Method for sharing and updating knowledge between personalized models

Abstract

A method for personalized knowledge sharing and updating among models is proposed. This method decomposes the incremental adapter parameters uploaded by each device into multiple parameter blocks using an edge server. Based on the response strength of each parameter block to the model output and its consistency with the direction of the parameter increment of the target device, a utility score is calculated for each parameter block. Dynamic programming is used to select the optimal subset of parameter blocks, which are then rescaled using a variance matching strategy to generate a merged adapter and distributed to connected devices. Each connected device continuously monitors class increment offset and tag frequency offset, avoids redundant fine-tuning through a caching and reuse mechanism, and dynamically estimates the optimal compression ratio to optimize uploads. This invention enables efficient knowledge sharing among devices under tag frequency offset conditions, significantly reducing communication and computational overhead while maintaining high model accuracy.

Description

Technical Field

[0001] This invention relates to a technology in the field of neural networks, specifically a method for knowledge sharing and updating among personalized models. Background Technology

[0002] Mobile devices commonly employ a "public backbone network + lightweight adapter" architecture to deploy personalized models. Limited by resources, this results in insufficient model generalization ability and performance degradation due to label frequency shifts. Existing single-device adaptive methods do not utilize the correlation of changes between multiple devices and suffer from high communication and computational overhead, making it difficult to balance low latency and lightweight updates. Current federated learning techniques cannot achieve selective knowledge sharing at the parameter level between devices under label frequency shift conditions, nor can they achieve selective merging and cache reuse of parameters between devices under the same conditions. Summary of the Invention

[0003] This invention addresses the aforementioned deficiencies in existing technologies and the inability of existing technologies to meet industrial requirements in terms of accuracy. It proposes a personalized method for knowledge sharing and updating between models, which enables efficient knowledge sharing between devices under tag frequency offset conditions, significantly reducing communication and computing overhead while maintaining high model accuracy.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for knowledge sharing and updating among personalized models. The method involves using an edge server to decompose the incremental adapter parameters uploaded by each device into multiple parameter blocks. Based on the response strength of each parameter block to the model output and its consistency with the direction of the parameter increment of the target device, a utility score is calculated for each parameter block. Dynamic programming is used to select the optimal subset of parameter blocks, which are then rescaled using a variance matching strategy to generate a merged adapter and distributed to connected devices. Each connected device continuously monitors class increment offset and tag frequency offset, avoids redundant fine-tuning through a cache reuse mechanism, and dynamically estimates the optimal compression ratio to optimize uploads.

[0006] The parameter block is obtained by decomposing it in the following way:

[0007] 1) For source devices that have completed local fine-tuning, the edge server calculates the parameter increments of its adapter relative to the initial common adapter. ,in For source device adapter parameter, These are the initial common adapter parameters.

[0008] 2) Increment the parameter The parameters are divided into multiple fine-grained blocks using both row-level and element-level partitioning strategies. Specifically, each row of the weight matrix is ​​treated as a block, and each scalar parameter is treated as a block, in order to balance interpretability, latency, and accuracy in different application scenarios.

[0009] The utility score mentioned above is specifically: Where: b is the parameter block number, , Source devices and target equipment parameter increment, It is the Frobenius norm. For from the first The linear propagation matrix from layer to classification head, For the classifier weights of the newly added categories, Calculate the number of parameters. Cosine similarity is used to measure the consistency of the update direction between the source device and the target device.

[0010] The aforementioned use of dynamic programming to select the optimal subset of parameter blocks refers to modeling the parameter block selection problem as a 0-1 knapsack problem with cost constraints. For each parameter block... Its effect is ,cost This represents the number of parameters within the block. Given the total parameter budget. To determine the compression ratio, dynamic programming is used to find the optimal subset and the utility matrix. The iterative formula is: Indicates selection There are parameter blocks, and the parameter budget is... At that time, the maximum utility that can be obtained by selecting a subset. By backtracking from Recovery of the optimal subset .

[0011] The variance-matching strategy rescaling refers to: for the retained optimal parameter block subset A variance-matching strategy is employed for rescaling to compensate for statistical bias caused by parameter discarding. Specifically, the rescaling factor is... ,in , These are the parameter increments for the target device and the source device, respectively. For indicator functions, For symbolic functions, This indicates the larger update magnitude when the symbols are consistent. The rescaling factor is adaptively adjusted based on the statistical relationship between the source and target devices. When the merged update and the target update have high consistency, the factor approaches 1; when conflicts exist, the factor can be greater than 1 to compensate for the reduction in effective update volume.

[0012] The real-time detection refers to the connected device dividing the data stream into continuous sliding windows to detect incremental offsets and label frequency offsets.

[0013] The aforementioned incremental shift refers to a shift triggered when the model's prediction confidence decreases. The average confidence level within the calculated window is then used. and average forecast margin The uncertainty indicator is: This is a hyperparameter. When U exceeds the preset threshold τ_u, a class increment offset is detected.

[0014] The label frequency shift refers to a condition triggered when the relative frequency of a known category changes. The Jensen-Shannon divergence between the predicted category distributions of the current window and the previous window is calculated. ,in , The predicted category distribution for the current window. This is the predicted category distribution of the previous window. This refers to the KL divergence. When the JS divergence exceeds a preset threshold... At that time, a tag frequency offset was detected.

[0015] The cache reuse mechanism refers to the following: when a class incremental offset or label frequency offset is detected, the connected device first checks if there is a merged update in the local model cache that matches the current offset type. If the applicability is verified, the cache update is directly activated to complete the adaptation after the verification is passed; otherwise, local fine-tuning is triggered.

[0016] The applicability verification includes:

[0017] a) For class increment offset, the verification condition is: This means that cache updates reduce the uncertainty of new classes without degrading the performance of old classes.

[0018] b) For tag frequency offset, the verification condition is: In other words, cache updates reduce the overall uncertainty of the current window.

[0019] The aforementioned dynamic estimation of the optimal compression ratio refers to the process where, when local fine-tuning is triggered, the connected device calculates the optimal compression ratio. Under the condition that the loss of model accuracy does not exceed the preset boundary, the parameter discard ratio is maximized. After fine-tuning, the device uploads the parameter increment, task-specific classification header and compression ratio metadata to the edge server. The edge server performs selective model merging and distributes the merged update to devices that may be affected by the same offset.

[0020] The compression ratio is determined by solving the following constrained optimization problem: The constraints of this problem are: , For the set of parameter blocks to be retained, For the set of discarded parameter blocks, As a lower bound for tolerable accuracy loss, This is the scaling factor for offline measurements. This constraint is based on Taylor expansion derivation and ensures that output variations caused by parameter discarding are kept within acceptable limits.

[0021] This invention relates to a personalized model-to-model knowledge sharing and updating system for implementing the above-mentioned method, comprising: a server-side parameter fusion unit, a device-side drift detection unit, a device-side compression ratio estimation unit, and a device-side cache reuse unit. Specifically: the server-side parameter fusion unit performs parameter block decomposition, utility evaluation, dynamic programming selection, and amplitude adjustment processing based on the adapter parameter increments uploaded by each device, obtaining a fused adapter and distributing it to high-similarity devices; the device-side drift detection unit performs sliding window uncertainty index or JS divergence calculation processing based on the input data stream to obtain the label frequency offset type; the device-side compression ratio estimation unit performs constraint optimization solving processing based on tolerable accuracy loss to obtain the optimal compression ratio and uploads the parameter increments, classification header, and metadata; the device-side cache reuse unit performs cache checking and applicability verification processing based on the drift detection results to obtain cache update activation or local fine-tuning trigger results.

[0022] Technical effect

[0023] Compared with the prior art, this invention decomposes the adapter parameter increment into parameter blocks, and then calculates the block-level utility score for value evaluation by comprehensively considering the response strength of the parameter blocks to the model output, the propagation matrix, the weight of the new class, and the consistency with the direction of the parameter increment of the target device. It also adopts a variance matching strategy and adaptively rescales the retained parameters based on the sign consistency and update magnitude of the parameter increments of the source device and the target device. 3. Different cache verification conditions are set for class increment offset and label frequency offset, where class increment offset requires the simultaneous satisfaction of the dual judgment of reduced uncertainty of the new class and no performance degradation of the old class. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention;

[0025] Figure 2 A schematic diagram showing the detailed process flow of the equipment;

[0026] Figure 3 This is a schematic diagram illustrating utility assessment and parameter selection.

[0027] Figure 4 This is a schematic diagram of edge server parameter fusion. Detailed Implementation

[0028] like Figure 1 and Figure 2 As shown, this embodiment relates to a method for knowledge sharing and updating among personalized models, including:

[0029] Step 1, the initial preparation stage, specifically includes:

[0030] 1.1 Initial State Setup: Data from different domains is assigned to the four connected devices Device1-4. Each device's personalized model undergoes initial fine-tuning based on the data in its assigned domain until the adapter parameters converge.

[0031] 1.2 Edge Server Model Pool Initialization: The edge server collects the adapter parameters of each device and calculates the parameters relative to the initial common adapter. The parameter increment.

[0032] 1.3 Pre-calculation of linear propagation matrix: For each layer of the model, the edge server pre-calculates the linear propagation matrix. (Obtained by multiplying the weight matrices of all subsequent layers), this matrix depends only on the frozen backbone network and can be computed and stored offline.

[0033] Step 2, Tag Frequency Offset Simulation and Device Adaptation, specifically includes:

[0034] 2.1 Offset Simulation: After running for a certain period of time (5 minutes), the first connected device, Device1, introduces new categories (5 new object categories), simulating incremental class offset. The local data distribution of the first connected device, Device1, changes, and the performance of the original model degrades on the new categories.

[0035] 2.2 Offset Detection and Buffer Check: The offset detector of the first connected device, Device1, calculates the uncertainty index using a sliding window of 50 samples. .when When it drops from 0.85 to 0.62, U exceeds the threshold. An incremental offset was detected. A check of the local cache revealed no matching update, triggering local fine-tuning.

[0036] 2.3 Local Fine-tuning: The first connected device, Device1, uses the new data (including newly added categories) to fine-tune the adapter, and the updated adapter is obtained after the fine-tuning is completed. Calculate parameter increments .

[0037] 2.4 Compression Ratio Estimation: The compression estimator of the first connected device, Device1, calculates the optimal compression ratio based on the constraints. For η=0.05 (allowing 5% accuracy loss), dynamic programming yields p_max≈15, meaning only about 6.5% of the original adapter parameters need to be transmitted.

[0038] 2.5 Parameter Upload: The first connected device, Device1, uploads the updated parameter increment, task header parameters, and compression ratio metadata to the edge server.

[0039] Step 3, selective model merging of edge servers, specifically includes:

[0040] 3.1 Parameter Increment Reception: The edge server receives the parameter increment uploaded by the first connected device Device1 and determines the current adapter status of other target devices (the second to fourth connected devices Device2, Device3, and Device4).

[0041] 3.2 Utility Score Calculation: For example... Figure 3 As shown, for each parameter block b, the edge server calculates its utility score relative to each target device. Taking the second connected device, Device2, as an example, the specific steps are as follows: For the increment of the first connected device Device1, For the increment of the second connected device Device2, It reflects the strength of the classifier weights for the newly added categories.

[0042] 3.3 Dynamic Programming Selection: Given a compression ratio Total parameter budget We use 0-1 knapsack dynamic programming to select the subset of parameter blocks that maximizes total utility. .

[0043] 3.4 Parametric rescaling: for The parameter block in the calculation calculates the rescaling factor. Since the update directions of the first connected device Device1 and the second connected device Device2 are highly consistent, the calculated r ≈ 0.98, which is close to 1, indicating that no significant scaling is required.

[0044] 3.5 Merging, updating, generating, and distributing: For example... Figure 4 As shown, the edge server combines the rescaled parameter blocks into a merged adapter. Depend on The parameters are composed of blocks. The merged adapter is then distributed to the second to fourth connected devices, Device2-4.

[0045] Step 4, target device cache update and verification, specifically includes:

[0046] 4.1 Buffer Reception: The second connected device, Device2, receives the merging adapter distributed by the edge server. Stored in the local model cache.

[0047] 4.2 Applicability Verification: When the second connected device, Device2, detects a class increment offset, it activates a cache check and calculates the specific verification conditions as follows: Once the verification is successful, the second connected device, Device2, will directly activate the cache update without requiring local fine-tuning.

[0048] 4.3 Performance Evaluation: After the second connected device, Device2, activated the cache update, the accuracy in the new category improved from 0.42 to 0.61, while the accuracy in the old category remained above 0.85, resulting in an overall accuracy improvement of 13.5%.

[0049] Through practical application experiments, the personalized model knowledge sharing and update implementation method of this invention was tested on NVIDIA RTX 4090 GPU servers and NVIDIA Jetson Orin Nano mobile devices, using the PyTorch framework. The results were evaluated on four public datasets: DomainNet, GLUE, RealWorld HAR, and Multilingual LibriSpeech. Specifically, the evaluation included:

[0050] Step 1, offline preparation stage, specifically includes:

[0051] 1.1 Initial state setup: Data from different domains is assigned to the four connected devices. The personalized model of each device is initially fine-tuned based on the data in its assigned domain, and the adapter parameters converge.

[0052] 1.2 Edge Server Model Pool Initialization: The edge server collects the adapter parameters of each device and calculates the parameter increments relative to the initial common adapter.

[0053] 1.3 Pre-calculation of linear propagation matrix: For each layer of the model, the edge server pre-calculates the linear propagation matrix. This matrix relies solely on the frozen backbone network and can be computed and stored offline.

[0054] Step 2, Tag Frequency Offset Simulation and Device Adaptation, specifically includes:

[0055] 2.1 Offset Simulation: After running for a certain period of time, the first connected device introduces a new category to simulate incremental offset, the local data distribution changes, and the performance of the original model degrades on the new category.

[0056] 2.2 Offset Detection and Cache Check: The offset detector of the first connected device calculates the uncertainty index using a sliding window. When the uncertainty exceeds the threshold, it detects a kind of incremental offset, checks the local cache and finds no matching update, triggering local fine-tuning.

[0057] 2.3 Local fine-tuning: The first connected device uses data containing the newly added categories to fine-tune the adapter. After the fine-tuning is completed, the updated adapter is obtained, and the parameter increment is calculated.

[0058] 2.4 Compression Ratio Estimation: The compression estimator of the first connected device calculates the optimal compression ratio based on the constraints, and obtains the maximum compression ratio through dynamic programming.

[0059] 2.5 Parameter Upload: The first connected device uploads the updated parameter increment, task header parameters, and compression ratio metadata to the edge server.

[0060] Step 3: Selective model merging on the edge server side, specifically including:

[0061] 3.1 Parameter Increment Reception: The edge server receives the parameter increment uploaded by the first connected device and determines the current adapter status of other target devices.

[0062] 3.2 Utility Score Calculation: For each parameter block, the edge server calculates its utility score relative to each target device.

[0063] 3.3 Dynamic Programming Selection: Given a compression ratio, select the subset of parameter blocks that maximizes the total utility using 0-1 knapsack dynamic programming.

[0064] 3.4 Parameter rescaling: Calculate the rescaling factor for the selected parameter block subset and generate the merged adapter.

[0065] 3.5 Merge Update Distribution: Distribute the merged adapter to other connected devices.

[0066] Step 4, target device cache update and verification, specifically includes:

[0067] 4.1 Cache Reception: The target device receives the merge adapter distributed by the edge server and stores it in the local model cache.

[0068] 4.2 Applicability Verification: When the target device itself detects a class incremental offset, it activates a cache check. If the verification is successful, the cache update is activated directly without local fine-tuning.

[0069] 4.3 Performance Evaluation: After the target device activates the cache update, the accuracy on the new category is significantly improved, while the accuracy on the old category remains stable. Evaluation on the DomainNet test set shows that in a four-device collaborative scenario, the proposed method achieves an image classification accuracy of 55.76% with an update latency of only 2.59 seconds, significantly outperforming existing baseline methods. Experimental results demonstrate that the proposed method achieves an accuracy of 55.76% in image classification tasks, a 4.5% improvement over the best baseline; an accuracy of 57.03% in natural language understanding tasks, a 15.7% improvement over the best baseline; an accuracy of 36.28% in human behavior recognition tasks, a 2.4% improvement over the best baseline; and a word error rate of 30.17% in automatic speech recognition tasks, a 13.5% reduction over the best baseline. The average update latency is reduced by 61% compared to traditional fine-tuning, and the communication parameters are compressed to 1 / 3000 of the original model, achieving a 5-fold compression rate improvement over the best baseline Ties-Merging. The technical effects are reproducible. The implementation process of the proposed method is illustrated using the publicly available multi-domain image classification dataset DomainNet as an example, and the experimental results are shown in Table 1.

[0070] Table 1

[0071] Compared to existing technologies, this invention achieves accurate identification of the most valuable parameters for tag frequency offset by calculating the utility score of parameter blocks containing cosine similarity on the edge server side, combined with a pre-computed linear propagation matrix. It also achieves efficient compression of communication parameters by modeling parameter block selection as a 0-1 knapsack problem and using dynamic programming to solve for the optimal subset, combined with a variance-matching rescaling strategy. On the edge device, a dual-type tag frequency offset detection and cache reuse verification mechanism ensure that only high-quality updates are applied, avoiding redundant fine-tuning. These technological innovations work together to enable effective optimization knowledge sharing between personalized models, achieving rapid adaptation under tag frequency offset conditions with extremely low communication and computational overhead, providing key technical support for large-scale edge intelligent systems.

[0072] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for knowledge sharing and updating among personalized models, characterized in that, include: The incremental adapter parameters uploaded by each device are decomposed into multiple parameter blocks by the edge server. The utility score of each parameter block is calculated based on the response strength of the parameter block to the model output and its consistency with the direction of the parameter increment of the target device. Dynamic programming is used to select the optimal subset of parameter blocks. After rescaling by the variance matching strategy, a merged adapter is generated and distributed to the connected devices. Each connected device detects the class increment offset and tag frequency offset in real time. Redundant fine-tuning is avoided through a cache reuse mechanism, and the optimal compression ratio is dynamically estimated to optimize the upload.

2. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The parameter block is obtained by decomposing it in the following way: 1) For source devices that have completed local fine-tuning, the edge server calculates the parameter increments of its adapter relative to the initial common adapter. ,in For source device adapter parameter, These are the initial common adapter parameters; 2) Increment the parameter The parameters are divided into multiple fine-grained blocks using both row-level and element-level partitioning strategies. Specifically, each row of the weight matrix is ​​treated as a block, and each scalar parameter is treated as a block, in order to balance interpretability, latency, and accuracy in different application scenarios.

3. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The utility score mentioned above is specifically: Where: b is the parameter block number, , Source devices and target equipment parameter increment, It is the Frobenius norm. For from the first The linear propagation matrix from layer to classification head, For the classifier weights of the newly added categories, Calculate the number of parameters. Cosine similarity is used to measure the consistency of the update direction between the source device and the target device.

4. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The aforementioned use of dynamic programming to select the optimal subset of parameter blocks refers to modeling the parameter block selection problem as a 0-1 knapsack problem with cost constraints, whereby for each parameter block... Its effect is ,cost Given the number of parameters within the block, and a total parameter budget. To determine the compression ratio, dynamic programming is used to find the optimal subset and the utility matrix. The iterative formula is: Indicates selection There are parameter blocks, and the parameter budget is... At that time, the maximum utility that can be obtained by selecting a subset is determined by backtracking from... Recovery of the optimal subset .

5. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The variance-matching strategy rescaling refers to: for the retained optimal parameter block subset A variance-matching strategy is employed for rescaling to compensate for statistical bias caused by parameter discarding. Specifically, the rescaling factor is... ,in , These are the parameter increments for the target device and the source device, respectively. For indicator functions, For symbolic functions, The larger update magnitude between the two when the symbols are consistent is indicated. This rescaling factor is adaptively adjusted according to the statistical relationship between the source device and the target device. When the consistency between the merged update and the target update is high, the factor approaches 1. When there is a conflict, the factor can be greater than 1 to compensate for the reduction in the effective update volume.

6. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The real-time detection refers to the connected device dividing the data stream into continuous sliding windows to detect incremental offset and label frequency offset. The aforementioned incremental shift refers to the calculation of the average confidence level within a window when the model's prediction confidence decreases. and average forecast margin The uncertainty indicator is: This is a hyperparameter; when U exceeds a preset threshold τ_u, a class increment offset is detected. The aforementioned label frequency offset refers to the calculation of the Jensen-Shannon divergence between the predicted category distributions of the current window and the previous window, triggered when the relative frequency of known categories changes. ,in , The predicted category distribution for the current window. This is the predicted category distribution of the previous window. For KL divergence, when the JS divergence exceeds a preset threshold... At that time, a tag frequency offset was detected.

7. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The cache reuse mechanism refers to the following: when a class incremental offset or label frequency offset is detected, the connected device first checks if there is a merged update in the local model cache that matches the current offset type, performs an applicability verification, and after the verification is passed, directly activates the cache update to complete the adaptation. Otherwise, trigger local fine-tuning; The applicability verification includes: a) For class increment offset, the verification condition is: In other words, cache updates reduce the uncertainty of new classes without degrading the performance of old classes; b) For tag frequency offset, the verification condition is: In other words, cache updates reduce the uncertainty of the current window as a whole.

8. The method for knowledge sharing and updating among personalized models according to claim 1, characterized in that, The aforementioned dynamic estimation of the optimal compression ratio refers to the process where, when local fine-tuning is triggered, the connected device calculates the optimal compression ratio. Under the condition that the loss of model accuracy does not exceed the preset boundary, the parameter discard ratio is maximized. After fine-tuning, the device uploads the parameter increment, task-specific classification header and compression ratio metadata to the edge server. The edge server performs selective model merging and distributes the merged update to devices that may be affected by the same offset.

9. The method for implementing knowledge sharing and updating among personalized models according to claim 1 or 8, characterized in that, The compression ratio is determined by solving the following constrained optimization problem: The constraints of this problem are: , For the set of parameter blocks to be retained, For the set of discarded parameter blocks, As a lower bound for tolerable accuracy loss, The scaling factor for offline measurements is a constraint derived from Taylor expansion, ensuring that output variations caused by parameter discarding are kept within acceptable limits.

10. A personalized model-to-model knowledge sharing and updating system for implementing the method of any one of claims 1-9, characterized in that, The system includes: a server-side parameter fusion unit, a device-side drift detection unit, a device-side compression ratio estimation unit, and a device-side cache reuse unit. Specifically: the server-side parameter fusion unit performs parameter block decomposition, utility evaluation, dynamic programming selection, and amplitude adjustment based on the adapter parameter increments uploaded by each device, obtaining a fused adapter and distributing it to high-similarity devices; the device-side drift detection unit performs sliding window uncertainty index or JS divergence calculation based on the input data stream to obtain the label frequency offset type; the device-side compression ratio estimation unit performs constraint optimization based on tolerable accuracy loss to obtain the optimal compression ratio and uploads the parameter increments, classification header, and metadata; and the device-side cache reuse unit performs cache checks and applicability verification based on the drift detection results to obtain cache update activation or local fine-tuning trigger results.

Images