Parallel split learning method, device and equipment based on meta-network gradient alignment

Abstract

The application provides a parallel split learning method, device and equipment based on meta-network gradient alignment, and relates to the technical field of distributed machine learning. The application receives intermediate features uploaded by each client through a server, constructs an unbiased query set with balanced categories through hierarchical sampling; calculates the gradient alignment score of the gradient of the client and the gradient of the query set by using a lightweight meta-network; generates normalized importance weights in combination with a dynamic temperature parameter; and performs weighted scaling on the client gradient in back propagation, thereby physically blocking noise gradient. The application can effectively alleviate gradient conflicts and training bias caused by non-independent and identically distributed and long-tail distribution without accessing original data, and improves the convergence stability and generalization ability of the model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and distributed machine learning technology, and more specifically, to a parallel segmentation learning method, apparatus, and device based on meta-network gradient alignment. Background Technology

[0002] Edge intelligence, as a key paradigm connecting terminal devices and cloud computing, is increasingly becoming an important development direction for distributed artificial intelligence. How to efficiently train deep learning models while ensuring data privacy and reducing communication overhead has become a core problem that urgently needs to be solved. Federated learning, through its "data stays put, model moves" mechanism, trains the complete model locally on the client and only uploads model parameter updates, effectively alleviating the problems of data privacy leakage and transmission bottlenecks. However, federated learning requires each client to have the ability to store and compute the complete neural network, which poses a significant hardware barrier for resource-constrained edge devices (such as miniature sensors and wearable devices). Although techniques such as model compression, pruning, or quantization can alleviate resource pressure to some extent, they often come at the cost of model accuracy, making it difficult to balance performance and efficiency.

[0003] To address these challenges, split learning and its derivative paradigm—split federated learning—have been proposed. This method splits a deep neural network along its layers into client-side sub-models (typically shallow feature extractors) and server-side sub-models (responsible for high-level semantic inference). The client only needs to perform lightweight forward computation and upload intermediate features to the server, significantly reducing memory usage (up to 99%) and computational load, making it more suitable for resource-constrained scenarios. However, traditional split learning employs a serial processing mechanism, where the server processes intermediate features uploaded by each client sequentially. This results in training latency increasing linearly with the number of clients, severely limiting system scalability.

[0004] To improve training efficiency, Parallel Split Learning (PSL) has emerged. This method fully utilizes the parallel computing power of the server, simultaneously receiving and processing intermediate feature streams uploaded by multiple clients. This means that the training time for each round is determined only by the slowest client, significantly shortening the overall training cycle and approaching the parallel efficiency of federated learning. However, this parallel mechanism of aggregation in the feature space introduces new technical challenges: in non-independent identically distributed (Non-IID) data environments, due to significant differences in local data distribution among different clients, the intermediate features they generate often lie on incompatible feature manifolds. This leads to severe divergence in gradient directions during backpropagation, even resulting in negative correlations and compromising model convergence stability. Even more serious is the fact that real-world data generally exhibits a long-tail distribution, meaning that a small number of "head class" samples constitute the vast majority, while a large number of "tail class" samples exist sparsely. In parallel training batches, the strong gradient signals generated by head class samples easily mask the weak gradients of tail classes, causing the aggregated gradient direction to be severely biased towards the mainstream class, resulting in a sharp decline in the model's generalization ability to rare classes.

[0005] Existing optimization strategies mostly focus on the parameter aggregation stage in federated learning, such as improving performance under Non-IID through adaptive weighting or personalized adjustments. However, these methods cannot be directly transferred to the feature interaction layer of parallel segmentation learning. At the same time, current research on segmentation learning mainly focuses on communication compression, segmentation point selection, or privacy enhancement, and has not yet formed a systematic solution to the problems of statistical incompatibility and long-tail bias in feature space.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] This invention aims to provide a parallel segmentation learning method, apparatus, device, and medium based on meta-network gradient alignment, in order to solve the technical problems of existing parallel segmentation learning in non-independent identically distributed (Non-IID) and long-tailed data distribution scenarios, such as inconsistent gradient directions, amplified training bias, and unstable model convergence caused by feature space aggregation.

[0008] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0009] A parallel segmentation learning method based on meta-network gradient alignment, applied to the server side, includes:

[0010] S1 receives intermediate features uploaded by multiple clients and constructs a class-balanced query set as a global unbiased optimization target through hierarchical sampling;

[0011] S2 utilizes a lightweight meta-network to compute the gradient alignment score between each client gradient and the query set gradient;

[0012] S3, Based on the gradient alignment score combined with the dynamic temperature parameter, generate the importance weight of the client;

[0013] S4. In backpropagation, the gradients of each client are weighted and scaled according to the importance weights to filter conflicting gradients, aggregate valid gradients, and update the global model broadcast to each client.

[0014] Preferably, the formula for constructing the query set is:

[0015] ;

[0016] in, For query set; Total number of categories; Represents the union; For sampling operations; Aggregated features that are intermediate features; As an intermediate feature; c represents the sample label; c represents the category. Set a sampling budget for each type.

[0017] Preferably, a lightweight meta-network is used to calculate the gradient alignment score between each client gradient and the query set gradient, specifically:

[0018] First, a lightweight meta-network is used to find the optimal weights so that the model can optimize the query set. Minimize the loss on;

[0019] Next, a first-order Taylor expansion is performed based on the loss of the query set to quantify the contribution of the client gradient to the global optimization. The formula is as follows:

[0020] ;

[0021] in, For query set loss; The weight corresponding to the k-th client; For the derivative sign; It is approximately proportional; Here are the global model parameters for round t; T is the transpose symbol. for The gradient of the query set loss, i.e., the global unbiased gradient; Let be the gradient of the k-th client in the t-th round, i.e., the client-local gradient; The global unbiased gradient corresponding to the query set; Let be the local loss gradient for the k-th client;

[0022] Then, the vector dot product of the global unbiased gradient corresponding to the query set and the local gradient of the client is calculated to obtain the gradient alignment score, expressed as:

[0023] ;

[0024] in, The gradient alignment score for the k-th client;

[0025] Finally, the gradient value of the client is determined based on the gradient alignment score: when This indicates that the client gradient is aligned with the global direction, thus increasing the weight; when This indicates that the client gradient conflicts with the global direction, thus reducing the weight.

[0026] Preferably, the formula for the importance weight is:

[0027] ;

[0028] in, The normalized importance weight for the k-th client; The dynamic temperature parameter for round t; Total number of clients; , Let be the gradient alignment score for the k-th and j-th clients.

[0029] Preferably, the gradients of each client are weighted and scaled according to the importance weights to filter conflicting gradients, aggregate valid gradients, and update the global model, specifically as follows:

[0030] Based on the aforementioned importance weights, a weighted aggregation loss function is constructed, the formula of which is:

[0031] ;

[0032] in, For weighted aggregation loss function; Total number of clients; The normalized importance weight for the k-th client; The local loss for the k-th client; These are the global model parameters for round t;

[0033] The weighted aggregate loss is backpropagated to the client using the chain rule, and the client gradient is scaled using the weights. The expression is as follows:

[0034] ;

[0035] in, The scaled gradient is backpropagated to the k-th client; The intermediate features uploaded by the kth client; Let be the local loss gradient for the k-th client;

[0036] For clients with conflicting gradient directions, importance weights If it approaches 0, then It also tends to 0 to achieve physical blocking of the noise gradient and complete robust gradient calibration.

[0037] The present invention also provides a parallel segmentation learning device based on meta-network gradient alignment, comprising:

[0038] An unbiased target construction unit is used to receive intermediate features uploaded by multiple clients and construct a class-balanced query set as a global unbiased optimization target through hierarchical sampling.

[0039] The gradient alignment score calculation unit is used to calculate the gradient alignment score between each client gradient and the query set gradient using a lightweight meta-network;

[0040] Importance weight generation unit, used to generate importance weights for the client based on the gradient alignment score combined with dynamic temperature parameters;

[0041] The weighted backpropagation unit is used to weight and scale the gradients of each client according to the importance weights during backpropagation, so as to filter conflicting gradients, aggregate valid gradients, update the global model and broadcast it to each client.

[0042] The present invention also provides a parallel segmentation learning device based on meta-network gradient alignment, including a processor and a memory. The memory stores a computer program that can be executed by the processor to implement the parallel segmentation learning method based on meta-network gradient alignment as described above.

[0043] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium resides, implement a parallel segmentation learning method based on meta-network gradient alignment as described above.

[0044] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention constructs a query set on the server side as an unbiased reference target and introduces a lightweight meta-network to dynamically evaluate the contribution of each client gradient to the global optimization objective, thereby generating importance weights. During backpropagation, it achieves physical weighting of gradients and noise blocking, making the parallel segmentation learning training process robust in scenarios with non-independent and identically distributed and long-tailed data distributions. This solves the training instability problem caused by statistical incompatibility in the feature space without accessing the original client data.

[0046] Since this invention is deployed on the server side, the client does not need to modify the original segmentation learning protocol, but only needs to upload intermediate features, thus maintaining low resource overhead and high parallel efficiency; the entire training process does not involve the exchange of original data, but only operates at the intermediate feature level, strictly protecting data privacy.

[0047] The gradient alignment score calculation used in this invention is based on a first-order Taylor expansion, which has a clear mathematical derivation basis. The annealing weight strategy adjusts the balance between exploration and utilization through dynamic temperature parameters, ensuring training stability and convergence. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of a parallel segmentation learning method based on meta-network gradient alignment provided in Example 1.

[0050] Figure 2 This is a schematic diagram of a parallel segmentation learning device based on meta-network gradient alignment provided in Embodiment 2.

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] Embodiment 1 of the present invention provides a parallel segmentation learning method based on meta-network gradient alignment, which can be implemented by a parallel segmentation learning device based on meta-network gradient alignment (hereinafter referred to as parallel segmentation learning device), specifically, executed by one or more processors within the parallel segmentation learning device.

[0055] In this embodiment, the parallel segmentation learning device can be an electronic device equipped with a processor. The processor carries a computer program for the parallel segmentation learning method based on meta-network gradient alignment, and the computer program can be executed. Examples include computers, smartphones, smart tablets, workstations, etc., which are not limited here.

[0056] This system architecture includes multiple clients and a central server. Each client deploys a shallow sub-model, locally holding a private dataset that is not independent and identically distributed (i.i.d.) and exhibits a long-tail distribution. Each client only performs forward propagation and uploads intermediate features to the server, without transmitting the original input data or model parameters. The central server integrates four functional modules: unbiased target construction, gradient alignment score calculation, importance weight generation, and weighted backpropagation. All data interaction is limited to intermediate feature tensors; the original data remains locally on the client, satisfying the privacy requirement of "data usable but not visible."

[0057] Specifically, such as Figure 1 As shown, a parallel segmentation learning method based on meta-network gradient alignment is applied to the server side, which includes steps S1 to S4.

[0058] S1 receives intermediate features uploaded by multiple clients and constructs a class-balanced query set as a global unbiased optimization objective through hierarchical sampling.

[0059] In each training round of parallel segmentation learning, the server receives intermediate feature data uploaded by K clients, performs format alignment and outlier filtering on the features, and obtains aggregated features. Based on the sample labels uploaded synchronously by the clients or the high-confidence pseudo-labels generated by the server, the class of each feature sample is labeled.

[0060] To address the feature distribution bias caused by data heterogeneity and long-tail distribution, this embodiment performs hierarchical sampling in the feature space.

[0061] Set a sampling budget for each category and stratify by category. For each category, select a subset of features belonging to category c from the aggregated features, and extract features from each category subset using a fixed sampling budget to ensure a consistent sampling quantity for each category. Merge the sampling results of all categories to obtain a category-balanced unbiased query set.

[0062] The formula for constructing the query set is:

[0063] ;

[0064] in, For query set; Total number of categories; Represents the union; For sampling operations; Aggregated features that are intermediate features; As an intermediate feature; c represents the sample label or a high-confidence pseudo-label; c represents the category. Set a sampling budget for each type.

[0065] This operation forces the class probability density in the query set to be... It satisfies a uniform distribution, that is: , This represents the sample category value.

[0066] To prove the validity of the query set, this embodiment is based on the following unbiased derivation:

[0067] In the original training stream, gradient expectation is dominated by a long-tailed distribution. However, in the query set constructed in this invention... Above, the expected gradient is transformed into:

[0068] ;

[0069] in, For query set Gradient expectation on; For loss function The gradient; These are the feature samples corresponding to category c; This is the model prediction function.

[0070] The above formula shows that, through resampling of the feature space, The weighted bias introduced by prior class probability was successfully eliminated, and the resulting gradient direction can be used as an unbiased estimate of the gradient under the global ideal distribution.

[0071] S2 uses a lightweight meta-network to compute gradient alignment scores between each client gradient and the query set gradient.

[0072] This step utilizes a meta-network to dynamically evaluate the quality of client gradients. To quantify the contribution of client gradients to the global optimization objective, this embodiment employs a two-layer optimization framework to quantify the directional consistency between the client's local gradients and the global unbiased gradients, providing a basis for weight allocation.

[0073] Specifically, firstly, a lightweight meta-network (such as a simplified MAML structure containing only 2-3 fully connected layers to reduce computational overhead) is used to find the optimal weights, enabling the model to optimize the query set. Minimize the loss on;

[0074] Next, a first-order Taylor expansion is performed based on the loss of the query set to quantify the contribution of the client gradient to the global optimization. The formula is as follows:

[0075] ;

[0076] in, For query set loss; The weight corresponding to the k-th client; For the derivative sign; It is approximately proportional; Here are the global model parameters for round t; T is the transpose symbol. for The gradient of the query set loss, i.e., the global unbiased gradient; Let be the gradient of the k-th client in the t-th round, i.e., the client-local gradient; The global unbiased gradient corresponding to the query set; Let be the local loss gradient for the k-th client;

[0077] Then, the vector dot product of the global unbiased gradient corresponding to the query set and the local gradient of the client is calculated to obtain the gradient alignment score, expressed as:

[0078] ;

[0079] in, Let be the gradient alignment score for the k-th client.

[0080] It can also validate the gradient alignment scores, remove outliers (such as extreme scores caused by client calculation errors), and retain valid scores for subsequent weight generation.

[0081] Finally, the gradient value of the client is determined based on the gradient alignment score: when This indicates that the client-side gradient is aligned with the global direction, and the weights should be increased; when (Such as noise or adversarial examples) indicates that the client gradient conflicts with the global direction, and the weights should be reduced.

[0082] S3, based on the gradient alignment score combined with the dynamic temperature parameter, generate the importance weight of the client.

[0083] To balance the "exploration" needs in the early stages of training with the "utilization" needs in the later stages, this embodiment does not directly use alignment scores. Instead, it introduces dynamic temperature parameters to modulate the Softmax function, generating client importance weights that take into account both discriminative power and training stability.

[0084] Therefore, the formula for the importance weight is:

[0085] ;

[0086] in, The normalized importance weight for the k-th client; Let t be the dynamic temperature parameter for the t-th round (e.g., decreasing from 1.0 to 0.1). Total number of clients; , Let be the gradient alignment score for the k-th and j-th clients.

[0087] For example, dynamic temperature parameters The formula can be set as:

[0088] ;

[0089] in, The initial temperature (an empirical value is usually taken as 5-10). t represents the temperature decay coefficient; t represents the training round. It is an exponential function.

[0090] In high In this phase, the weight distribution is smooth, allowing more clients to participate in the update and enhancing generalization; in low... In this stage, the weight distribution is sharp, and only the high-fidelity gradient that is highly consistent with the global direction is retained, which accelerates convergence.

[0091] S4. In backpropagation, the gradients of each client are weighted and scaled according to the importance weights to filter conflicting gradients, aggregate valid gradients, and update the global model broadcast to each client.

[0092] Obtaining dynamic weights Then, the server performs weighted scaling on the gradients of each client according to the importance weights to filter conflicting gradients, aggregate effective gradients, update the global model, complete a training loop, and synchronize it to the client.

[0093] Specifically, a weighted aggregation loss function is first constructed based on the aforementioned importance weights, as shown in the formula:

[0094] ;

[0095] in, For weighted aggregation loss function; Total number of clients; The normalized importance weight for the k-th client; The local loss for the k-th client; These are the global model parameters for round t;

[0096] The weighted aggregate loss is backpropagated to the client using the chain rule, and the client gradient is scaled using the weights. The expression is as follows:

[0097] ;

[0098] in, The scaled gradient is backpropagated to the k-th client; The intermediate features uploaded by the kth client; Let be the local loss gradient for the k-th client;

[0099] For clients with conflicting gradient directions, importance weights If it approaches 0, then It also tends to 0 to achieve physical blocking of the noise gradient and complete robust gradient calibration.

[0100] Then, stochastic gradient descent (SGD) or an adaptive optimizer (such as Adam) is used to update the global model parameters. The updated global model parameters are broadcast to all clients, synchronizing the current round's weight and alignment score statistics, providing a foundation for the next round of client-side local training and intermediate feature uploading.

[0101] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0102] This invention performs dynamic evaluation and calibration of the gradient quality on the server side, effectively solving the problems of gradient divergence and training bias caused by statistical incompatibility of feature space in environments with non-independent and identically distributed (Non-IID) and long-tailed data distributions.

[0103] The entire process requires no modification to the client's original protocol, only the uploading of intermediate features, maintaining low resource overhead and high parallel efficiency. At the same time, all operations are completed at the intermediate feature level, without involving the exchange of original data, thus strictly protecting data privacy. The gradient alignment calculation adopted is based on the first-order Taylor expansion, which has a clear mathematical derivation basis. The annealing weight strategy adjusts the balance between exploration and utilization through dynamic temperature parameters, ensuring training stability and convergence.

[0104] In practical applications, such as smart medical edge device clusters or industrial IoT sensor networks, this invention can effectively improve the robustness and generalization ability of models under heterogeneous long-tail data while protecting patient privacy or enterprise data security.

[0105] Example 2

[0106] like Figure 2 As shown, the second embodiment of the present invention also provides a parallel segmentation learning device based on meta-network gradient alignment, comprising:

[0107] An unbiased target construction unit is used to receive intermediate features uploaded by multiple clients and construct a class-balanced query set as a global unbiased optimization target through hierarchical sampling.

[0108] The gradient alignment score calculation unit is used to calculate the gradient alignment score between each client gradient and the query set gradient using a lightweight meta-network;

[0109] Importance weight generation unit, used to generate importance weights for the client based on the gradient alignment score combined with dynamic temperature parameters;

[0110] The weighted backpropagation unit is used to weight and scale the gradients of each client according to the importance weights during backpropagation, so as to filter conflicting gradients, aggregate valid gradients, update the global model and broadcast it to each client.

[0111] Example 3

[0112] The third embodiment of the present invention also provides a parallel segmentation learning device based on meta-network gradient alignment, which includes a memory and a processor. The memory stores a computer program that can be executed by the processor to implement the parallel segmentation learning method based on meta-network gradient alignment as described above.

[0113] Example 4

[0114] The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, they implement the parallel segmentation learning method based on meta-network gradient alignment as described above.

[0115] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the invention. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0116] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0117] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "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. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0118] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0119] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0120] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0121] The use of "first" and "second" in the embodiments is merely to distinguish similar objects and does not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A parallel segmentation learning method based on meta-network gradient alignment, applied to the server side, characterized in that, include: S1 receives intermediate features uploaded by multiple clients, and through... Hierarchical sampling constructs a class-balanced query set as the global unbiased optimization objective; S2 utilizes a lightweight meta-network to compute the gradient alignment score between each client gradient and the query set gradient; S3, Based on the gradient alignment score combined with the dynamic temperature parameter, generate the importance weight of the client; S4. In backpropagation, the gradients of each client are weighted and scaled according to the importance weights to filter conflicting gradients, aggregate valid gradients, and update the global model and broadcast it to each client. Specifically, a lightweight meta-network is used to calculate the gradient alignment score between each client gradient and the query set gradient, as follows: First, a lightweight meta-network is used to find the optimal weights so that the model can optimize the query set. Minimize the loss on; Next, a first-order Taylor expansion is performed based on the loss of the query set to quantify the contribution of the client gradient to the global optimization. The formula is as follows: ； in, For query set loss; The weight corresponding to the k-th client; For the derivative sign; It is approximately proportional; Here are the global model parameters for round t; T is the transpose symbol. for The gradient of the query set loss, i.e., the global unbiased gradient; Let be the gradient of the k-th client in the t-th round, i.e., the client-local gradient; The global unbiased gradient corresponding to the query set; Let be the local loss gradient for the k-th client; Then, the vector dot product of the global unbiased gradient corresponding to the query set and the local gradient of the client is calculated to obtain the gradient alignment score, expressed as: ； in, The gradient alignment score for the k-th client; Finally, the gradient value of the client is determined based on the gradient alignment score: when This indicates that the client gradient is aligned with the global direction, thus increasing the weight; when This indicates a conflict between the client-side gradient and the global direction, leading to a reduction in weights. The formula for the importance weight is: ； in, The normalized importance weight for the k-th client; Here are the dynamic temperature parameters for the t-th round; Total number of clients; , The gradient alignment score for the k-th and j-th clients; The gradients of each client are weighted and scaled according to the aforementioned importance weights to filter conflicting gradients, aggregate valid gradients, and update the global model, specifically as follows: Based on the aforementioned importance weights, a weighted aggregation loss function is constructed, the formula of which is: ； in, For weighted aggregation loss function; Total number of clients; The normalized importance weight for the k-th client; The local loss for the k-th client; The weighted aggregate loss is backpropagated to the client using the chain rule, and the client gradient is scaled using the weights. The expression is as follows: ； in, The scaled gradient is backpropagated to the k-th client; The intermediate features uploaded by the kth client; For clients with conflicting gradient directions, importance weights If it approaches 0, then It also tends to 0 to achieve physical blocking of the noise gradient and complete robust gradient calibration.

2. The parallel segmentation learning method based on meta-network gradient alignment according to claim 1, characterized in that... The formula for constructing the query set is: ； in, For query set; This represents the total number of categories; Represents the union; For sampling operations; Aggregated features that are intermediate features; As an intermediate feature; c represents the sample label; c represents the category. Set a sampling budget for each type.

3. A parallel segmentation learning device based on meta-network gradient alignment, used to implement the parallel segmentation learning method based on meta-network gradient alignment as described in any one of claims 1-2, characterized in that, include: An unbiased target building unit is used to receive intermediate features uploaded by multiple clients, through... Hierarchical sampling constructs a class-balanced query set as the global unbiased optimization objective; The gradient alignment score calculation unit is used to calculate the gradient alignment score between each client gradient and the query set gradient using a lightweight meta-network; Importance weight generation unit, used to generate importance weights for the client based on the gradient alignment score combined with dynamic temperature parameters; The weighted backpropagation unit is used to weight and scale the gradients of each client according to the importance weights during backpropagation, so as to filter conflicting gradients, aggregate valid gradients, update the global model and broadcast it to each client.

4. A parallel segmentation learning device based on meta-network gradient alignment, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that can be executed by the processor to implement a parallel segmentation learning method based on meta-network gradient alignment as described in any one of claims 1-2.

Images