, proceed to step (2).

[0089] (9) Output the current neural network model, output and save the current neural network model architecture and its parameters.

[0090] (10) Input the radome sample to be tested. The radome sample is the data obtained after performing frequency-selective surface simulation of the radome element using CST. Input the radome sample to be tested into the trained neural network model.

[0091] (11) Output the prediction results and generate the prediction value based on the read-in antenna radome sample.

[0092] Please see Figure 2 ,in Figure 2 (a) is a schematic diagram of the overall architecture of the neural network. Figure 2 (b) is a schematic diagram of the adaptive fusion module for the m-th task. Figure 2 This is a schematic diagram of the overall neural network architecture. The following formulas illustrate the internal workings of the neural network. The radome is input as a sample to be detected after the overall neural network architecture is completed. This sample is then combined with the neural network to generate predicted values. The radome sample is obtained through simulation data.

[0093] For m learning tasks, the overall architecture of the neural network includes:

[0094] 1) A fully shared encoder network Where L is the number of layers in the network;

[0095] 2) m×m×L task-specific encoder networks Implement the gradient adjustment module;

[0096] 3) m×m adaptive task-specific fusion parameters

[0097] 4) m task-specific decoder networks {g i} i∈[m] , used for the corresponding task header.

[0098] Input represents the input. This represents a shared module at layer l. β represents the adjustment module for tasks i and j at layer l, and β represents the final layer fusion weights. gi The decoder representing the i-th task. This represents the prediction for the i-th task. The adaptive fusion module for tasks: taking the m-th task as an example, only the m-th row and m-th column of the architecture can be fused.

[0099] For fully shared networks The formula for calculating the output of each layer is as follows:

[0100]

[0101] For implementing task-specific encoder networks Its output at each layer is composed of

[0102]

[0103] The input to the fusion module is shared knowledge. Task-specific gradient information from the previous layer Composition. The core of architecture search is determination. This indicates that the gradient conflict between the task (i,j) detected at layer r and the corresponding modulation module is activated if and only if it is 1.

[0104] An adaptive task-specific fusion layer connecting the encoder and decoder networks. It adaptively integrates shared knowledge and task-specific information.

[0105] by Figure 2 Taking the m-th task as an example (b), the output calculation formula for the m-th task header is:

[0106]

[0107] Where the fusion matrix {β i,i The diagonal elements of} do not exist because the tasks themselves do not conflict.

[0108] Please see Figure 3 Step (5) specifically involves:

[0109] When two tasks conflict, the modulation module mitigates the conflict between task i and task j. The modulation module can use either highly fine-grained neurons or less fine-grained modules. Formally, when task i and task j conflict at layer l of the network, cos(g) i,l ,g j,l ) < 0, where gi,l and gj,l represent the gradients of task i and task j at layer l, respectively;

[0110] According to the definition of the cosine between two gradient vectors:

[0111]

[0112] Two of the vectors conflict, which means that... <g i,l ,g j,l > < 0, now assume tasks i and j share Q parameters at layer l, and make <g i,l ,g j,l >=-c, where c is a positive number. The adjustment module consists of P parameters. Now, let's consider the gradient g after adding the adjustment module. i,l g j,l Divided into two parts: and and Therefore, the inner product of two gradient vectors can be expressed as:

[0113]

[0114] Gradient conflict occurred based on the previous assumptions, and <g i,l ,g j,l >=-c, by increasing the dimensionality of the vectors in two two-dimensional planes and folding the two gradients so that they have a common descent direction, the gradient conflict can be alleviated, so that all tasks have a common descent gradient.

[0115] This invention proposes a novel multi-task architecture learning method from the perspective of Multi-Objective Optimization (MAMO). It achieves its objective by finding a Pareto optimal solution to balance competing tasks and generating a gradient mediation unit module to mitigate gradient conflicts. Based on the gradient mediation unit, MAMO can adaptively fuse shared knowledge and dynamically learn to switch to specific task modules. It can also be integrated into other state-of-the-art (SOTA) methods to improve their performance.

[0116] The advantages of this invention can be further illustrated by the following simulation experiments:

[0117] Simulation conditions for the test dataset:

[0118] The frequency-selective surface of the radome was simulated using CST. The unit module consists of a substrate and a ring. The specific simulation settings are shown in Table 1. The simulated S21 transmission coefficient was imported into the surrogate model for training to obtain predicted values ​​of the transmission coefficient at various angles.

[0119] Table 1 Simulation Settings

[0120]

[0121]

[0122] Simulation results:

[0123] The neural network architecture search method for the radome proxy model of this invention is employed. Please refer to [link to relevant documentation]. Figure 4 , Figure 4This is a comparison of the simulated and predicted S21 curves at 15 GHz for substrate thicknesses h of 1 mm, 3 mm, and 5 mm, and polarization angles Phi of 0 degrees and 90 degrees. Figure 4 It can be seen that although there is a certain error between the simulation curve and the predicted curve of the surrogate model, the surrogate model of this invention can achieve a good prediction effect under a certain error. The maximum absolute error between the simulation value and the predicted value of S21 is shown in Tables 2 and 3, where Table 2 shows the maximum absolute error of the real part of the transmission coefficient and Table 3 shows the maximum absolute error of the imaginary part of the transmission coefficient. It can be seen from the tables that the absolute error is small and meets the error accuracy range. Therefore, under the allowable error, by training the surrogate model and then using the surrogate model input parameters to directly obtain the output value of S21, the development efficiency of the radome can be effectively improved and simulation time can be saved.

[0124] Table 2 Maximum absolute error of the real part (Re) of the transmission coefficient

[0125]

[0126] Table 3 Maximum absolute error of the imaginary part (Im) of the transmission coefficient

[0127]

[0128] The parts not described in detail in this implementation scheme are common and well-known methods in the industry, and will not be described in detail here. The above examples are merely illustrative of the present invention and do not constitute a limitation on the scope of protection of the present invention. Any methods identical to the present invention...

[0129] All designs, or similar designs, fall within the scope of protection of this invention.

Claims

1. An antenna radome proxy model oriented neural network architecture search method, characterized in that, The method comprises the following steps: Step (1): input the training set, the validation set and the prediction set into the neural network model in sequence, and search for space parameters; Step (2): training the neural network model by using the training data set; Step (3): calculating the multi-task conflict matrix by using the trained neural network model; Step (4): calculating the indication matrix according to the multi-task conflict matrix; Step (5): generating the modulation module according to the calculated indication matrix; Step (6): judging whether the current neural network model meets the maximum model limit; if the parameter quantity of the current model is lower than the limit, the subsequent steps are continued; if the parameter quantity of the current model is equal to the limit, the construction process of the current model needs to be terminated and step (9) is turned to; Step (7): adding the modulation module generated in step (5) to the neural network model; Step (8): training the newly added modulation module by using the training data set, and turning to step (2) after completing the pre-training; Step (9): outputting and saving the current neural network model architecture and its parameters; Step (10): inputting the antenna cover sample to be detected into the trained neural network model; Step (11): generating the prediction value according to the read-in antenna cover sample to be detected, and outputting the prediction result; In the step (2), the training set and the shared layer network are initialized, the shared layer network is trained by using the training set, and the early stopping strategy is used until convergence; The training neural network in step (2) is specifically: Using the training set D train The model parameters are trained and their loss is monitored using an early stopping mechanism that automatically terminates training if the loss does not improve over 50 training epochs. In step (3), the trained neural network model obtained in step (2) is used to calculate the multi-task gradient conflict matrix C using the validation set D val The back-propagated gradient is calculated on the model trained in step (2), wherein the gradient with respect to task i is grad i , the gradient with respect to task j is grad j ; The gradient conflict between the task gradients is calculated by the following formula: conflict = cos(grad i , grad j ) The multi-task gradient conflict matrix C is obtained by calculating the gradient conflict between each two tasks, and C is a symmetric matrix; where the conflict matrix element C ij The calculation formula is as follows: C ij = cos(grad i , grad j ) wherein cos(grad i , grad j ) represents the cosine value between the gradient vector i and the gradient vector j, and the value size represents the conflict degree between the task i and the task j, and the smaller the value is, the more conflict.

2. The method of claim 1, wherein, In the step (1), the data parameters of the training data set D train , the verification data set D val , and the prediction data set D pre are input into the neural network model in sequence, and the neural network structure parameters of the shared layer structure f S , the modulation module h i,j , the maximum model capacity limit, and the gradient conflict threshold, and the neural network training parameters of the training round, the learning step, and the detection number are input into the neural network model in sequence.

3. The method of claim 1, wherein, The step (4) is specifically: according to the multi-task gradient conflict matrix C and the gradient conflict threshold input in the step (1), the indicating matrix is calculated The specific calculation is as follows: wherein representing the conflict threshold, D val If the data is divided into batches, there will be batch number of indicator matrices In multi-batch detection, the conflict matrix of the whole validation set is obtained by summing the indicator matrices, and the calculation formula is as follows: The final indication matrix is generated by using the average value strategy or the maximum value strategy, and the average value strategy calculation formula is as follows: where I average The average value of the matrix I is represented by I and the maximum value strategy calculation formula is as follows: where I top represents the maximum value of the matrix I, the average value strategy and the maximum value strategy mean whether the number of modulation modules generated after each detection of a collision is half of the search space or only one modulation module is generated.

4. The method of claim 3, wherein, The step (5) is specifically: according to the indication matrix provided in step (4), the modulation module for modulating the gradient conflict is generated in the corresponding layer; When two tasks conflict, the modulation module mitigates the conflict between task i and task j. The modulation module can use either highly fine-grained neurons or low-grained modules. Formally, when task i and task j conflict at layer l of the network, cos(g) i,l ,g j,l ) < 0, g i,l With g j,l represent the gradients of task i and task j at layer l, respectively; According to the cosine between two gradient vectors, it is defined as: where two vectors conflict, meaning that <g i,l , j,l > < 0, now assume that tasks i and j share Q parameters on the l-th layer and that <g i,l , j,l > = -c, where c is a positive number, the adjusted module consists of P parameters, the gradient g i,l , j,l after adding the adjusted module is divided into two parts: and and The inner product of the two gradient vectors is represented as: Based on the above assumption that gradient conflicts occur, and <g i,l ,g j,l > = -c, the two two-dimensional plane vectors are folded by dimension elevation to fold the two gradients, so that they have a common descending direction, alleviate the gradient conflict, so that all tasks have a common descending gradient.

5. The method of claim 4, wherein, The step (6) is specifically: According to the maximum model parameter quantity limit in step (1), the maximum parameter quantity allowed is determined, and the total parameter quantity of the current model is compared with the set maximum model parameter quantity limit; The related calculation formula of adding the modulation module to the neural network in the step (7) is as follows: For the input and output of the full shared layer, the calculation formula is as follows: wherein representing an input of the lth layer, representing an output of the lth layer, f s ( ) represents the l+1th shared layer; For the input and output of the modulation module between task i and task j, the calculation formula is as follows: wherein represents the input of the lth modulation module, represents the output of the modulation module.

6. The method of claim 2, wherein, The pre-training method of the modulation module in the step (8) is as follows: After adding the modulation module, freeze the shared layers and the parameters of the modulation module added before, initialize the newly added module using the Xavier initialization strategy, and use the training set D train Pre-train it, also using the early stopping mechanism described in step (2).

Images