Extensible expert network fine tuning method based on meta-core architecture and coefficient sharing
By decomposing the convolution kernel of a deep neural network into meta-kernels and meta-kernel coefficients, constructing a hybrid expert architecture and sharing coefficients, the problem of excessively rapid parameter growth and insufficient adaptability in existing fine-tuning methods is solved, and the model is efficiently expanded and its performance is improved in resource-constrained environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-05
AI Technical Summary
Existing efficient parameter fine-tuning methods such as LoRA are prone to overfitting when faced with limited data, and the number of parameters increases linearly with the rank, making them unfriendly to memory-constrained fine-tuning environments. Fine-tuning methods based on filter subspaces are difficult to fully capture features in complex tasks and the number of parameters increases rapidly.
A scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing is adopted. By decomposing the convolution kernel of the deep neural network into meta-kernels and meta-kernel coefficients, a hybrid expert architecture is constructed. Dynamic gating network is used to improve parameter efficiency and expressive power, and meta-kernel coefficients are shared among experts to control the growth of parameter quantity.
It achieves improved scalability and adaptability of the model in resource-constrained environments, has strong adaptability to dynamic samples, and the growth of the number of parameters is much lower than linear growth, adapting to complex and diverse datasets and improving model performance.
Smart Images
Figure CN121981199A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning technology, specifically relating to a scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing. Background Technology
[0002] In recent years, large-scale pre-trained models, represented by deep convolutional neural networks and Transformers, have made groundbreaking progress in fields such as computer vision and natural language processing. However, when these general-purpose pre-trained models are applied to specific downstream tasks (such as image classification and object detection), full fine-tuning is usually required, i.e., updating all parameters of the model. This process requires enormous computational resources, storage space, and GPU memory overhead. Parameter-Efficient Fine-Tuning (PEFT) solves this problem by updating only a small subset of parameters (typically in the millions), adapting to downstream applications while maintaining the generalization performance of the pre-trained model.
[0003] Among existing parameter-efficient fine-tuning methods, Low-Rank Adaptation (LoRA) and its variants are widely used and representative techniques. LoRA achieves high efficiency in storage and computation by freezing pre-trained weights through training additional low-rank matrices; however, LoRA is prone to overfitting with limited data. Furthermore, the number of parameters in LoRA increases linearly with rank, further increasing the number of training parameters, which remains unfavorable for memory-constrained fine-tuning environments. Although LoRA can reduce the number of updated parameters to no more than 1% of that in full-parameter fine-tuning, in deeper and larger networks, LoRA still needs to increase the rank within a certain range to maintain fine-tuning performance, inevitably increasing the number of trainable parameters.
[0004] Another type of fine-tuning method based on convolutional filter decomposition (i.e., filter subspace-based fine-tuning) is based on the idea of constructing an inherent low-rank dimension using LoRA. It proposes decomposing the convolutional kernel into atoms and atomic coefficients, and fine-tuning only the atoms. Atoms have a relatively small number of parameters and can be considered as bases. Through linear combinations of atomic coefficients with these bases, convolutional filter decomposition can reconstruct the parameter space with a lower dimension. This fine-tuning paradigm can update high-dimensional weights with fewer parameters, but it may struggle to fully capture features in complex tasks. In filter subspace-based fine-tuning methods, atoms are further decomposed into sub-atoms and sub-atomic coefficients to increase the number of trainable parameters. While this significantly improves performance, it also rapidly increases the number of parameters. Summary of the Invention
[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing includes the following steps:
[0008] S1. Decompose the original convolution kernel of a given deep neural network into a meta-kernel and meta-kernel coefficients;
[0009] S2. Learn and fix the kernel coefficients from the pre-trained weights;
[0010] S3. Add the pre-trained weights as a bypass to the decomposed convolutional kernel;
[0011] S4. Construct the expert component of a hybrid expert architecture, creating multiple expert branches with shared coefficients;
[0012] S5. Design and integrate dynamic gating networks;
[0013] S6. Set the training strategy, including optimizer, learning rate, batch size, number of meta-kernels, and number of experts;
[0014] S7. Forward computation and loss acquisition: Data is propagated forward through the model, and the prediction results are obtained by fusing expert outputs through a gated network. The calculated loss value provides a basis for parameter updates.
[0015] S8. Gradient calculation and parameter update: Calculate the gradient through backpropagation and update the network parameters using the optimizer.
[0016] S9. Repeat steps S7 to S8 until the total training loss value converges.
[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0018] 1. This invention achieves a dual breakthrough in parameter efficiency and expressive power: by constructing a hybrid expert (MoE) framework based on dynamic routing, a fine-tuning method based on convolutional kernel decomposition is introduced into expert network design. Leveraging the low parameter count advantage of convolutional kernel decomposition and the dynamic adaptability of MoE, while reducing the number of trainable parameters through convolutional kernel decomposition, the multi-expert mechanism of MoE enhances the model's expressive power.
[0019] 2. Dynamic sample adaptability: Gated networks can automatically adjust expert weights according to the characteristics of input samples, achieving "expert specialization" and ensuring that the model has stronger adaptability when dealing with different samples. At the same time, it provides a flexible task allocation mechanism for model expansion, significantly improving the model's adaptability and final performance on complex and diverse datasets.
[0020] 3. Excellent scalability and deployment friendliness: Since the kernel coefficients are shared among experts, when the number of experts needs to be increased to improve the model capacity, the growth rate of the total number of trainable parameters of the model is much lower than linear growth (i.e., sublinear growth is achieved). This feature makes it possible to expand the model in resource-constrained environments, providing an efficient fine-tuning solution for deploying large-scale models on edge devices. Attached Figure Description
[0021] Figure 1 This is a flowchart of the method of the present invention.
[0022] Figure 2 This is a schematic diagram of the hybrid expert network DCK_MoE based on convolution kernel decomposition in the embodiment.
[0023] Figure 3 This is a schematic diagram comparing the kernel decomposition process with the original convolution process. Detailed Implementation
[0024] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0025] This invention presents a scalable expert network fine-tuning method based on a meta-kernel architecture and coefficient sharing. Leveraging the dynamic multi-expert characteristics of MoE (Modal-Expert) models, the method enhances the model's adaptability to input samples. It introduces a fine-tuning method based on convolutional filter decomposition into the construction of the expert network within MoE, proposing the DCK_MoE (Decomposed Convolutional Kernel_Mixture-of-Experts) framework. To avoid the parameter surge problem caused by the multi-expert structure in MoE, it borrows the method of sharing atomic coefficients among models in the field of continuous learning, achieving sublinear growth in the number of parameters by sharing meta-kernel coefficients among experts.
[0026] Example; This example uses fine-tuning a pre-trained ResNet-18 model on the ImageNet dataset to illustrate the implementation steps of the present invention in detail; Figure 1 As shown, it includes the following steps:
[0027] S1. Decompose the original convolutional kernel of a given deep neural network into a kernel and kernel coefficients; specifically:
[0028] First, the target convolutional layer in the model is selected. In this embodiment, all convolutional layers in the ResNet-18 network are decomposed.
[0029] Given the ResNet-18 network, the first... The original convolutional kernel of the layer is represented as ,in Indicates the input channel. Indicates the output channel. Indicates the kernel size;
[0030] Decompose the original convolutional kernel into metakernels. With kernel coefficient ,in, , , Indicates the number of nuclei. Indicates the kernel size; the first kernel in the network The decomposition calculation of the convolution kernel of the layer is as follows:
[0031]
[0032] in, This indicates tensor multiplication, not convolution. Indicates the first Layer kernel coefficients, Indicates the first The core of the layer;
[0033] When using convolution kernel decomposition, at the... In layer computation, input With the nucleus Kernel coefficients The computation consists of two steps: spatial convolution only and channel fusion; input ,in and These represent the input height and width, respectively.
[0034] enter With the nucleus Perform spatial convolution only to obtain intermediate features intermediate features With kernel coefficient Channel fusion is performed to obtain the final output. ;
[0035] enter With the nucleus The spatial convolution process performed is as follows:
[0036] In the In the layer, input With the nucleus During the spatial convolution process, the input Each channel and each core After convolution, we get the following: Intermediate features of each channel ;
[0037] intermediate features With kernel coefficient The channel fusion process is as follows:
[0038] In the In the layer, intermediate features With kernel coefficient Perform channel weight allocation; kernel coefficients Depend on Components composition, ,in , Represented as intermediate features of The weights assigned to each channel; after indivual The allocation yields the final output. ;in and These represent the output height and width, respectively. Since this process does not involve convolution calculations, it... arrive Its characteristic is that there are no changes in height and width.
[0039] like Figure 3 The diagram shows a comparison between the convolution kernel decomposition process and the original convolution process. The diagram illustrates the process when... , , A comparison between temporal convolution kernel decomposition and the original convolution process.
[0040] S2. Learn kernel coefficients from pre-trained weights And fix it; specifically:
[0041] By utilizing the ideas of dictionary learning and sparse coding, the kernel coefficients learned from the pre-trained weights can approach the optimal channel combination.
[0042] The goal of dictionary learning and sparse coding is to learn from the original weights. In this case, a sparse coding matrix is used. and a dictionary Perform matrix multiplication to fit the original weights as closely as possible. And requires a sparse coding matrix As sparse as possible; the optimization objective can be written as:
[0043]
[0044] in, Let represent the Lagrange multipliers, where the first term represents maximizing the fit to the original weights, and the second term represents the penalty for sparsity. Indicates pre-trained weights, Indicates the number of input channels. Indicates the number of output channels. The kernel size; , representing the nucleus, , representing the kernel coefficient;
[0045] In the first stage of the learning process, the kernel coefficients are fixed and the kernel is trained; in the second stage, the kernel is fixed and the kernel coefficients are trained; through multiple alternating optimizations in these two stages, the kernel coefficients are obtained. The optimal solution is obtained; after convergence, the optimal kernel coefficients are obtained and kept frozen during subsequent fine-tuning, retaining only the trainability of the kernel.
[0046] S3. Add the pre-trained weights as a bypass to the decomposed convolutional kernel; specifically:
[0047] The first The weights of the decomposed convolutional kernels in the layer are related to the first layer. The pre-trained weights of the layers are added together to obtain the final weights of the first layer. The convolution kernel of the layer is specifically calculated as follows:
[0048]
[0049] in, Indicates the first The final fine-tuning weights of the layer, Indicates the first The pre-trained weights of the layer Indicates the first Weights decomposed using convolution kernels within a layer.
[0050] S4. Construct a multi-expert branch with shared coefficients; specifically:
[0051] Based on the convolutional kernel structure built in steps S1 to S3, the expert part of the hybrid expert (MoE) architecture is constructed. ;
[0052] create (For example, N=9) expert branches with identical structures; each expert Each contains its own independent set of trainable kernels. But all Each expert shares the same set of fixed kernel coefficients obtained in step S2. .
[0053] S5. Design and integrate dynamic gating networks; specifically:
[0054] The gating network is responsible for dynamically assigning expert weights based on the input samples. The structure of the gating network in this embodiment is shown in Table 1 below. In this embodiment, the gating network has four convolutional layers to extract input features, and finally, after global average pooling, the features are distributed to N experts according to their weights through a fully connected layer.
[0055] Table 1 Gating Network Structure
[0056]
[0057] Given input back, Entering the gated network will get the correct The weights assigned to each expert are calculated as follows:
[0058]
[0059] at the same time They will also enter separately An expert received this Features extracted by experts The calculation process is as follows:
[0060]
[0061] in, Represents a neural network; when the input and Layer fine-tuning weights After convolution processing, the output result is obtained. ;
[0062] The weights generated by the gating network Features extracted by various experts We perform a weighted summation to obtain the fused feature representation. The calculation process is as follows:
[0063]
[0064] in, Represents scalar multiplication;
[0065] Subsequently, the fusion features will be used. Input to fully connected layer The final prediction output of the network is obtained by performing linear transformation and classification. The calculation process is as follows:
[0066]
[0067] During training, all kernel coefficients are fixed. Only fine-tuning the core Gating network and linear layer .
[0068] S6. Set the training strategy, including optimizer, learning rate, batch size, number of kernels, etc., specifically:
[0069] The cross-entropy loss function is used; the batch size is set to 256; Adam is used as the optimizer; the initial learning rate is 0.001; and the weight decay is 0.0001.
[0070] The number of cores, m, can be configured as 9, 12, 15, etc.
[0071] The number of experts, N, can be configured as 3, 6, 9, etc.
[0072] The convolutional kernels of the gated network are initialized using the Kaiming initialization method based on the number of output channel connections, and batch normalization is added to the output of each convolutional layer.
[0073] S7. Forward computation and loss acquisition: Data is propagated forward through the model, and the prediction results are obtained by fusing expert outputs through a gating network. The loss value is calculated to provide a basis for parameter updates. The training data is input into the model, and the forward propagation of the fine-tuning process is performed as follows:
[0074] The input sample is sent to the gating network designed in step S5 for processing, and the sample is obtained. Weights assigned to each expert ;
[0075] The input sample is sent to all of the steps described in step S4. Each expert branch processes in parallel; each expert Use its independent core and shared kernel coefficients Perform calculations, each expert Output its processed features ;
[0076] Using the obtained weights For corresponding features Perform a weighted summation to output the weighted sum of each expert's features. ;
[0077] Weighted features The process continues through the fully connected layer of the deep neural network until the final prediction is obtained.
[0078] The forward propagation process described above is executed to obtain the predicted output. The loss value is then calculated based on the predicted result and the true label. .
[0079] S8. Gradient calculation and parameter update: The gradient is calculated through backpropagation, and the network parameters are updated using the optimizer; specifically:
[0080] Calculate the loss value Gradients over all trainable parameters in the model; trainable parameters include: kernels of all expert branches. Gating network Weights and fully connected layers The weights. And the kernel coefficients. and pre-trained weights The system remains frozen throughout the fine-tuning process, and its gradient is not calculated.
[0081] The gradient is calculated using back propagation, based on the loss value. The partial derivatives of the trainable parameters are calculated; this process is automatically performed by deep learning frameworks such as PyTorch and TensorFlow.
[0082] The Adam optimizer is used to update all trainable parameters based on the calculated gradients.
[0083] In each training batch, the loss value is calculated according to the chain rule. For the current trainable network parameters gradient back( Represents the parameters of the neural network. (This is the gradient of a single batch of data with respect to the corresponding parameters). The Adam optimizer updates the parameters according to the following steps:
[0084] The calculation of the first and second moment estimates is as follows:
[0085]
[0086]
[0087] in, This represents the current iteration step. and The hyperparameters for the exponential decay rate of the moment estimation are set to 0.9 and 0.999 by default, respectively. and These are the first moment (mean) estimate and the second moment (uncentered variance) estimate of the gradient, respectively.
[0088] because and Initializing the training step size to 0 will cause the estimation to be biased towards zero, so bias correction is performed to obtain an unbiased estimate:
[0089]
[0090]
[0091] in, It is the gradient first moment (mean) estimate after bias correction. It is the gradient second moment (uncentered variance) estimate after bias correction;
[0092] The corrected moment estimate is used to calculate the parameter update amount, and the network parameters are updated accordingly:
[0093]
[0094] in, The global learning rate, It is a constant (usually 1) ), used to maintain numerical stability.
[0095] S9. Repeat steps S7 to S8 until the total training loss value is reached. Convergence; using the finally converged model parameters, define the fine-tuned hybrid expert network DCK_MoE based on convolutional kernel decomposition, and apply this network to the test set sample data to obtain the final task recognition result. For example... Figure 2 The diagram shown is a schematic of DCK_MoE in this embodiment.
[0096] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0097] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing, characterized in that, Includes the following steps: S1. Decompose the original convolution kernel of a given deep neural network into a meta-kernel and meta-kernel coefficients; S2. Learn and fix the kernel coefficients from the pre-trained weights; S3. Add the pre-trained weights as a bypass to the decomposed convolutional kernel; S4. Construct the expert component of a hybrid expert architecture, creating multiple expert branches with shared coefficients; S5. Design and integrate dynamic gating networks; S6. Set the training strategy, including optimizer, learning rate, batch size, number of meta-kernels, and number of experts; S7. Forward computation and loss acquisition: Data is propagated forward through the model, and the prediction results are obtained by fusing expert outputs through a gated network. The calculated loss value provides a basis for parameter updates. S8. Gradient calculation and parameter update: Calculate the gradient through backpropagation and update the network parameters using the optimizer. S9. Repeat steps S7 to S8 until the total training loss value converges.
2. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 1, characterized in that, Step S1 is as follows: In a given deep neural network, the first... The original convolutional kernel of the layer is represented as ,in Indicates the input channel. Indicates the output channel. Indicates the kernel size; Decompose the original convolutional kernel into metakernels. With kernel coefficient ,in, , , Indicates the number of nuclei. Represents the kernel size; the third kernel in a deep neural network The decomposition calculation of the convolution kernel of the layer is as follows: in, Represents tensor multiplication. Indicates the first Layer kernel coefficients, Indicates the first The core of the layer; During kernel decomposition, at the th In layer computation, input With the nucleus Kernel coefficients The computation consists of two steps: spatial convolution only and channel fusion; input ,in and These represent the input height and width, respectively. enter With the nucleus Perform spatial convolution only to obtain intermediate features intermediate features With kernel coefficient Channel fusion is performed to obtain the final output. ; enter With the nucleus The specific process of performing spatial convolution is as follows: In the In the layer, input With the nucleus During the spatial convolution process, the input Each channel and each core After convolution, we get the following: Intermediate features of each channel ; intermediate features With kernel coefficient The channel fusion process is as follows: In the In the layer, intermediate features With kernel coefficient Perform channel weight allocation; kernel coefficients Depend on Components composition, ,in , Represented as intermediate features of The weights assigned to each channel; after indivual The allocation yields the final output. ,in and These represent the output height and width, respectively.
3. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 2, characterized in that, Step S2 is as follows: By utilizing the ideas of dictionary learning and sparse coding, the kernel coefficients learned from the pre-trained weights can approach the optimal channel combination. The goal of dictionary learning and sparse coding is to learn from the original weights. In this case, a sparse coding matrix is used. and a dictionary Perform matrix multiplication to fit the original weights as closely as possible. And requires a sparse coding matrix As sparse as possible; this optimization objective is expressed as: in, Let represent the Lagrange multipliers, where the first term represents maximizing the fit to the original weights, and the second term represents the penalty for sparsity. Indicates pre-trained weights, Indicates the number of input channels. Indicates the number of output channels. The kernel size; , representing the nucleus, , representing the kernel coefficient; In the first stage of the learning process, the kernel coefficients are fixed and the kernel is trained; in the second stage, the kernel is fixed and the kernel coefficients are trained; through multiple alternating optimizations in these two stages, the kernel coefficients are obtained. The optimal solution is obtained; after convergence, the optimal kernel coefficients are obtained and kept frozen during subsequent fine-tuning, retaining only the trainability of the kernel.
4. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 3, characterized in that, Step S3 is as follows: The first The weights of the decomposed convolutional kernels in the layer are related to the first layer. The pre-trained weights of the layers are added together to obtain the final weights of the first layer. The convolution kernel of the layer is specifically calculated as follows: in, Indicates the first The final fine-tuning weights of the layer, Indicates the first The pre-trained weights of the layer Indicates the first Weights decomposed using convolution kernels within a layer.
5. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 4, characterized in that, Step S4 is as follows: Based on the convolutional kernel structure built in steps S1 to S3, the expert part of the hybrid expert architecture is constructed. ; Create multiple expert branches with identical structures; Each expert Each contains a separate set of trainable kernels. But all Each expert shares the same set of fixed kernel coefficients obtained in step S2. .
6. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 1, characterized in that, Step S5 is as follows: Design a gated network structure; given an input back, Entering the gated network will get the correct The weights assigned to each expert are calculated as follows: at the same time Enter separately An expert received this Features extracted by experts The calculation process is as follows: in, Represents a neural network; when the input and Layer fine-tuning weights After convolution processing, the output result is obtained. ; The weights generated by the gating network Features extracted by various experts We perform a weighted summation to obtain the fused feature representation. The calculation process is as follows: in, Indicates scalar multiplication; Subsequently, the fusion features will be used. Input to fully connected layer The final prediction output of the network is obtained by performing linear transformation and classification. The calculation process is as follows: During training, all kernel coefficients are fixed. Only fine-tuning the core Gating network and linear layer .
7. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 6, characterized in that, Step S6 is as follows: The cross-entropy loss function is used; the batch size is set to 256; Adam is used as the optimizer; the initial learning rate is 0.001; and the weight decay is 0.0001. Configure the number of cores, m; Configure the number of experts N; The convolutional kernels of the gated network are initialized using the He Kaiming initialization method based on the number of output channel connections, and batch normalization layers are added to the output of each convolutional layer for processing.
8. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 7, characterized in that, Step S7 is as follows: The input sample is sent to the gating network designed in step S5 for processing, and the sample is obtained. Weights assigned to each expert ; The input sample is sent to all of the steps described in step S4. Each expert branch processes in parallel; each expert Use its independent core and shared kernel coefficients Perform calculations, each expert Output its processed features ; Using the obtained weights For corresponding features Perform a weighted summation to output the weighted sum of each expert's features. ; Weighted features The process continues through the fully connected layer of the deep neural network until the final prediction is obtained. The loss value is calculated based on the prediction results and the actual labels. .
9. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 8, characterized in that, Step S8 is as follows: Loss value is calculated through backpropagation. The gradient of the trainable parameters is calculated, and the trainable parameters are updated using an optimization algorithm based on adaptive moment estimation. In each training batch, the loss value is calculated according to the chain rule. For the current trainable network parameters gradient Then, the Adam optimizer updates the parameters according to the following steps: The calculation of the first and second moment estimates is as follows: in, This represents the current iteration step. and The hyperparameters for the exponential decay rate of the moment estimation are set to 0.9 and 0.999, respectively. and These are the first-order moment estimate and the second-order moment estimate of the gradient, respectively; because and Initializing the training step size to 0 will cause the estimation to be biased towards zero, so bias correction is performed to obtain an unbiased estimate: in, It is the gradient first moment estimate after bias correction. It is the gradient second moment estimate after bias correction; The corrected moment estimate is used to calculate the parameter update amount, and the network parameters are updated accordingly: in, The global learning rate, It is a constant used to maintain numerical stability.
10. The scalable expert network fine-tuning method based on meta-kernel architecture and coefficient sharing according to claim 9, characterized in that, Step S9 is as follows: Repeat steps S7 to S8 until the total training loss value is reached. convergence; The fine-tuned hybrid expert network DCK_MoE based on convolutional kernel decomposition is defined using the finally converged model parameters. This network is then applied to the test set sample data to obtain the final task recognition results.