A test-time personalized model generation method and system for dynamic networks

Abstract

This application belongs to the fields of distributed machine learning, decentralized learning, test-time personalization, and hypernetwork model generation technology, and particularly relates to a test-time personalized model generation method and system for dynamic networks. Specifically, it includes: a coupled dual-model construction module for building a shared pre-trained encoder, hypernetwork generator, and task model personalized parameters on the training client; a descriptor generation module for generating client descriptors based on local data features; a decentralized aggregation module for aggregating neighbor hypernetwork parameters based on dynamic communication topology; a hypernetwork training module for back-updating hypernetwork parameters based on multi-step update differences of the task model; and a test-time generation module for aggregating the training client hypernetwork parameters and generating a fine-tuning-free personalized model for newly added test clients.

Description

Technical Field

[0001] This application belongs to the fields of distributed machine learning, decentralized learning, test-time personalization, and hypernetwork model generation technology, and particularly relates to a method and system for rapidly generating personalized models for newly joined test clients in a dynamic decentralized network. Background Technology

[0002] With the development of autonomous driving, drone swarms, multi-agent collaboration, edge intelligence, and mobile terminal intelligence applications, an increasing number of learning tasks need to be completed collaboratively by multiple clients in an environment without a central server. Decentralized learning allows each client to communicate only with its network neighbors and update the model based on local data distribution, thereby reducing dependence on a central server and enhancing system robustness.

[0003] In real-world scenarios, the data distribution across different clients typically varies significantly. To adapt to this data heterogeneity, personalized collaborative learning methods usually learn a personalized model tailored to each training client's local data. However, traditional decentralized personalized learning primarily focuses on clients already present during the training phase, i.e., a fixed set of training clients. When new clients join during the testing phase, these new clients often expect to receive a usable personalized model immediately upon connecting to the network.

[0004] Existing technologies typically address personalization issues through retraining, local fine-tuning based on the initial model, meta-learning initialization, or generating model parameters via hypernetworks. However, retraining and local fine-tuning require multiple rounds of gradient updates on the test client, resulting in high computational overhead and latency, making it difficult to meet real-time, on-demand personalization needs. Some hypernetwork-based solutions require a central server to continuously collect client representations or data descriptors, which are difficult to apply directly in decentralized dynamic topologies. Other solutions rely on trainable client embeddings, but the test client lacks the corresponding embeddings upon initial joining, making it impossible to directly generate effective personalized parameters.

[0005] Furthermore, decentralized dynamic networks suffer from issues such as unstable communication links, client mobility, and changes in neighbor relationships over time. If each training client independently maintains a hypernetwork without a consistent aggregation mechanism, the hypernetworks received by the testing client from different training clients may be inconsistent, thus affecting the quality of model generation. Therefore, a personalized model generation method is needed that requires no central server, no fine-tuning during testing, can adapt to dynamic topologies, and can utilize features from unlabeled test data. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this application provides a method and system for generating personalized models for decentralized dynamic networks without fine-tuning during testing. This method uses a coupled dual-model architecture to simultaneously use a shared pre-trained encoder for client descriptor extraction and task model feature representation, and obtains a hypernetwork that can directly generate personalized models during the testing phase through decentralized hypernetwork aggregation training. The technical solution adopted by this application to solve its technical problem is as follows: A method for generating personalized test-time models for dynamic networks includes the following steps: S1: Construct a coupled dual-model architecture on multiple training clients, the coupled dual-model architecture including a shared pre-trained encoder, a hypernetwork generator, and task model personalized parameters; S2: Each training client receives hypernetwork parameters from neighboring training clients based on the decentralized communication topology of the current round, and performs weighted aggregation of the received hypernetwork parameters and local hypernetwork parameters. S3: Each training client uses local data features and the shared pre-trained encoder to generate a client descriptor, and the local hypernetwork generator generates corresponding personalized parameters for the task model based on the client descriptor; S4: Each training client generates local task model personalized parameters based on the aggregated hypernetwork, and performs several steps of task model update using local data. The difference between the task model personalized parameters before and after the update is used to construct the hypernetwork training signal to update the local hypernetwork parameters in reverse. S5: During the testing phase, a new test client connects to one or more training clients, aggregates the trained hypernetwork parameters of the connected training clients, and inputs the local data features of the test client into the aggregated hypernetwork to generate a personalized test client model that can be used without fine-tuning.

[0007] Preferably, in the coupled dual-model architecture, the hypernetwork and the task model share the same pre-trained encoder; wherein, the hypernetwork is represented as... The task model is represented as , Indicates a shared pre-trained encoder. Indicates the first Hypernetwork generator parameters for each training client, Indicates the first Personalized parameters for the task model of each training client.

[0008] Preferably, in the first In the first round of training, Each training client follows a dynamic communication matrix We perform weighted aggregation of the hypernetwork parameters of the neighboring training clients to obtain the aggregated hypernetwork parameters: ; in, The decentralized topology weight matrix varies with training rounds and is determined by client movement, link availability, or random rotation mechanisms. After each round of topology determination, each client calculates and normalizes the weights based on its current set of communicable neighbors. The element in the i-th row and j-th column ,in In the qth round, the relationship with the 1st round is... The number of neighbors that a training client can directly communicate with; A value greater than zero indicates that the i-th training client and the j-th training client can communicate directly in this round. Then the i-th training client and the j-th training client cannot communicate in this round. For the hypernet generator parameters of neighbor j.

[0009] Preferably, the client descriptor is derived from the features of the training client's local data. Shared pre-trained encoder Generate, represented as: ; This indicates that the parameters are shared by the pre-trained encoder. The descriptor generation function of the representation is the client descriptor, the client descriptor Used to characterize the local data distribution of the training client, enabling the hypernetwork generator to generate data based on the client descriptor. Output personalized parameters for the task model that adapt to the data distribution of this client.

[0010] Preferably, step S4 includes the following steps: S401. Based on the aggregated hypernetwork parameters and client descriptor Generate personalized parameters for the initial task model: ; Represents the parameters of the aggregated hypernetwork The hypernetwork generating function is characterized; S402. Personalize the parameters based on the initial task model. Starting from the client Execute on local data batch Step gradient descent to obtain the updated personalized parameters of the task model. ; S403. Personalize parameters based on the initial task model. and the updated task model personalized parameters Calculate the parameter difference: ; S404. The parameter difference As the target and training signal for the supernetwork generation, it is used to guide the backpropagation update of the local supernetwork parameters.

[0011] Preferably, the hypernetwork training signal is backpropagated from the task model's personalized parameter space to the hypernetwork parameter space via a chain rule, and then... Update local hypernetwork parameters; among which, Indicates the supernet learning rate. Indicates the first The training objective of the hypernetwork for each training client. For gradient.

[0012] Preferably, each training client only sends or receives hypernetwork generator parameters during the communication phase, without needing to send complete task model parameters or local data, thereby reducing the communication burden in the decentralized dynamic network and avoiding reliance on a central server for model generation or model aggregation.

[0013] Preferably, the personalized model generation in the testing phase of step S5 includes the following steps: S501. When a new test client is added. When accessing a decentralized network, the test client Establish a communication connection with at least one trained client according to the connection relationship, and obtain the corresponding connection weights through normalization based on the connection relationship. : ; A value greater than zero indicates a test client With the client Connected; S502. The test client Receive connected training client The trained hypernetwork parameters The aggregated network parameters corresponding to the test client are obtained by weighting and aggregating the data according to the connection weights. ; S503. The test client Local data characteristics Input consists of the aggregated hypernetwork parameters In the represented hypernetwork, personalized parameters are generated for the test client: ; Indicates the aggregation of hypernetwork parameters The hypernetwork generating function is characterized. This indicates that the parameters are shared by the pre-trained encoder. The descriptor generation function for the representation; S504. Share encoder parameters With the personalized parameters Combine them to obtain the test client. Personalized models: .

[0014] Preferably, the test client only needs local data features when generating a personalized model. It does not rely on local data labels; the test client obtains a personalized model through one or a small number of forward calculations, without the need to perform backpropagation or gradient fine-tuning during the testing phase.

[0015] A test-time personalized model generation system for dynamic networks includes a coupled dual-model construction module, a descriptor generation module, a decentralized aggregation module, a hypernetwork training module, and a test-time generation module. Coupled dual-model building module, used to build a shared pre-trained encoder, hypernetwork generator and task model personalized parameters on the training client; The descriptor generation module is used to generate client descriptors based on local data characteristics; The decentralized aggregation module is used to aggregate neighbor supernetwork parameters based on dynamic communication topology; The hypernetwork training module is used to update the hypernetwork parameters in reverse by updating the difference in multiple steps based on the task model; The test-time generation module is used to aggregate the hypernetwork parameters of newly added test clients and generate a personalized model that does not require fine-tuning.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This application significantly enhances the rapid deployment capability of newly added test clients. During the testing phase, new clients do not need to re-participate in the complete training process, nor do they need to perform multiple rounds of gradient fine-tuning based on local data. Instead, they receive and aggregate the hypernetwork parameters of trained clients, and use the aggregated hypernetwork to directly generate personalized model parameters adapted to the features of their local data. Therefore, new clients can quickly obtain personalized models for testing or inference after connecting to the decentralized network, better meeting the application requirements of frequent addition of new devices, rapidly changing task requirements, and real-time response in open environments.

[0017] This application also reduces the personalization cost during the testing phase. Traditional personalized learning methods typically require new clients to save, update, or fine-tune the complete task model, resulting in high computational overhead and waiting time. This application, however, transforms the personalization process during testing into a single or limited forward generation process of the hypernetwork, avoiding the additional overhead of retraining and multi-step local optimization. Furthermore, the personalized model generated by the testing client primarily relies on local data features and does not require labeled data, thus remaining applicable even when the testing client only possesses unlabeled data, further lowering the deployment threshold.

[0018] This application is better suited to dynamic decentralized network environments. During the training phase, it does not rely on a central server to uniformly maintain or distribute the global model. Each training client only exchanges hypernetwork parameters with its neighboring clients and performs weighted aggregation based on the connection weights of each round. When the communication topology changes due to device movement, link fluctuations, or temporary client access, the client can still update the hypernetwork based on the current neighbor relationships. New clients in the testing phase can also aggregate the trained hypernetworks of neighboring training clients based on the connection relationships formed upon their access, thereby gaining personalized generation capabilities and enhancing the applicability and robustness of the method in open, dynamic, distributed scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] This application provides a test-time personalized model generation method for dynamic networks. The method uses a coupled dual-model architecture to simultaneously utilize a shared pre-trained encoder for client descriptor extraction and task model feature representation, and obtains a hypernetwork that can directly generate personalized models during the testing phase through decentralized hypernetwork aggregation training. The process is as follows: Figure 1 As shown, the main steps include: S1: Construct a coupled dual-model architecture. The supernetwork and the task model share the same pre-trained encoder; where the supernetwork is represented as... The task model is represented as , Indicates a shared pre-trained encoder. Indicates the first Hypernetwork generator parameters for each training client, Indicates the first Personalized parameters for the task model of each training client.

[0022] S2: Perform decentralized hypernetwork aggregation. In the... In the first round of training, Each training client follows a dynamic communication matrix The hypernetwork parameters of the neighbor training clients are weighted and aggregated to obtain the aggregated hypernetwork parameters. ;in, The decentralized topology weight matrix varies with training rounds and is determined by client movement, link availability, or random rotation mechanisms. After each round of topology determination, each client calculates and normalizes the weights based on its current set of communicable neighbors. The element in the i-th row and j-th column ,in In the qth round, the relationship with the 1st round is... The number of neighbors that a training client can directly communicate with; A value greater than zero indicates that the i-th training client and the j-th training client can communicate directly in this round. Then the i-th training client and the j-th training client cannot communicate in this round. For the hypernet generator parameters of neighbor j.

[0023] S3: Generate client descriptors. The client descriptors are generated from the local data features of the training client. Shared pre-trained encoder Generate, represented as The client descriptor is used to characterize the local data distribution of the training client, enabling the hypernetwork generator to... The output is customized to the task model parameters that adapt to the data distribution of the client, so new test clients do not need to learn the client embedding in advance.

[0024] S4: Use the task model to update the difference and train the supernetwork, including the following steps: S401. Based on the aggregated hypernetwork parameters and client descriptor Generate personalized parameters for the initial task model: ; For the first An intermediate variable in round calculation.

[0025] S402. Personalize the parameters based on the initial task model. Starting from the client Local data batch Execution Step gradient descent to obtain the updated personalized parameters of the task model. .

[0026] S403. Personalize parameters based on the initial task model. and Step-updated parameters Calculate the parameter difference: .

[0027] S404. The parameter difference The training signal and the target signal generated by the supernetwork are then used as the training signal corresponding to the difference. The training signal is then backpropagated to the supernetwork generator parameters using the chain rule to update the local supernetwork. ; in, To exceed the network learning rate, This represents the training objective of the local hypernetwork. For gradient.

[0028] S5: Generate personalized models without fine-tuning during testing, including the following steps: S501. When a new test client is added. When accessing a decentralized network, the test client Establish a communication connection with at least one trained client according to the connection relationship, and obtain the corresponding connection weights through normalization based on the connection relationship. : ; A value greater than zero indicates a test client With the client Connected.

[0029] S502. The test client Receive connected training client The trained hypernetwork parameters The aggregated network parameters corresponding to the test client are obtained by weighting and aggregating the data according to the connection weights. ; S503. The test client Local data characteristics Input consists of the aggregated hypernetwork parameters In the represented hypernetwork, personalized parameters are generated for the test client: .

[0030] S504. Share encoder parameters With the personalized parameters Combine them to obtain the test client. Personalized models: .

[0031] A test-time personalized model generation system for dynamic networks includes a coupled dual-model construction module, a descriptor generation module, a decentralized aggregation module, a hypernetwork training module, and a test-time generation module. Coupled dual-model building module, used to build a shared pre-trained encoder, hypernetwork generator and task model personalized parameters on the training client; The descriptor generation module is used to generate client descriptors based on local data characteristics; The decentralized aggregation module is used to aggregate neighbor supernetwork parameters based on dynamic communication topology; The hypernetwork training module is used to update the hypernetwork parameters in reverse by updating the difference in multiple steps based on the task model; The test-time generation module is used to aggregate the hypernetwork parameters of newly added test clients and generate a personalized model that does not require fine-tuning.

[0032] 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 limited to the embodiments shown herein, but should be accorded the widest scope of protection consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating personalized test-time models for dynamic networks, characterized in that, Includes the following steps: S1: Construct a coupled dual-model architecture on multiple training clients, the coupled dual-model architecture including a shared pre-trained encoder, a hypernetwork generator, and task model personalized parameters; S2: Each training client receives hypernetwork parameters from neighboring training clients based on the decentralized communication topology of the current round, and performs weighted aggregation of the received hypernetwork parameters and local hypernetwork parameters. S3: Each training client uses local data features and the shared pre-trained encoder to generate a client descriptor, and the local hypernetwork generator generates corresponding personalized parameters for the task model based on the client descriptor; S4: Each training client generates local task model personalized parameters based on the aggregated hypernetwork, and performs several steps of task model update using local data. The difference between the task model personalized parameters before and after the update is used to construct the hypernetwork training signal to update the local hypernetwork parameters in reverse. S5: During the testing phase, a new test client connects to one or more training clients, aggregates the trained hypernetwork parameters of the connected training clients, and inputs the local data features of the test client into the aggregated hypernetwork to generate a personalized test client model that can be used without fine-tuning.

2. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, In the coupled dual-model architecture, the hypernetwork and the task model share the same pre-trained encoder; wherein, the hypernetwork is represented as... The task model is represented as , Indicates a shared pre-trained encoder. Indicates the first Hypernetwork generator parameters for each training client, Indicates the first Personalized parameters for the task model of each training client.

3. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, In the In the first round of training, Each training client follows a dynamic communication matrix The hypernetwork generator parameters of the neighbor training clients are weighted and aggregated to obtain the aggregated hypernetwork parameters: ； in, The decentralized topology weight matrix varies with training rounds and is determined by client movement, link availability, or random rotation mechanisms. After each round of topology determination, each client calculates and normalizes the weights based on its current set of communicable neighbors. The element in the i-th row and j-th column ,in In the qth round, the relationship with the 1st round is... The number of neighbors that each training client can directly communicate with; A value greater than zero indicates that the i-th training client and the j-th training client can communicate directly in this round. Then the i-th training client and the j-th training client cannot communicate in this round. For the hypernet generator parameters of neighbor j.

4. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, The client descriptor is derived from the features of the training client's local data. Shared pre-trained encoder Generate, represented as: ； This indicates that the parameters are shared by the pre-trained encoder. The client descriptor is represented by the descriptor generation function; Used to characterize the local data distribution of the training client, enabling the hypernetwork generator to generate data based on the client descriptor. Output personalized parameters for the task model that adapt to the data distribution of this client.

5. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, Step S4 includes the following steps: S401. Based on the aggregated hypernetwork parameters and client descriptor Generate personalized parameters for the initial task model: ； Represents the parameters of the aggregated hypernetwork The hypernetwork generating function is characterized; S402. Personalize the parameters based on the initial task model. Starting from the client Execute on local data batch Step gradient descent to obtain the updated personalized parameters of the task model. ; S403. Personalize parameters based on the initial task model. and the updated task model personalized parameters Calculate the parameter difference: ； S404. The parameter difference As the target and training signal for the supernetwork generation, it is used to guide the backpropagation update of the local supernetwork parameters.

6. The method for generating personalized test models for dynamic networks according to claim 5, characterized in that, The training signal of the hypernetwork is backpropagated from the personalized parameter space of the task model to the parameter space of the hypernetwork through a chain rule, and then follows... Update local hypernetwork parameters; among which, Indicates the supernet learning rate. Indicates the first The training objective of the hypernetwork for each training client.

7. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, During the communication phase, each training client only sends or receives hypernetwork generator parameters and does not need to send complete task model parameters or local data.

8. The method for generating personalized test models for dynamic networks according to claim 1, characterized in that, The personalized model generation in the testing phase of step S5 includes the following steps: S501. When a new test client is added. When accessing a decentralized network, the test client Establish a communication connection with at least one trained client according to the connection relationship, and obtain the corresponding connection weights through normalization based on the connection relationship. : ; A value greater than zero indicates a test client With the client Connected; S502. The test client Receive connected training client The trained hypernetwork parameters The aggregated network parameters corresponding to the test client are obtained by weighting and aggregating the data according to the connection weights. ； S503. The test client Local data characteristics Input consists of the aggregated hypernetwork parameters In the represented hypernetwork, personalized parameters are generated for the test client: ； Indicates the aggregation of hypernetwork parameters The hypernetwork generating function is characterized. This indicates that the parameters are shared by the pre-trained encoder. The descriptor generation function for the representation; S504. Share encoder parameters With the personalized parameters Combine them to obtain the test client. Personalized models: 。 9. The method for generating personalized test models for dynamic networks according to any one of claims 1-8, characterized in that, The test client only needs local data features to generate personalized models. It does not rely on local data labels; the test client obtains a personalized model through one or a small number of forward calculations, without the need to perform backpropagation or gradient fine-tuning during the testing phase.

10. A test-time personalized model generation system for dynamic networks, applicable to the test-time personalized model generation method for dynamic networks as described in any one of claims 1-8, characterized in that, It includes a coupled dual-model construction module, a descriptor generation module, a decentralized aggregation module, a hypernetwork training module, and a test-time generation module; Coupled dual-model building module, used to build a shared pre-trained encoder, hypernetwork generator and task model personalized parameters on the training client; The descriptor generation module is used to generate client descriptors based on local data characteristics; The decentralized aggregation module is used to aggregate neighbor supernetwork parameters based on dynamic communication topology; The hypernetwork training module is used to update the hypernetwork parameters in reverse by updating the difference in multiple steps based on the task model; The test-time generation module is used to aggregate the hypernetwork parameters of newly added test clients and generate a personalized model that does not require fine-tuning.

Images