CN118313445A - A federated incremental learning method and system based on constrained gradient updating - Google Patents

Abstract

The invention relates to the technical field of federal learning, and discloses a federal incremental learning method and a federal incremental learning system based on constrained gradient update, wherein the method comprises the following steps: receiving a personalized global model of the last task and cross-task collaboration loss sent by a server; when new task data is acquired, updating model parameters of the previous task along the direction orthogonal to the gradient space of the previous task based on the basis vector of the gradient space of the previous task, the prediction label and the real label of the new task and cross-task cooperation loss to obtain model parameters of the new task in the round of training; performing weighted average on the class prototypes of the new tasks to obtain class average prototypes of the new tasks in the training of the round; and repeating the updating of the model parameters and the class average prototypes, performing preset round training, and after each round of training is completed, sending the model parameters and the class average prototypes of the new tasks to the server. The invention avoids the problems of disastrous forgetting and model drifting when the client has a new task.

Description

A federated incremental learning method and system based on constrained gradient updating

Technical Field

The present invention relates to the field of federated learning technology, and in particular to a federated incremental learning method and system based on constrained gradient updating.

Background technique

Federated learning has developed into a well-known distributed learning paradigm. The classic federated learning scenario assumes that the client trains and tests the model on statically distributed local data, that is, the tasks and quantity of each client's local data are pre-set and remain unchanged. In fact, the task distribution of the client's local data will change dynamically over time, and each client will continuously generate new task data, so the global model needs to continuously update knowledge to adapt to new tasks. This is federated incremental learning.

In federated learning, clients are lightweight mobile devices with limited storage capacity. If new data is continuously saved, it will lead to a memory crisis. However, if the client only saves the newly appeared data and trains the model only on the data of the new task, the global model will not be able to distinguish between the new task and the earlier task well, resulting in catastrophic forgetting of the earlier task. Therefore, the catastrophic forgetting of the earlier tasks in federated incremental learning needs to be solved urgently.

At present, the distillation learning method is usually used to retain the knowledge of historical tasks by constraining the soft labels of historical tasks to be close to the output of the current task model. The problem with the above technical solution is that when the number of tasks on the client is large or the task offset is large, the soft labels of historical tasks will be incorrectly degraded, which will increase the noise in the current task learning process, poor distillation effect, and cause catastrophic forgetting of early tasks.

Summary of the invention

In view of this, the present invention provides a federated incremental learning method and system based on constrained gradient updating to solve the problem of catastrophic forgetting of early tasks caused by the client continuously generating data for new tasks.

In a first aspect, the present invention provides a federated incremental learning method based on constrained gradient updating, which is applied to a client, and the method comprises:

Receive the personalized global model and cross-task collaborative loss of the previous task sent by the server. The personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by all clients based on their local data training of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks indicate all tasks whose training sequence precedes the previous task.

When the local data of the new task is obtained, based on the basis vectors of the gradient space of the previous task, the predicted label and true label of the new task, and the cross-task collaborative loss, the model parameters of the previous task are updated along the direction orthogonal to its gradient space to obtain the model parameters of the new task in this round of training;

Take a weighted average of the class prototypes of the new task to obtain the class average prototype of the new task in this round of training;

Repeat the above updating of model parameters and class average prototypes, perform preset rounds of training, and after each round of training is completed, send the model parameters and class average prototypes of the new task to the server.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient update. Since the personalized global model obtained by the previous task training can classify the previous tasks of all clients and the historical tasks of the previous tasks, and the cross-task collaborative loss obtained by the previous task can improve the performance of the personalized global model, when the local data of the new task is obtained, the personalized global model and cross-task collaborative loss of the previous task sent by the server can be used to limit the gradient update direction to the direction orthogonal to the gradient space of the previous task from the perspective of the gradient space, and the model parameters are updated to protect the previously learned knowledge. Then, the client sends the model parameters and class average prototype obtained in each round of training to the server, and uses the knowledge learned by the new task for the next round of training. By receiving and updating the personalized global model and cross-task collaborative loss of the previous task, and repeatedly sending the updated results of each round of training, when the client continuously generates data for new tasks, catastrophic forgetting of early tasks can be avoided, and the model drift problem between clients can be avoided.

In an optional implementation, before updating the model parameters of the previous task along a direction orthogonal to its gradient space to obtain the model parameters of the new task in the current round of training, the method further includes:

Convert local data of new tasks into high-dimensional vectors;

Based on the model parameters of the previous task and the high-dimensional vector of the new task, the predicted label of the new task is determined.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient updating. By determining the predicted label of the new task, it is possible to prepare for updating the model parameters of the new task based on the difference between the real label and the predicted label of the new task.

In an optional embodiment, the method further includes:

After the new task is trained for a preset number of rounds, the representation matrix of the new task is obtained based on the gradient of the new task obtained in the last round of training and the historical task basis vectors. The historical task basis vectors contain the basis vectors of all historical tasks of the new task.

Perform singular value decomposition on the representation matrix to obtain the basis vectors of the new task and form the orthogonal gradient space of the new task;

The basis vector of the new task is concatenated with the basis vector of the historical tasks to obtain the basis vector containing all current tasks. The basis vector containing all current tasks forms a new orthogonal gradient space for training the next task.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient updating. Since the historical task basis vector is obtained by concatenating the basis vector of the previous task and the basis vectors of all historical tasks of the previous task, the basis vector containing all current tasks is obtained based on the historical task basis vector and the basis vector of the new task. This can learn the knowledge of new tasks without interfering with the knowledge of early tasks, thereby avoiding catastrophic forgetting of early tasks.

In an optional implementation, before receiving the personalized global model of the previous task and the cross-task collaborative loss sent by the server, the method further includes:

For each client, receive the initial model sent by the server;

Based on the local data and initial model of the initial task of the client, the first round of training of the initial task is performed to obtain the initial model parameters and initial prediction labels;

Based on the initial task of the client and the class prototype of the initial task, the initial class average prototype of the initial task of the client in the first round of training is obtained;

Determine the initial gradient corresponding to the initial task based on the predicted label of the initial model, the true label of the initial model, and the initial model parameters;

Based on the initial gradient, the orthogonal initial basis vector of the gradient space of the initial task is obtained;

The initial model parameters and initial class average prototypes obtained in the first round of training for the initial task are sent to the server.

An embodiment of the present invention provides a federated class incremental learning method based on constrained gradient updating, which receives and trains an initial model sent by a server, and sends the trained initial model parameters and initial class average prototype to the server, so that when a new task appears, it can be updated based on the initial model, thereby avoiding catastrophic forgetting of early tasks.

In an optional embodiment, the method further includes:

After each round of training for a new task, the optimization target is updated based on the cross-task collaborative loss of the previous round of training for the new task, the predicted label and the true label of the new task. The optimization target is used to indicate the expected value of the training loss obtained by the client during the training process;

Based on the updated optimization objective, the next round of training for the new task is carried out.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient update, which integrates the knowledge beneficial to the previous task into the training process of the new task by designing a cross-task collaborative loss in the optimization objective, so as to alleviate the model drift problem between clients.

In a second aspect, the present invention provides a federated incremental learning method based on constrained gradient updating, which is applied to a server, and the method comprises:

Send the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by all clients based on their local data training of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks refer to all tasks that are trained before the previous task.

After obtaining new task data, the receiving client updates the model parameters and class average prototype in each round of training, and sends the updated model parameters and class average prototype after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated in a direction orthogonal to its gradient space. The class average prototype is obtained based on the weighted average of the class prototypes of the new task.

An embodiment of the present invention provides a federated class incremental learning method based on constrained gradient updating. By sending a personalized global model and cross-task collaborative loss of a previous task to a client, the client can perform training based on the personalized global model and cross-task collaborative loss of the previous task under a new task, and after each round of training of the client, the client receives the updated model parameters and class average prototype to obtain the personalized global model and cross-task collaborative loss of the new task, so that when the client continuously updates the task, it can learn the new task on the basis of preserving and not interfering with the early task, thereby avoiding catastrophic forgetting of the early task.

In an optional implementation, before sending the personalized global model of the previous task and the cross-task collaborative loss to each client, the method further includes:

For each client, after the client performs a preset round of training on the previous task, receiving the model parameters and the class average prototype of the previous task from the client, where the model parameters and the class average prototype of the previous task are obtained through the last round of training on the previous task;

Determine a plurality of first distances between the client and other clients, the first distance representing the similarity between the class average prototype of the previous task of the client and the class average prototype of the previous task of other clients;

Based on the multiple first distances and the model parameters of the previous task, personalized aggregation is performed to obtain a personalized global model of the previous task of the client;

From each historical task, determine a plurality of cross-task clients whose second distance to the client is greater than a distance threshold, where the second distance represents the degree of similarity between the class average prototype of the client's previous task and the target class average prototype of any cross-task client;

A cross-task collaboration loss of the previous task is determined based on target model parameters of the multiple cross-task clients, second distances corresponding to the multiple cross-task clients, and a model parameter of the previous task.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient updating. Through cross-task collaborative loss, it can improve the performance of a personalized global model based on knowledge that is beneficial to the personalized global model in historical tasks of other clients, thereby alleviating the problem of model drift between clients.

In a third aspect, the present invention provides a federated incremental learning system based on constrained gradient updating, the system comprising a client and a server, and performing the following steps:

The server sends the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks refer to all tasks that were trained before the previous task.

The client receives the personalized global model and cross-task collaborative loss of the previous task sent by the server;

When the local data of a new task is obtained, the client updates the model parameters of the previous task in a direction orthogonal to its gradient space based on the basis vectors of the gradient space of the previous task, the predicted label and true label of the new task, and the cross-task collaborative loss, and obtains the model parameters of the new task in this round of training;

The client performs weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training;

The client repeats the above model parameter and class average prototype update, performs preset rounds of training, and sends the new task model parameters and class average prototype to the server after each round of training.

After receiving new task data, the server updates the model parameters and class average prototypes in each round of training, and sends the updated model parameters and class average prototypes after completing each round of training.

In a fourth aspect, the present invention provides a federated incremental learning device based on constrained gradient updating, which is applied to a client, and the device includes:

The first receiving module is used to receive the personalized global model and cross-task collaborative loss of the previous task sent by the server. The personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss represents the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks represent all tasks whose training sequence is prior to the previous task.

The update module is used to update the model parameters of the previous task in a direction orthogonal to its gradient space based on the basis vectors of the gradient space of the previous task, the predicted label and true label of the new task, and the cross-task collaborative loss when the local data of the new task is obtained, so as to obtain the model parameters of the new task in this round of training;

The acquisition module is used to perform weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training;

The first training module is used to repeat the updating of the above model parameters and class average prototypes, perform preset rounds of training, and send the model parameters and class average prototypes of new tasks to the server after each round of training is completed.

In a fifth aspect, the present invention provides a federated incremental learning device based on constrained gradient updating, which is applied to a server, and the device includes:

A sending module is used to send the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss represents the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks represent all tasks whose training sequence is prior to the previous task.

The first receiving module is used to receive the model parameters and class average prototypes updated by the client in each round of training after acquiring new task data, and send the updated model parameters and class average prototypes after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated in a direction orthogonal to its gradient space, and the class average prototype is obtained based on the weighted average of the class prototypes of the new task.

In a sixth aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to execute a federated incremental learning method based on constrained gradient update according to the first aspect or any corresponding embodiment thereof.

In the seventh aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute a federated incremental learning method based on constrained gradient updating according to the first aspect or any corresponding embodiment thereof.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a federated incremental learning method based on constrained gradient updating applied to a client according to an embodiment of the present invention;

FIG2 is a schematic diagram of a constrained gradient update according to an embodiment of the present invention;

3 is a flow chart of a federated incremental learning method based on constrained gradient updating applied to a server according to an embodiment of the present invention;

4 is a flowchart of the interaction between a client and a server in a federated incremental learning system based on constrained gradient updating according to an embodiment of the present invention;

FIG5 is a flow chart of a federated incremental learning method based on constrained gradient updating according to an embodiment of the present invention;

6 is a structural block diagram of a federated incremental learning device based on constrained gradient updating applied to a client according to an embodiment of the present invention;

7 is a structural block diagram of a federated incremental learning device based on constrained gradient updating applied to a server according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient updating, which is applied to clients and servers. It solves the problem in related technologies that federated incremental learning through distillation learning causes catastrophic forgetting of early tasks. By limiting the gradient update direction and increasing the cross-task collaboration loss between clients, it is possible to learn knowledge of new tasks without interfering with the knowledge of early tasks, thereby avoiding catastrophic forgetting of early tasks and avoiding the problem of model drift caused by differences between clients.

According to an embodiment of the present invention, an embodiment of a federated incremental learning method based on constrained gradient updating is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

In this embodiment, a federated incremental learning method based on constrained gradient updating is provided, which is applied to a client. FIG1 is a flow chart of a federated incremental learning method based on constrained gradient updating applied to a client according to an embodiment of the present invention. As shown in FIG1 , the process includes the following steps:

Step S101, receiving the personalized global model and cross-task collaborative loss of the previous task sent by the server, the personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task, the cross-task collaborative loss represents the similarity between the model parameters of the previous task of all clients and the model parameters of the historical task, and the historical task represents all tasks whose training sequence precedes the previous task. Specifically, the personalized global model is a neural network model, such as LeNet-5, AlexNet (deep convolutional neural network), ResNet (residual neural network), MobileNet or MLP (Multilayer Perceptron), etc., and the embodiment of the present invention does not limit this. Taking ALexNet as an example, its structure includes an input layer, five convolutional layers, each convolutional layer is followed by a maximum pooling layer and a local response normalization layer, the convolutional layer is followed by three fully connected layers, the convolutional layer and the fully connected layer use ReLU activation function, and finally a Softmax layer is used to output the classification result to achieve the classification task. The present invention is described by taking a lightweight mobile device as a client and a specific application service provider as a server as an example, for example, the server provides various classification tasks for data such as images, videos or texts. As a data collector, the client will continuously generate data for new tasks, but at the same time, it is limited by the limited content and cannot save the data of all tasks. Therefore, the client needs to continuously train the model on the data of the newly added tasks, and at the same time, it is also necessary to ensure that the trained model can show good classification ability on the data of the earlier tasks that have been deleted or inaccessible. If the model is trained on the data of the newly added tasks using the traditional training method, it will easily cause the classification accuracy of the server's global model on the data of the earlier tasks to drop rapidly, resulting in the catastrophic forgetting problem of the earlier tasks. The present invention receives the personalized global model and cross-task collaborative loss of the previous task sent by the server, and updates it on the basis of the personalized global model of the previous task, so that the new tasks can be classified while the earlier tasks can also be accurately classified, avoiding the catastrophic forgetting problem of the earlier tasks. In addition, since the cross-task collaborative loss is obtained based on the similarity between the model parameters of the previous tasks of all clients and the model parameters of the historical tasks, the model drift problem caused by the differences between clients can be avoided.

Step S102, when the local data of the new task is obtained, based on the basis vectors of the gradient space of the previous task, the predicted label and the true label of the new task, and the cross-task collaborative loss, the model parameters of the previous task are updated along the direction orthogonal to its gradient space, and the model parameters of the new task in this round of training are obtained. Specifically, when the client has a new task, that is, the local data of the new task is obtained, the model parameters of the previous task can be updated along the direction orthogonal to its gradient space by the following formula (1) to obtain the gradient of the gradient space of the new task.

Where t represents the previous task; t+1 represents the new task; ▽ represents the gradient; θ ^t+1 represents the model parameters of the new task; k represents the kth client; represents the true label of the kth client in the new task, represents the predicted label of the kth client in the new task; Represents the cross entropy loss between the true label and the predicted label of the new task; U ^t represents the basis vector of the gradient space of the previous task.

Optionally, the cross entropy loss between the true label and the predicted label of the new task can be determined by the following formula (2).

In some optional embodiments, based on the cross-task collaboration loss of the previous task and the cross entropy loss between the true label and the predicted label of the new task, the model parameters in the personalized global model of the previous task can be updated in a direction orthogonal to its gradient space through the error and back propagation algorithm to obtain the model parameters of the new task in this round of training, so that the updated model parameters can learn the knowledge of the new task.

In some optional implementations, FIG. 2 is a schematic diagram of a constrained gradient update according to an embodiment of the present invention. As shown in FIG. 2 , represents the model parameters of the previous task, Represents the model parameters of the new task. After the client completes training of each task, it can be considered that the gradient of the personalized global model obtained through training retains the knowledge of all historical tasks that have been learned, and the gradients of the current task and the historical tasks constitute a gradient space. When the client obtains the local data of a new task, the gradient needs to be updated. Since the gradient space of the previous task is composed of the gradient of the previous task and the gradients of all historical tasks of the previous task, if the gradient update direction of the new task is not in the orthogonal direction of the gradient space of the previous task, it will cause disturbance to the knowledge of the previous task, that is, the updated gradient space forgets the knowledge of the earlier tasks. Therefore, by constraining the update direction of the gradient to the orthogonal direction of the gradient space of the previous task, the disturbance to the knowledge of the historical tasks is minimized, so that the model parameters are updated according to the updated gradient to prevent the knowledge of the historical tasks from being forgotten and alleviate the problem of early forgetting.

Step S103, weighted average the class prototype of the new task to obtain the class average prototype of the new task in this round of training. Specifically, the class prototype is the intermediate result of the local data of the new task output by the feature extractor, that is, the result obtained after the local data of the new task is calculated by the last fully connected layer. By obtaining the class average prototype, preparation is made for the subsequent determination of the personalized global model of the new task.

Optionally, the class average prototype of the new task in this round of training can be determined by the following formula (3).

Among them, t represents the previous task; t+1 represents the new task; k represents the kth client, represents the average class prototype of the kth client in the new task; N represents the number of local data of the new task; e represents the class prototype of the new task in this round of training; i represents the local data of the new task.

Step S104, repeat the update of the above model parameters and class average prototype, perform preset rounds of training, and after each round of training is completed, send the model parameters and class average prototype of the new task to the server. Specifically, each client has multiple tasks, and each task needs to be trained for a preset round. Steps S101 to S104 represent the execution process of any round of training for any task except the initial task. In this round of training, the personalized global model and cross-task collaborative loss obtained in the previous round of training are received, and then the current round of training is performed based on the personalized global model and cross-task collaborative loss obtained in the previous round of training, and the model parameters and class average prototype of this round of training are obtained and sent to the server. Among them, the present invention does not limit the training rounds of any task. By performing preset rounds of training on the new task and sending the training results to the server, the learning effect of the local data of the new task is guaranteed.

An embodiment of the present invention provides a federated incremental learning method based on constrained gradient update. Since the personalized global model obtained by the previous task training can classify the previous tasks of all clients and the historical tasks of the previous tasks, and the cross-task collaborative loss obtained by the previous task can improve the performance of the personalized global model, when the local data of the new task is obtained, the personalized global model and cross-task collaborative loss of the previous task sent by the server can be used to limit the gradient update direction to the direction orthogonal to the gradient space of the previous task from the perspective of the gradient space, and the model parameters are updated to protect the previously learned knowledge. Then, the client sends the model parameters and class average prototype obtained in each round of training to the server, and uses the knowledge learned by the new task for the next round of training. By receiving and updating the personalized global model and cross-task collaborative loss of the previous task, and repeatedly sending the updated results of each round of training, when the client continuously generates data for new tasks, catastrophic forgetting of early tasks can be avoided, and the model drift problem between clients can be avoided.

In this embodiment, a federated incremental learning method based on constrained gradient updating is provided, which is applied to a server. FIG3 is a flow chart of a federated incremental learning method based on constrained gradient updating applied to a server according to an embodiment of the present invention. As shown in FIG3 , the process includes the following steps:

Step S301, sending the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks indicate all tasks whose training sequence precedes the previous task. Specifically, all clients will execute the previous task. Therefore, in each round of training of the previous task, each client can receive the personalized global model corresponding to the previous round of training of the training, so that the next round of training can be performed based on the personalized global model obtained in the previous round of training. Since the server obtains the personalized global model of the previous task based on the multiple model parameters and multiple class average prototypes obtained by all clients in the previous task. Therefore, the personalized global model has the classification capability for the previous task of all clients. Since the server obtains the cross-task collaborative loss based on the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks, the cross-task collaborative loss can reduce the differences between clients to avoid the model drift problem. Optionally, for the first round of training of each task, the server may send to the client the personalized global model and cross-task collaborative loss obtained in the last round of training of the previous task of the task. For the second round of training of each task until the end of the preset round of training, the server may send to the client the personalized global model obtained in the previous round of training of the current round, or may send to the client the model parameters obtained in the previous round of training of the current round, which is not limited in the embodiment of the present invention.

Step S302, after receiving the new task data, the client updates the model parameters and the class average prototype in each round of training, and sends the updated model parameters and class average prototype after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated along the gradient direction orthogonal to its gradient space, and the class average prototype is obtained based on the weighted average of the class prototype of the new task. Specifically, since there are multiple clients, and each client has multiple tasks. For each task, a preset round of training is required. The client will send the model parameters and class average model obtained in each round of training. The server will receive all model parameters and all class average prototypes obtained by all clients in this round of training for this task, and prepare for the next round of training.

In some optional implementations, FIG5 is a flow chart of a federated incremental learning method based on constrained gradient updating according to an embodiment of the present invention. As shown in FIG5 , any round of training of any task except the initial task is taken as an example for explanation. Represents the local data of the first client in this round of training for this task. Indicates the model parameters of the first client in this round of training for this task. represents the average prototype of the class of the first client in this round of training for this task, D represents an optional distance metric function, p ^t represents the cross-task collaboration loss, represents the personalized global model of this round of training for this task, and k represents the kth client. First, all clients perform local training on this task, obtain the model parameters and class average prototype of this round of training, and send them to the server. Then, the server calculates the distance between the historical task and the current task through the class average prototype to determine the cross-task collaboration loss. Then, the server performs personalized aggregation on the model parameters of this round of training to obtain the personalized global model of this round of training. Then, the server sends the cross-task collaboration loss and personalized global model obtained in this round of training to the client for the next round of training.

An embodiment of the present invention provides a federated class incremental learning method based on constrained gradient updating. By sending a personalized global model and cross-task collaborative loss of a previous task to a client, the client can perform training based on the personalized global model and cross-task collaborative loss of the previous task under a new task, and after each round of training of the client, the client receives the updated model parameters and class average prototype to obtain the personalized global model and cross-task collaborative loss of the new task, so that when the client continuously updates the task, it can learn the new task on the basis of saving the early task, thereby avoiding catastrophic forgetting of the early task.

In this embodiment, a federated class incremental learning system based on constrained gradient update is provided, and the system includes a client and a server. FIG4 is an interaction flow chart of a client and a server in a federated class incremental learning system based on constrained gradient update according to an embodiment of the present invention. As shown in FIG4, the server sends a personalized global model and cross-task collaborative loss of the previous task to each client. Then, the client receives the personalized global model and cross-task collaborative loss of the previous task, and performs training on the newly added task when there is a new task. Based on the basis vectors of the gradient space of the previous task, the predicted label and the true label of the new task, and the cross-task collaborative loss, the client updates the model parameters of the previous task along the gradient direction orthogonal to its gradient space, and obtains the model parameters of the new task in this round of training. Then, the client performs weighted averaging on the class prototype of the new task to obtain the class average prototype of the new task in this round of training. Then, the client repeats the update of the above model parameters and class average prototype, and sends the updated model parameters and class average prototype to the server after each round of training. Then, the server receives the updated model parameters and class average prototype of the client after each round of training.

In some optional implementations, the federated incremental learning system based on constrained gradient update includes a client and a server, and the detailed process of the system includes the following steps:

Step S401: For each client, receive the initial model sent by the server. Specifically, the initial model is a model parameter randomly initialized and can be represented by θ ^init .

Step S402: The client performs the first round of training for the initial task based on the local data and initial model of the initial task of the client to obtain the initial model parameters and initial prediction labels. Specifically, all clients initialize their local data. Assuming that each client has T tasks and they appear in sequence, the tth task can be expressed as in, represents the local data set of the t-th task; k represents the k-th client; i represents the ith local data; t represents the t-th task; N represents the number of local data for each task; T represents the total number of tasks for the client. The local data of the initial task can be expressed as Then, the first round of training for the initial task proceeds as follows: First, the local data of the initial task is read as a high-dimensional vector consisting of floating-point numbers By Data Calculate the model parameters θ ^init to get the predicted label Then, the true label can be calculated by the above formula (2) and predicted labels Then, the model parameters θ ^init are modified and updated through the error and back propagation algorithm to obtain the updated initial model parameters

Step S403, based on the initial task of the client and the class prototype of the initial task, obtain the initial class average prototype of the initial task of the client in the first round of training. Specifically, referring to the above step S103, the initial class average prototype can be obtained by the above formula (3).

Step S404, based on the predicted label of the initial model, the true label of the initial model and the initial model parameters, determine the initial gradient corresponding to the initial task. Specifically, the initial gradient can be obtained by the following formula (4).

Among them, ▽ represents the gradient; θ ¹ represents the model parameters of the first round of training for the initial task; represents the cross entropy loss between the true label and the predicted label of the initial task in the first round of training, k represents the kth client, represents the true label of the kth client in the first round of training of the initial task; A high-dimensional vector representing the local data of the k-th client in the initial task.

In step S405, the client obtains the orthogonal initial basis vectors of the gradient space of the initial task based on the initial gradient. Specifically, the initial gradient obtained by the initial task in the first round of training can be represented by the representation matrix shown in equation (5). Then, the representation matrix is subjected to singular value decomposition to obtain the following equation (6). Then, the left singular matrix is defined as the orthogonal initial basis vectors of the gradient space of the initial task in the first round of training as shown in equation (7). Optionally, other matrix decomposition methods can also be used to obtain basis vectors, which is not limited in the embodiments of the present application.

Among them, Z represents the representation matrix of the first training of the initial task; L represents L vectors.

Z ¹ ＝U ¹ S ¹ (V ¹ ) ^T (6)

Among them, Z represents the representation matrix of the first training of the initial task; U represents the left singular matrix of the first training of the initial task; S represents the singular values of the first training of the initial task; V represents the right singular matrix of the first training of the initial task.

Among them, U represents the orthogonal initial basis vector of the gradient space of the initial task in the first round of training; span() represents the expansion space function; k represents the kth client.

Step S406, the client sends the initial model parameters and the initial class average prototype obtained in the first round of training of the initial task to the server. Specifically, the initial model parameters and the initial class average prototype obtained in the first round of training of the initial task are sent to the server, so that the second round of training of the initial task can be updated based on the knowledge learned in the first round of training of the initial task.

It should be noted that the above steps S401 to S406 are the implementation process of the first round of training of the initial task of all clients. After each client executes the above steps S401 to S406, the server receives the initial model parameters and the initial class average prototype obtained in the first round of training of the initial task, and determines the personalized global model of the first round of training of the initial task. Then, the server sends the personalized global model obtained in the first round of training to the client for the second round of training of the initial task. Then, starting from the second round of training of the initial task until the last round of training of the initial task is completed, the client refers to the above steps S401 to S406, receives the personalized global model obtained in the previous round of training on the server for this round of training, and sends the model parameters and class average prototype obtained in each round of training to the server after each round of training. The server also determines the personalized global model of the current round after each round of training and sends it to the client for the next round of training. Among them, the training rounds of the initial task can be 5 times, 10 times or 20 times, etc., which are not limited in the embodiment of the present invention. After completing the preset rounds of training for the initial task, if the client has a new task, the client and the server no longer perform the above steps S401 to S406 for the new task, but perform the following steps to achieve training and learning for the new task.

In some optional implementations, after the client completes the preset round of training of the previous task and a new task appears, the federated incremental learning method based on constrained gradient update provided by the embodiment of the present application is described by taking the following steps S407 to S424 as an example. Among them, the server processes the training results of the previous task that has completed training through the following steps S407 to S411, and the client and the server start training for the new task from step S412.

Step S407, for each client, after the client performs a preset round of training on the previous task, the server receives the model parameters and class average prototype of the previous task of the client, and the model parameters and class average prototype of the previous task are obtained by the last round of training of the previous task. Specifically, the previous task can be any task except the initial task. Since each round of training of each task will obtain the model parameters and class average prototype corresponding to this round of training. Therefore, after the client performs a preset round of training on the previous task, the model parameters and class average prototype received by the server are obtained by the last round of training of the previous task. By receiving the model parameters and class average prototype of the previous task, the server can determine the personalized global model that has learned the knowledge of the previous task and prepare for the training of the next task.

Step S408, the server determines a plurality of first distances between the client and other clients, wherein the first distance represents the similarity between the class average prototype of the previous task of the client and the class average prototype of the previous task of other clients. Specifically, the server can determine the first distance using the following formula (8).

Where t represents the previous task; k represents the kth client; j is the index of other clients; represents the first distance; D represents an optional distance metric function; p represents the class average prototype; K represents all clients. Optionally, other distance metric functions in Euclidean space can also be used to determine the distance between class average prototypes, which is not limited in the present embodiment.

In step S409, the server performs personalized aggregation based on the multiple first distances and the model parameters of the previous task to obtain a personalized global model of the previous task of the client. Specifically, the server can perform personalized aggregation through the following formula (9) to obtain a personalized global model of the previous task. By performing personalized aggregation based on the first distance, the target model parameters and the target class average prototype, the obtained personalized global model can not only classify the tasks of the client itself, but also has the ability to classify the tasks of other clients, thereby improving the classification ability of the model.

Among them, t represents the previous task; k represents the kth client; Represents a personalized global model of the previous task; represents the model parameters of the previous task; j is the index of other clients; K represents all clients; represents the first distance;

In step S410, the server determines from each historical task a plurality of cross-task clients whose second distance to the client is greater than a distance threshold, wherein the second distance represents the degree of similarity between the class average prototype of the client's previous task and the target class average prototype of any cross-task client. Specifically, the server can use the historical prototype selection mechanism to find the historical task that has made the greatest improvement on the previous task, that is, determine the cross-task client by the second distance obtained from each historical task based on the above formula (8) to improve the performance of the personalized global model of the previous task on the historical task. For example, the server selects two other clients with the smallest distance for client k in each historical task. When the client executes the tth task, the number of selected other client historical tasks is α=2*(t-1). Optionally, the cross-task client can also be determined in other ways, which is not limited in the embodiments of the present invention.

In step S411, the server determines the cross-task collaboration loss of the previous task based on the target model parameters of the multiple cross-task clients, the second distances corresponding to the multiple cross-task clients, and the model parameters of the previous task. Specifically, since the cross-task client has a high degree of similarity with the client, the cross-task collaboration loss is determined based on the target model parameters of the cross-task client by the following formula (10), which can reduce the model drift problem caused by the differences between the clients.

Among them, t represents the previous task; k represents the kth client; represents the model parameters of the previous task; θ ^old represents the target model parameters of the cross-task client determined from the historical tasks; represents the second distance corresponding to the cross-task client; α represents the number of historical tasks; i represents the i-th historical task.

In step S412, the server sends the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks indicate all tasks whose training sequence is prior to the previous task. For details, please refer to step S301 of the embodiment shown in FIG3 , which will not be described in detail here.

In step S413, the client receives the personalized global model and cross-task collaborative loss of the previous task sent by the server. The personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks indicate all tasks whose training sequence is prior to the previous task. For details, please refer to step S101 of the embodiment shown in FIG1 , which will not be described in detail here.

Step S414: The client converts the local data of the new task into a high-dimensional vector. Specifically, the high-dimensional vector is composed of floating-point numbers. By converting the local data of the new task into a high-dimensional vector, preparation is made for determining the prediction label of the new task.

In step S415, the client determines the predicted label of the new task based on the model parameters of the previous task and the high-dimensional vector of the new task. Specifically, the predicted label of the new task is determined with reference to the above step S602.

Step S416: Based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss, the client updates the model parameters of the previous task along the gradient direction orthogonal to its gradient space to obtain the model parameters of the new task in this round of training. For details, please refer to step S102 of the embodiment shown in FIG1 , which will not be described in detail here.

Step S417: The client performs weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training. For details, please refer to step S103 of the embodiment shown in FIG1 , which will not be described in detail here.

Step S418, after each round of training of the new task is completed, the optimization target is updated based on the cross-task collaborative loss of the previous round of training of the new task, the predicted label and the true label of the new task, and the optimization target is used to indicate the expected value of the training loss obtained by the client during the training process. Specifically, the optimization target can indicate the minimum loss value that the client expects to achieve for task training. The optimization target can be determined by the following formula (11).

Among them, t+1 represents a new task; k represents the kth client; Represent a personalized global model for new tasks; represents the cross entropy loss based on the predicted label and the true label of the new task, which can be calculated by the above formula (2); It represents the cross-task collaboration loss obtained by the previous round of training of the new task obtained by the above formula (10) with reference to the above step S411.

Step S419, the client performs the next round of training for the new task based on the updated optimization target. Specifically, since the optimization target can reflect the minimum value expected to be achieved by the training loss, indicating the next round of training based on the optimization target can improve the training effect of the next round of training.

In step S420, the client repeats the updating of the above model parameters and the class average prototype, performs a preset round of training, and sends the model parameters and the class average prototype of the new task to the server after each round of training. For details, please refer to step S104 of the embodiment shown in FIG1, which will not be repeated here.

Step S421, after the new task is trained for a preset round, the client obtains the representation matrix of the new task based on the gradient of the new task obtained in the last round of training and the historical task basis vector, and the historical task basis vector contains the basis vectors of all historical tasks of the new task. Specifically, the historical task basis vector is obtained by splicing the basis vectors of all historical tasks of the new task. Therefore, based on the gradient and historical task basis vector obtained by the client in the last round of training of the new task, the representation matrix of the new task can be obtained by the following formula (12), so that the knowledge of the historical tasks can be not interfered with on the basis of learning the new task, avoiding catastrophic forgetting of the early tasks.

Wherein, t represents the previous task; t+1 represents the new task; U ^t represents the historical task basis vector; Z ^t+1 represents the representation matrix obtained based on the gradient obtained in the last round of training of the new task with reference to the above step S405; The representation matrix of the historical tasks that represent the new task.

In step S422, the client performs singular value decomposition on the representation matrix to obtain the basis vector of the new task and form the orthogonal gradient space of the new task. Specifically, the representation matrix can be subjected to singular value decomposition as shown in the following formula (13), and then the basis vector of the new task is obtained by the following formula (14).

Z ^t+1 =U ^t+1 S ^t+1 (V ^t+1 ) ^T (13)

Among them, t+1 represents the new task; Z ^t+1 represents the representation matrix of the new task in this round of training; U ^t+1 represents the left singular matrix of the new task in this round of training; S ^t+1 represents the singular value of the new task in this round of training; V ^t+1 represents the right singular matrix of the new task in this round of training;

Among them, t+1 represents the new task; U ^t+1 represents the left singular matrix of the new task in this round of training, that is, the basis vector of the new task.

In step S423, the client concatenates the basis vector of the new task with the basis vector of the historical tasks to obtain the basis vector containing all current tasks, and forms a new orthogonal gradient space by the basis vector containing all current tasks for training the next task. Specifically, the basis vector containing all current tasks is shown in the following formula (15). By concatenating the basis vector of the new task with the basis vector of the historical tasks, the knowledge of the new task is concatenated with the knowledge of the historical tasks to obtain the basis vector containing all current tasks. When adding a new task, training can be performed based on the basis vector containing all current tasks, so that the knowledge of new and old tasks can be stored at the same time, avoiding the catastrophic forgetting problem of early tasks.

Among them, t represents the previous task; t+1 represents the new task; U ^t represents the basis vector of the previous task; U ^t+1 represents the basis vector of the new task; U ^t+1 represents the left singular matrix of the new task in this round of training; k represents the kth client.

Step S424, the server receives the updated model parameters and class average prototypes sent by the client after acquiring new task data in each round of training, and after completing each round of training, the model parameters are updated based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss, and the model parameters of the previous task are updated along the gradient direction orthogonal to its gradient space, and the class average prototype is obtained based on the weighted average of the class prototypes of the new task. For details, please refer to step S302 of the embodiment shown in Figure 3, which will not be repeated here.

It should be noted that for the first round of training of the initial task, the above steps S401 to S406 are executed. For the second round of training of the initial task until the last round of training, the client refers to the above steps S401 to S406, receives the personalized global model obtained by the previous round of training from the server for this round of training, and sends the model parameters and class average prototype obtained by each round of training to the server after each round of training. The server sends the personalized global model obtained by the previous round of training to the client after each round of training of the initial task. For any task other than the initial task, the server determines the personalized global model and cross-task collaboration loss of the previous task of the task through the above steps S407 to S411. Then, the server executes step S412, and the client executes steps S413 to S420 to perform the first round of training for the new task. Then, for the second round of training of the new task until the last round of training of the new task, the server updates the model parameters and class average prototype obtained in the previous round of training of the current round with reference to the above steps S407 to S412, and sends the obtained personalized global model and cross-task collaborative loss to the client. The client receives the personalized global model and cross-task collaborative loss obtained in the previous round of training of the current round, and repeats the above steps S413 to S420. The server repeats the above step S424 to update the model parameters and class average prototype obtained in the current round of training. When the client performs the last round of training of the new task, it is also necessary to additionally perform steps S421 to S423.

The embodiment of the present invention provides a federated incremental learning method based on constrained gradient update. Since the personalized global model obtained by the previous task training can classify the previous tasks of all clients and the historical tasks of the previous task, the cross-task collaborative loss obtained by the previous task can improve the performance of the personalized global model. Therefore, when the local data of the new task is obtained, the gradient update direction can be limited to the gradient space of the previous task from the perspective of the gradient space based on the personalized global model and cross-task collaborative loss of the previous task sent by the server, and the model parameters are updated to protect the previously learned knowledge. Then, the client sends the model parameters and class average prototype obtained in each round of training to the server, and uses the knowledge learned by the new task for the next round of training. By receiving and updating the personalized global model and cross-task collaborative loss of the previous task, and repeatedly sending the updated results of each round of training, when the client continuously generates data for new tasks, catastrophic forgetting of early tasks can be avoided, and the model drift problem between clients can be avoided.

In this embodiment, a federated incremental learning device based on constrained gradient update is also provided, which is used to implement the above embodiments and preferred implementation modes, and will not be repeated hereafter. As used below, the term "module" can implement a combination of software and/or hardware for a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

This embodiment provides a federated incremental learning device based on constrained gradient updating, which is applied to a client, as shown in FIG6 , and includes:

The first receiving module 601 is used to receive the personalized global model of the previous task and the cross-task collaborative loss sent by the server. The personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss represents the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks represent all tasks whose training sequence is prior to the previous task.

Update module 602 is used to update the model parameters of the previous task in a direction orthogonal to its gradient space based on the basis vectors of the gradient space of the previous task, the predicted label and true label of the new task, and the cross-task collaborative loss when new task data is acquired, so as to obtain the model parameters of the new task in this round of training.

The acquisition module 603 is used to perform weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training.

The first training module 604 is used to repeat the updating of the above model parameters and class average prototypes, perform preset rounds of training, and send the model parameters and class average prototypes of new tasks to the server after each round of training is completed.

In some optional implementations, before updating module 602, the device further includes:

The conversion module is used to convert the local data of the new task into a high-dimensional vector.

The first determination module is used to determine the prediction label of the new task based on the model parameters of the previous task and the high-dimensional vector of the new task.

In some optional embodiments, the device further comprises:

The second determination module is used to obtain the representation matrix of the new task after the new task has been trained for a preset round of times, based on the gradient of the new task obtained in the last round of training and the historical task basis vectors, where the historical task basis vectors contain the basis vectors of all historical tasks of the new task.

The decomposition module is used to perform singular value decomposition on the representation matrix to obtain the basis vectors of the new task and form the orthogonal gradient space of the new task.

The splicing module is used to splice the basis vector of the new task with the basis vector of the historical tasks to obtain the basis vector containing all current tasks for training the next task.

In some optional implementations, before the first receiving module 601, the device further includes:

The second receiving module is used to receive the initial model sent by the server for each client.

The second training module is used to perform the first round of training for the initial task based on the local data and initial model of the initial task of the client, and obtain the initial model parameters and initial prediction labels.

The third determination module is used to obtain an initial class average prototype of the initial task of the client in the first round of training based on the initial task of the client and the class prototype of the initial task.

The fourth determination module is used to determine the initial gradient corresponding to the initial task based on the predicted label of the initial model, the true label of the initial model and the initial model parameters.

The fifth determination module is used to obtain an orthogonal initial basis vector of the gradient space of the initial task based on the initial gradient.

The sending module is used to send the initial model parameters and initial class average prototype obtained by the initial task in the first round of training to the server.

This embodiment provides a federated incremental learning device based on constrained gradient updating, which is applied to a server, as shown in FIG7 , and includes:

The sending module 701 is used to send the personalized global model and cross-task collaborative loss of the previous task to each client. The personalized global model is obtained by the server personalizing and aggregating multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task. The cross-task collaborative loss represents the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks. The historical tasks represent all tasks whose training sequence is prior to the previous task.

The first receiving module 702 is used to receive the updated model parameters and class average prototypes sent by the client after the client obtains new task data in each round of training, and after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated along a direction orthogonal to its gradient space, and the class average prototype is obtained based on the weighted average of the class prototypes of the new task.

In some optional implementations, before sending module 701, the device further includes:

The second receiving module is used to receive the model parameters and class average prototype of the previous task of each client after the client performs a preset round of training on the previous task, and the model parameters and class average prototype of the previous task are obtained through the last round of training of the previous task.

The third determination module is used to determine a plurality of first distances between the client and other clients, wherein the first distance represents the similarity between the class average prototype of the previous task of the client and the class average prototype of the previous task of other clients.

The aggregation module is used to perform personalized aggregation based on multiple first distances and model parameters of the previous task to obtain a personalized global model of the previous task of the client.

The fourth determination module is used to determine, from each historical task, multiple cross-task clients whose second distance to the client is greater than a distance threshold, where the second distance represents the similarity between the class average prototype of the client's previous task and the target class average prototype of any cross-task client.

The fifth determination module is used to determine the cross-task collaboration loss of the previous task based on the target model parameters of the multiple cross-task clients, the second distances corresponding to the multiple cross-task clients, and the model parameters of the previous task.

The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

In this embodiment, a federated incremental learning device based on constrained gradient update is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and/or other devices that can provide the above functions.

An embodiment of the present invention also provides a computer device having a federated incremental learning device based on constrained gradient updating as shown in FIG. 6 and FIG. 7 above.

Please refer to Figure 8, which is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention. As shown in Figure 8, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. The various components are connected to each other using different buses for communication, and can be installed on a common motherboard or installed in other ways as needed. The processor can process instructions executed in the computer device, including instructions stored in or on the memory to display graphical information of the GUI on an external input/output device (such as a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and/or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). In Figure 8, a processor 10 is taken as an example.

The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required by at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims (10)

1. A federated incremental learning method based on constrained gradient update, characterized in that it is applied to a client and comprises: Receive a personalized global model and a cross-task collaborative loss of a previous task sent by a server, wherein the personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task, and the cross-task collaborative loss represents the degree of similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks, wherein the historical tasks represent all tasks whose training sequence is prior to the previous task; When the local data of the new task is obtained, based on the basis vectors of the gradient space of the previous task, the predicted label and the true label of the new task, and the cross-task collaborative loss, the model parameters of the previous task are updated along a direction orthogonal to its gradient space to obtain the model parameters of the new task in this round of training; Take a weighted average of the class prototypes of the new task to obtain the class average prototype of the new task in this round of training; Repeat the updating of the above model parameters and class average prototypes, perform preset rounds of training, and after each round of training is completed, send the model parameters and class average prototypes of the new task to the server. 2. The method according to claim 1, characterized in that before the model parameters of the previous task are updated along a direction orthogonal to its gradient space to obtain the model parameters of the new task in the current round of training, the method further comprises: Converting local data of the new task into a high-dimensional vector; Based on the model parameters of the previous task and the high-dimensional vector of the new task, a prediction label of the new task is determined. 3. The method according to claim 1, characterized in that the method further comprises: After the new task is trained for a preset number of rounds, a representation matrix of the new task is obtained based on the gradient of the new task obtained in the last round of training and the historical task basis vectors, wherein the historical task basis vectors include basis vectors of all historical tasks of the new task; Performing singular value decomposition on the representation matrix to obtain basis vectors of a new task, thereby forming an orthogonal gradient space of the new task; The basis vector of the new task is concatenated with the basis vector of the historical tasks to obtain the basis vector containing all current tasks, and a new orthogonal gradient space is formed by the basis vector containing all current tasks for training the next task. 4. The method according to claim 1, characterized in that before receiving the personalized global model of the previous task and the cross-task collaborative loss sent by the server, the method further comprises: For each client, receiving the initial model sent by the server; Based on the local data of the initial task of the client and the initial model, perform a first round of training for the initial task to obtain initial model parameters and initial prediction labels; Based on the initial task of the client and the class prototype of the initial task, obtaining an initial class average prototype of the initial task of the client in the first round of training; Determining an initial gradient corresponding to the initial task based on the predicted label of the initial model, the true label of the initial model, and the initial model parameters; Based on the initial gradient, obtaining an orthogonal initial basis vector of the gradient space of the initial task; The initial model parameters and the initial class average prototype obtained in the first round of training for the initial task are sent to the server. 5. The method according to claim 1, characterized in that the method further comprises: After each round of training of the new task is completed, based on the cross-task collaborative loss of the previous round of training of the new task, the predicted label and the true label of the new task, the optimization target is updated, and the optimization target is used to indicate the expected value of the training loss obtained by the client during the training process; Based on the updated optimization objective, the next round of training for the new task is performed. 6. A federated incremental learning method based on constrained gradient update, characterized in that it is applied to a server and comprises: Sending a personalized global model and a cross-task collaborative loss of a previous task to each client, wherein the personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on their local data of the previous task, and the cross-task collaborative loss indicates the similarity between the model parameters of the previous task of all clients and the model parameters of the historical tasks, wherein the historical tasks indicate all tasks whose training sequence is prior to the previous task; After acquiring new task data, the receiving client updates the model parameters and class average prototype in each round of training, and sends the updated model parameters and class average prototype after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated along a direction orthogonal to its gradient space. The class average prototype is obtained based on the weighted average of the class prototypes of the new task. 7. The method according to claim 6, characterized in that before sending the personalized global model of the previous task and the cross-task collaborative loss to each client, the method further comprises: For each client, after the client performs a preset round of training on the previous task, receiving the model parameters and the class average prototype of the previous task of the client, where the model parameters and the class average prototype of the previous task are obtained by the last round of training of the previous task; Determine a plurality of first distances between the client and other clients, wherein the first distances represent the similarity between the class average prototype of the previous task of the client and the class average prototype of the previous task of other clients; Based on the multiple first distances and the model parameters of the previous task, personalized aggregation is performed to obtain a personalized global model of the previous task of the client; From each historical task, determine a plurality of cross-task clients whose second distance from the client is greater than a distance threshold, wherein the second distance represents a similarity between an average prototype of a class of a previous task of the client and an average prototype of a target class of any cross-task client; A cross-task collaboration loss of a previous task is determined based on the target model parameters of the multiple cross-task clients, the second distances corresponding to the multiple cross-task clients, and the model parameters of the previous task. 8. A federated incremental learning system based on constrained gradient update, characterized in that the system includes a client and a server, and performs the following steps: The server sends a personalized global model of the previous task and a cross-task collaborative loss to each client, wherein the personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on local data of their respective previous tasks, and the cross-task collaborative loss represents the degree of similarity between the model parameters of the previous tasks of all clients and the model parameters of historical tasks, wherein the historical tasks represent all tasks whose training sequence precedes the previous task; The client receives the personalized global model and cross-task collaboration loss of the previous task sent by the server; When the local data of the new task is obtained, the client updates the model parameters of the previous task in a direction orthogonal to its gradient space based on the basis vectors of the gradient space of the previous task, the predicted label and the true label of the new task, and the cross-task collaborative loss, to obtain the model parameters of the new task in this round of training; The client performs weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training; The client repeats the updating of the model parameters and the class average prototype, performs a preset round of training, and sends the model parameters and the class average prototype of the new task to the server after each round of training is completed; The server receives the model parameters and class average prototypes updated by the client in each round of training after acquiring new task data, and sends the updated model parameters and class average prototypes after completing each round of training. 9. A federated incremental learning device based on constrained gradient update, characterized in that it is applied to a client and comprises: A first receiving module is used to receive a personalized global model and a cross-task collaborative loss of a previous task sent by a server, wherein the personalized global model is obtained by the server performing personalized aggregation on multiple model parameters and multiple class average prototypes obtained by training all clients based on local data of their respective previous tasks, and the cross-task collaborative loss represents the degree of similarity between the model parameters of the previous tasks of all clients and the model parameters of historical tasks, wherein the historical tasks represent all tasks whose training sequence is prior to the previous task; An update module is used to update the model parameters of the previous task in a direction orthogonal to its gradient space based on the basis vectors of the gradient space of the previous task, the predicted label and the true label of the new task, and the cross-task collaborative loss when the local data of the new task is obtained, so as to obtain the model parameters of the new task in this round of training; The acquisition module is used to perform weighted averaging on the class prototypes of the new task to obtain the class average prototype of the new task in this round of training; The first training module is used to repeat the updating of the above-mentioned model parameters and class average prototypes, perform preset rounds of training, and send the model parameters and class average prototypes of the new task to the server after each round of training is completed. 10. A federated incremental learning device based on constrained gradient update, characterized in that it is applied to a server, and the device comprises: A sending module, used to send a personalized global model and a cross-task collaborative loss of a previous task to each client, wherein the personalized global model is obtained by the server performing personalized aggregation on a plurality of model parameters and a plurality of class average prototypes obtained by training all clients based on local data of their respective previous tasks, and the cross-task collaborative loss indicates the similarity between the model parameters of the previous tasks of all clients and the model parameters of historical tasks, wherein the historical tasks indicate all tasks whose training sequence precedes the previous task; The first receiving module is used to receive the updated model parameters and class average prototypes sent by the client in each round of training after acquiring new task data, and the updated model parameters and class average prototypes sent after completing each round of training. The model parameters are based on the basis vectors of the gradient space of the previous task, the predicted labels and true labels of the new task, and the cross-task collaborative loss. The model parameters of the previous task are updated along a direction orthogonal to its gradient space. The class average prototype is obtained based on the weighted average of the class prototypes of the new task.