WO2022041947A1

WO2022041947A1 - Method for updating machine learning model, and communication apparatus

Info

Publication number: WO2022041947A1
Application number: PCT/CN2021/100003
Authority: WO
Inventors: 杨水根; 晋英豪; 秦东润; 周彧
Original assignee: 华为技术有限公司
Priority date: 2020-08-24
Filing date: 2021-06-15
Publication date: 2022-03-03
Also published as: CN114091679A

Abstract

A method for updating a machine learning model, and a communication apparatus, which relate to the technical field of artificial intelligence and the technical field of communications. The method comprises: a network device determining corresponding model training configuration information for a terminal device according to the computing capability of the terminal device; receiving, after sending the model training configuration information to the terminal device, a model update parameter sent by the terminal device, wherein the model update parameter is a model parameter that is updated after the terminal device trains a first machine learning model according to the model training configuration information; and updating a second machine learning model according to the received model update parameter. Therefore, according to the computing capability of each terminal device, the difference in the time for each terminal device to report a model update parameter to a network device is reduced, and the network device can finish model updating in as short a time as possible on the basis of the model update parameter reported by each terminal device, thereby improving the convergence speed of model updating, and improving the update efficiency of a machine learning model.

Description

A method and communication device for updating a machine learning model

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the priority of the Chinese patent application filed on August 24, 2020 with the application number of 202010858858.7 and the application title of "A method and communication device for updating a machine learning model", the entire contents of which are incorporated by reference in this application.

technical field

The present application relates to the technical field of artificial intelligence, and in particular, to a method and a communication device for updating a machine learning model.

Background technique

Wireless communication networks are developing in the direction of network diversification, broadbandization, integration and intelligence. Wireless transmission uses higher and higher frequency spectrums, wider bandwidths, and more and more antennas. Traditional communication methods are too complex and difficult to guarantee performance. Furthermore, with the explosion of smart terminals and various applications, wireless communication network behavior and performance factors are more dynamic and unpredictable than in the past. Operating increasingly complex wireless communication networks with low cost and high efficiency is a challenge facing operators today.

With the development of artificial intelligence (AI) technology and machine learning (ML) technology, AI/ML will also assume more and more important tasks in wireless communication networks. Currently, in wireless communication networks, AI/ML is being introduced in the fields of physical layer, media intervention control, wireless resource control, wireless resource management, and operation and maintenance between the terminal device and the network side. Terminal equipment and network equipment (such as base stations), as part of the wireless communication network, can introduce AI/ML to process related communication transactions. The trained machine learning model handles the relevant communication transactions.

The data in terminal devices generally involve user privacy. In order to avoid leakage of user privacy data, each terminal device generally uses its own data locally to train the machine learning model pre-distributed by the network device, and then The obtained model update parameters are sent to the network device, and then the network device aggregates the model update parameters sent by each participant (ie, each terminal device that performs model training), and then directly updates the local machine learning model of the network device. In this process, the user data in each terminal device is generally different, because the ability of each terminal device to perform model training is generally different, and the configuration information of each terminal device for model training is configured by each terminal itself or by the terminal. The time it takes for each terminal device to perform machine learning model training is generally different, so the time it takes for the network device to receive the model update parameters uploaded by each terminal device is generally quite different, and the network device performs the local model. Update is a model update parameter that needs to be reported by all participants. Therefore, due to the time difference between the respective model update parameters reported by each terminal device, the time required for the network device to update the local model is also affected. The convergence speed is slower and the update efficiency is lower.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a method and a communication device for updating a machine learning model, which are used to improve the convergence speed of updating the machine learning model, so as to improve the updating efficiency of the machine learning model.

In a first aspect, a method for updating a machine learning model is provided, and the method can be applied to a network device or a chip inside the network device. Taking the method applied to a network device as an example, in this method, the network device determines the corresponding model training configuration information for the terminal device according to the computing capability of the terminal device, and receives the model training configuration information after sending the model training configuration information to the terminal device. The model update parameter sent by the terminal device is then updated according to the received model update parameter to the second machine learning model in the network device. The model update parameter sent by the terminal device is an update parameter obtained by the terminal device performing local training on the local first machine learning model according to the model training configuration information sent by the network device.

The machine learning model local to the terminal device is called the first machine learning model, the machine learning model local to the network device is called the second machine learning model, and the first machine learning model is distributed by the network device for the terminal device. The first machine learning model and the second machine learning model are of the same type of machine learning model, or the first machine learning model and the second machine learning model are different types of machine learning models. The model update parameter information is used for local model update, and the first machine learning model and the second machine learning model are the same type of machine learning model.

In the embodiment of the present application, the network device allocates corresponding model training configuration information to each terminal device according to the computing capability of each terminal device, so that the model training configuration information used by each terminal device to train the local machine learning model is the same as the In contrast to the way in which each terminal device independently selects the model training configuration information in the related art, in this solution, the network side uniformly calculates the difference of each terminal device according to the computing capability of each terminal device itself. The corresponding model training configuration information can be configured so as to reduce the time difference caused by different capabilities of each terminal device during model training, thereby ensuring that each terminal device can complete the model training within the same time as possible, so that each terminal device can report The time of the respective model update parameters is roughly the same, which reduces the difference in the time when each terminal device reports the model update parameters, thereby reducing the time difference between the network device receiving the model update parameters sent by each terminal device, so that the network device based on each The model update parameters reported by the terminal device can be completed in a short time as much as possible to improve the convergence speed of the model update, thereby improving the update efficiency of the machine learning model.

In a possible implementation manner, the network device may receive the first computing power indication information from the terminal device, or may receive the second computing power indication information from the terminal device after sending a computing capability acquisition request to the terminal device, or may Receive third computing power indication information from other network devices.

Wherein, the first computing power indication information, the second computing power indication information, and the third computing power indication information are all information used to indicate the computing capability of the terminal device, that is to say, in this embodiment, three types of acquisition terminal equipment are provided. In this way, the flexibility of the method for acquiring the computing power of the terminal device can be improved.

In a possible implementation manner, the model training configuration information includes at least one of hyperparameters, precision, and training time information.

In this solution, the network device can configure one or more model training configuration information for the terminal device according to the computing capability of the terminal device, and the configuration flexibility is high. In addition, the configured model training configuration information is routinely used by terminal devices for model training, which generally meets the configuration requirements of most terminal devices for local model training, and has good versatility.

In a possible implementation manner, the network device further sends training feature information to the terminal device, where the training feature information is used to indicate the training feature set used by the terminal device to train the first machine learning model in the terminal device.

In this solution, the network device sends the training feature information to the terminal device, so that each terminal device participating in the local training can use the same training feature information to perform local training, thereby reducing the time when each terminal device performs local training based on different training feature information. difference in time spent.

In a possible implementation manner, the network device further sends accuracy evaluation information to the terminal device, where the accuracy evaluation information includes at least one of a method for evaluating the accuracy or a test sample for evaluating the accuracy.

In this solution, by specifying the accuracy evaluation information to the terminal device, each terminal device participating in the local model training can use the same accuracy evaluation information to evaluate the accuracy of the locally trained machine learning model, because the same accuracy is used. The evaluation method can try to make each terminal device meet the specified accuracy requirements under the same accuracy evaluation standard, thereby reducing the difference in the time spent by each terminal device for local training.

In a possible implementation manner, the network device further receives accuracy indication information from the terminal device, where the accuracy indication information is used to instruct the terminal device to perform a local first machine learning model on the local first machine learning model by using the model training configuration information sent by the network device. Accuracy achieved after training.

In this solution, in addition to feeding back the model update parameters to the network device, the terminal device can also feed back the accuracy of the corresponding model training to the network device, so that the network device can know the training effect of the terminal device, which can be used as a follow-up When configuring the model training configuration information for the terminal device, it is used as a reference to maximize the training effect.

In a possible implementation manner, before the network device receives the model update parameters sent by the terminal device, it also determines a time point for acquiring the model update parameters of the terminal device, and sends an acquisition request to the terminal device at the time point, The obtaining request is used to instruct the terminal device to send the model update parameter of the terminal device to the network device.

In this solution, the network device can explicitly control the time for requesting model update parameters from each terminal device. On the basis of reducing the time difference between each terminal device completing local training through the model training configuration information, it can further reduce the reporting model of each terminal device. The time difference between the update parameters is reduced, thereby reducing the time difference between the network device actually obtaining the model update parameters sent by each terminal device.

In a possible implementation manner, before the network device receives the model update parameters sent by the terminal device, it also determines a time point for acquiring the model update parameters of the terminal device, and sends reporting time information to the terminal device. The information is used to indicate that the model update parameters are sent to the network device at the determined aforementioned time point.

In this solution, the network device can clearly control the specific time when each terminal device reports the model update parameters. On the basis of reducing the time difference between each terminal device completing local training through the model training configuration information, it can further reduce the reporting model of each terminal device. The time difference between the update parameters is reduced, thereby reducing the time difference between the network device actually obtaining the model update parameters sent by each terminal device.

In a possible implementation manner, the network device determines the transmission duration for each of the multiple terminal devices to send their respective model update parameters to the network device, and determines to obtain the model update of the terminal device according to the transmission duration corresponding to each terminal. The acquisition time of the parameter. The plurality of terminal devices may include the aforementioned terminal devices, or may not include the aforementioned terminal devices.

In this solution, the network device actively requests model update parameters from each terminal device, and sends an acquisition request for requesting model update parameters to each corresponding terminal device at a time matching each terminal device. In addition, the time required for updating parameters of the transmission model of most or even all participants can be comprehensively considered, and the time for each terminal device to report its own model update parameters can be more accurately controlled, thereby reducing the network device acquiring the data sent by each terminal device. The time difference of the model update parameters, thereby improving the convergence speed of the local model update and improving the model update efficiency.

In a possible implementation manner, the above obtaining request is also used to indicate that the specified model update parameters need to be obtained.

In this solution, the network device can instruct the terminal to upload specific model update parameters, not necessarily all model update parameters, which can reduce the amount of data and time required for the terminal device to transmit the model update parameters to the network device, thereby minimizing invalidation The transmission improves the effectiveness of transmission, saves network transmission resources, and reduces air interface resource overhead.

In a possible implementation manner, the network device further receives parameter availability indication information from the terminal device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

In this solution, the availability of the model update parameters in the terminal device can be indicated by the parameter availability indication information. In this way, the network device can specify the availability of various model update parameters in the terminal device according to the parameter availability indication information sent by the terminal device. , which increases the cognitive consistency between the network device and the terminal device, and the network device can also be more specific when acquiring the model update parameters sent by the terminal device.

In a second aspect, a method for updating a machine learning model is provided, and the method can be applied to a terminal device or a chip inside the terminal device. Taking the method applied to a terminal device as an example, in this method, the terminal device receives the model training configuration information sent by the network device, wherein the model training configuration information is determined according to the computing capability of the terminal device, and further, the terminal device is Perform local training on the first machine learning model in the terminal device according to the received model training configuration information to obtain model update parameters, and then send the obtained model update parameters to the network device, so that the network device updates the parameters according to the model A local update is made to the second machine learning model in the network device.

The understanding of the first machine learning model and the second machine learning model can be understood according to the description of the first machine learning model and the second machine learning model in the first aspect.

In the embodiment of the present application, the model training configuration information for the terminal device to perform local machine learning model training is determined by the network device according to the computing capability of the terminal device itself, so the model training configuration information is related to the computing capability of the terminal device itself. Matching, it can ensure that each terminal device can complete the model training in the same time as possible. In this way, the network device can configure the corresponding model training configuration information for each terminal device participating in the local training according to its own computing capability, so that the time for each terminal device to report its own model update parameters is roughly the same, reducing the number of terminal devices. The difference in the time when the device reports the model update parameters, thereby reducing the time difference between the network device receiving the model update parameters sent by each terminal device, so that the network device can update the model based on the model update parameters reported by each terminal device in as short a time as possible. It can be completed in time to improve the convergence speed of model update, thereby improving the update efficiency of the machine learning model.

In a possible implementation manner, before receiving the model training configuration information sent by the network device, the terminal device receives the computing capability acquisition request sent by the network device, and according to the computing capability acquisition request, sends to the network device an instruction to indicate the terminal The second computing power indication information of the computing power of the device.

In a possible implementation manner, on the basis of receiving the model training configuration information sent by the network device, the terminal device also receives training feature information from the network device, where the training feature information is used to indicate the machine learning in the terminal device. The training feature set used by the model for training, and then the first machine learning model in the terminal device is locally trained according to the model training configuration information and the training feature information.

In a possible implementation manner, the terminal device further receives accuracy evaluation information from the network device, where the accuracy evaluation information includes at least one of a method for evaluating the accuracy or a test sample for evaluating the accuracy, and then according to the accuracy The evaluation information determines the accuracy achieved by the trained first machine learning model.

In a possible implementation manner, the terminal device also sends accuracy indication information to the network device, where the accuracy indication information is used to instruct the terminal device to use the model training configuration information sent by the network device to train its local machine learning model after training. achieved accuracy.

In a possible implementation manner, the terminal device further receives an acquisition request from the network device, where the acquisition request is used to instruct the terminal device to send the model update parameters of the terminal device to the network device, and then send the network device to the network device according to the acquisition request. Send model update parameters.

In a possible implementation manner, the terminal device further receives reporting time information from the network device, and sends the model update parameter to the network device at the time point indicated by the reporting time information.

In a possible implementation manner, the terminal device further sends parameter availability indication information to the network device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

For the technical effects that can be achieved by any implementation manner of the foregoing second aspect, reference may be made to the description of the beneficial effects in the foregoing first aspect, which will not be repeated here.

In a third aspect, a method for updating a machine learning model is provided, and the method can be applied to a network device, or can be applied to a chip in the network device. Taking the method applied to a network device as an example, in this method, the network device selects and obtains the model update parameter of the first terminal device according to the transmission duration of each terminal device in the plurality of terminal devices sending their respective model update parameters to the network device. at the selected time point, and then send an acquisition request to the first terminal device at the selected time point, or send to the terminal device the reporting time information for instructing the first terminal device to send the model update parameters at this time point, and receive the first terminal device The sent model update parameter, and then locally update the second machine learning model in the network device according to the model update parameter.

In this embodiment of the present application, the network device actively requests each terminal device for the model update parameters in each terminal device, and the time point of the request is the amount of data that the network device actually needs to transmit according to each terminal device and the transmission chain situation (the quality) is determined. Specifically, the network device selects the time point for acquiring the model update parameters of each terminal device according to the transmission duration required for each terminal device to send the respective model update parameters to the network device, and according to the difference between the transmission durations of each terminal device Differentially request the respective model update parameters from each terminal device at different time points, which can minimize the time difference between the network device receiving the model update parameters sent by each terminal device due to the difference in transmission time, so that each terminal device can The model update parameters sent by the device can reach the network device at the same time (or approximately the same short period of time) as much as possible, reducing the time difference between each terminal device transmitting the model update parameters to the network device, thereby reducing the network device acquiring the model of each terminal device. The time difference between the update parameters, so that the network device can locally update the local machine learning model according to the model update parameters of each terminal device in a short time, so as to improve the convergence speed of the local update, and then improve the update of the machine learning model. effectiveness.

In a possible implementation manner, the network device further receives parameter availability indication information sent by the first terminal device.

In this solution, the availability of the model update parameters in the terminal device can be indicated by the parameter availability indication information. In this way, the network device can specify the availability of various model update parameters in the terminal device according to the parameter availability indication information sent by the terminal device. , which increases the cognitive consistency between the network device and the terminal device, and the network device can also be more explicit when acquiring the model update parameters sent by the terminal device.

In a possible implementation manner, the network device indicates the specified model update parameter to the first terminal device. Optionally, the model update parameters specified by the network device may be instructed through the acquisition request.

In a fourth aspect, a method for updating a machine learning model is provided, and the method can be applied to a terminal device, or can be applied to a chip in the terminal device. Taking the method applied to the first terminal device as an example, in this solution, the first terminal device receives the acquisition request sent by the network device, and the time point when the acquisition request is sent is when the network device sends the request from each of the multiple terminal devices to the network device. The transmission duration for the network device to send the respective model update parameters is determined, and the acquisition request is sent at the determined time point, and then the model update parameters are sent to the network device according to the acquisition request, or the report time information sent by the network device is received, and The model update parameter is sent to the network device at the time point indicated by the reporting time information, so that the network device locally updates the local machine learning model of the network device according to the model update parameter sent by the first terminal device.

In a possible implementation manner, the first terminal device further sends parameter availability indication information to the network device.

In a possible implementation manner, the first terminal device further receives indication information sent by the network device, where the indication information is used to indicate a model update parameter specified by the network device.

In a possible implementation manner, the indication information used to indicate the model update parameter specified by the network device is the above-mentioned acquisition request.

For the technical effects that can be achieved by any of the implementation manners of the above-mentioned fourth aspect, reference may be made to the description of the beneficial effects in the above-mentioned third aspect, which will not be repeated here.

A fifth aspect provides a communication apparatus, the communication apparatus may be a network device, or a chip set inside the network device, and the communication apparatus includes the first aspect or any possible implementation manner of the first aspect. Modules for the methods described in . Exemplarily, the communication device includes a processing unit and a communication unit, wherein:

a processing unit, configured to determine model training configuration information corresponding to the terminal device according to the computing capability of the terminal device;

A communication unit, configured to send the model training configuration information to the terminal device, and receive the model update parameter sent by the terminal device, wherein the model update parameter is the model parameter updated by the terminal device after training the first machine learning model according to the model training configuration information ;

The processing unit is further configured to update the second machine learning model according to the model update parameter.

In a possible implementation manner, the communication unit is further used for:

receiving first computing power indication information from the terminal device, where the first computing power indication information is used to indicate the computing capability of the terminal device; or,

After sending a computing capability acquisition request to the terminal device, receive second computing power indication information from the terminal device, where the second computing power indication information is used to indicate the computing capability of the terminal device; or,

Receive third computing power indication information from other network devices, where the third computing power indication information is used to indicate the computing capability of the terminal device.

In a possible implementation manner, the communication unit is further configured to send training feature information to the terminal device, where the training feature information is used to instruct the terminal device to use for training the first machine learning model training feature set.

In a possible implementation manner, the communication unit is further configured to send accuracy evaluation information to the terminal device, where the accuracy evaluation information includes at least one of a method for evaluating accuracy or a test sample for evaluating accuracy kind.

In a possible implementation manner, the communication unit is further configured to receive accuracy indication information from the terminal device, where the accuracy indication information is used to instruct the terminal device to use the model training configuration information to The accuracy achieved by a machine learning model after training.

In a possible implementation manner, the processing unit is further configured to determine a time point for acquiring the model update parameter of the terminal device; then, the communication unit is further configured to:

Send an acquisition request to the terminal device at the time point, where the acquisition request is used to instruct the terminal device to send the model update parameter of the terminal device to the network device; or,

Sending reporting time information to the terminal device, where the reporting time information is used to indicate that model update parameters are sent to the network device at the time point.

In a possible implementation manner, the processing unit is specifically used for:

determining the transmission duration for each terminal device in the plurality of terminal devices to send the respective model update parameters to the network device;

The time point is determined and acquired according to each transmission duration corresponding to each terminal device.

In a possible implementation manner, the obtaining request is further used to indicate that the specified model update parameters need to be obtained.

In a possible implementation manner, the communication unit is further configured to receive parameter availability indication information from the terminal device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

For the technical effects that can be achieved by any of the implementation manners of the above fifth aspect, reference may be made to the description of the beneficial effects in the above first aspect, which will not be repeated here.

In a sixth aspect, a communication device is provided, the communication device may be a terminal device, or a chip provided inside the terminal device, and the communication device includes the second aspect or any possible implementation manner of the second aspect. Modules for the methods described in . Exemplarily, the communication device includes a communication unit and a processing unit, wherein:

a communication unit, configured to receive model training configuration information sent by the network device, where the model training configuration information is determined according to the computing capability of the terminal device;

a processing unit, configured to train the first machine learning model according to the model training configuration information to obtain model update parameters;

The communication unit is further configured to send the model update parameter to the network device, where the model update parameter is used by the network device to update the second machine learning model.

receiving a computing capability acquisition request sent by the network device;

According to the computing capability acquisition request, second computing power indication information is sent to the network device, where the second computing power indication information is used to indicate the computing capability of the terminal device.

In a possible implementation manner, the communication unit is further configured to receive training feature information from the network device, where the training feature information is used to instruct the terminal device to perform training on the first machine learning model. The used training feature set; then, the processing unit is further configured to train the first machine learning model according to the model training configuration information and the training feature information.

In a possible implementation manner, the communication unit is further configured to receive accuracy evaluation information from the network device, where the accuracy evaluation information includes at least one of a method for evaluating accuracy or a test sample for evaluating accuracy One; then, the processing unit is further configured to determine the accuracy achieved by the trained machine learning model according to the accuracy evaluation information.

In a possible implementation manner, the communication unit is configured to send accuracy indication information to the network device, where the accuracy indication information is used to instruct the terminal device to use the model training configuration information to The accuracy achieved by the learned model after training.

In a possible implementation manner, the communication unit is further configured to receive an acquisition request from the network device, and send the model update parameter to the network device according to the acquisition request, where the acquisition request is used for The terminal device is instructed to send the model update parameter of the terminal device to the network device.

In a possible implementation manner, the communication unit is further configured to receive report time information from the network device, and send the model update parameter to the network device at the acquisition time indicated by the report time information .

In a possible implementation manner, the communication unit is further configured to send parameter availability indication information to the network device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

For the technical effects that can be achieved by any of the implementation manners of the above sixth aspect, reference may be made to the description of the beneficial effects in the above second aspect, which will not be repeated here.

In a seventh aspect, a communication device is provided, the communication device may be a network device, or a chip set inside the network device, and the communication device includes the third aspect or any possible implementation manner of the third aspect. Modules for the methods described in . Exemplarily, the communication device includes a processing unit and a communication unit, wherein:

a processing unit, configured to select a time point for acquiring the model update parameters of the first terminal device according to the transmission duration of each terminal device in the plurality of terminal devices sending their respective model update parameters to the network device;

The communication unit is configured to send an acquisition request to the first terminal device at the time point, or send to the terminal device the reporting time information for instructing the first terminal device to send the model update parameters at the acquisition time, wherein the acquisition request uses in requesting the first terminal device to send model update parameters to the network device; and receiving the model update parameters sent by the first terminal device;

The processing unit is further configured to update the second machine learning model in the network device according to the model update parameter.

In a possible implementation manner, the communication unit is further configured to receive parameter availability indication information from the first terminal device.

In a possible implementation manner, the communication unit is further configured to receive indication information from the network device for indicating the specified model update parameter.

In a possible implementation manner, the indication information is carried in the acquisition request.

For the technical effects that can be achieved by any one of the implementation manners of the above seventh aspect, reference may be made to the description of the beneficial effects in the above third aspect, which will not be repeated here.

In an eighth aspect, a communication device is provided, the communication device may be a terminal device, or a chip set inside the terminal device, and the communication device includes the fourth aspect or any possible implementation manner of the fourth aspect. Modules for the methods described in . Exemplarily, the communication device includes a communication unit and a processing unit, wherein:

A communication unit, configured to receive an acquisition request sent by a network device or report time information, wherein the time point at which the acquisition request is sent is when the network device sends the transmission of the respective model update parameters to the network device according to each terminal device in the plurality of terminal devices fixed duration;

a processing unit, configured to determine model update parameters to be sent according to the acquisition request;

The communication unit is further configured to send the determined model update parameter to the network device, or send the model update parameter to the network device at the time point indicated by the reporting time information, and the model The update parameter is used by the network device to update the second machine learning model in the network device.

In a possible implementation manner, the communication unit is further configured to send parameter availability indication information to the network device.

For the technical effects that can be achieved by any one of the implementation manners of the above-mentioned eighth aspect, reference may be made to the description of the beneficial effects in the above-mentioned fourth aspect, which will not be repeated here.

In a ninth aspect, a communication device is provided, comprising: at least one processor; and a communication interface communicatively connected to the at least one processor; the at least one processor executes instructions stored in a memory so that the communication device passes The communication interface performs the method as described in the first aspect or any possible implementation of the first aspect.

Optionally, the memory is located outside the device.

Optionally, the apparatus includes the memory connected to the at least one processor, the memory storing instructions executable by the at least one processor.

In a tenth aspect, a communication device is provided, comprising: at least one processor; and a communication interface communicatively connected to the at least one processor; the at least one processor executes instructions stored in a memory, so that the communication device passes The communication interface performs the method as described in the second aspect or any possible implementation of the second aspect.

Optionally, the memory is located outside the device.

In an eleventh aspect, a communication device is provided, comprising: at least one processor; and a communication interface communicatively connected to the at least one processor; the at least one processor executes instructions stored in a memory to enable the communication device The method as described in the third aspect or any possible implementation manner of the third aspect is performed through the communication interface.

Optionally, the memory is located outside the device.

A twelfth aspect provides a communication device, comprising: at least one processor; and a communication interface communicatively connected to the at least one processor; the at least one processor causes the communication device to execute instructions stored in a memory by the at least one processor The method as described in the fourth aspect or any possible implementation manner of the fourth aspect is performed through the communication interface.

Optionally, the memory is located outside the device.

A thirteenth aspect provides a computer-readable storage medium, comprising a program or an instruction, when the program or instruction is run on a computer, the first aspect or any possible implementation manner of the first aspect is as described in the first aspect. method is executed.

A fourteenth aspect provides a computer-readable storage medium, comprising a program or an instruction, which, when the program or instruction is run on a computer, makes the second aspect or any possible implementation of the second aspect as described in the second aspect method is executed.

A fifteenth aspect provides a computer-readable storage medium, comprising a program or an instruction, which, when the program or instruction is run on a computer, makes the third aspect or any possible implementation of the third aspect described in the third aspect method is executed.

A sixteenth aspect provides a computer-readable storage medium, comprising a program or an instruction, when the program or instruction is run on a computer, the fourth aspect or any possible implementation manner of the fourth aspect is provided. method is executed.

A seventeenth aspect provides a chip, which is coupled to a memory and configured to read and execute program instructions stored in the memory, so that the first aspect or any possible implementation manner of the first aspect is described in the first aspect. method is executed.

An eighteenth aspect provides a chip, which is coupled to a memory and used to read and execute program instructions stored in the memory, so that the second aspect or any possible implementation manner of the second aspect is described in the second aspect. method is executed.

A nineteenth aspect provides a chip, which is coupled to a memory and configured to read and execute program instructions stored in the memory, so that the third aspect or any of the possible implementations of the third aspect is described in the third aspect. method is executed.

A twentieth aspect provides a chip, which is coupled to a memory and used to read and execute program instructions stored in the memory, so that the fourth aspect or any of the possible implementations of the fourth aspect is described in the method is executed.

A twenty-first aspect provides a computer program product comprising instructions, when executed on a computer, to cause the method described in the first aspect or any of the possible implementations of the first aspect to be performed.

A twenty-second aspect provides a computer program product comprising instructions that, when run on a computer, cause the method described in the second aspect or any of the possible implementations of the second aspect to be performed.

A twenty-third aspect provides a computer program product comprising instructions, which when run on a computer, cause the above third aspect or any of the possible implementations of the third aspect to be performed.

A twenty-fourth aspect provides a computer program product comprising instructions, which when run on a computer, cause the method described in the fourth aspect or any of the possible implementations of the fourth aspect to be performed.

It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

Description of drawings

Figure 1 is a schematic diagram of applying federated learning to ML model training;

FIG. 2 is a schematic diagram of an application scenario of an embodiment of the present application;

FIG. 3 is a schematic diagram of a device architecture of a separate access network according to an embodiment of the present application;

4 is a flowchart of a method for updating a machine learning model provided by an embodiment of the present application;

5 is a flowchart of another method for updating a machine learning model provided by an embodiment of the present application;

FIG. 6 is a schematic structural diagram of a communication device in an embodiment of the present application;

7 is a schematic structural diagram of another communication device in an embodiment of the present application;

FIG. 8 is a schematic structural diagram of another communication device in an embodiment of the present application;

FIG. 9 is a schematic structural diagram of another communication device in an embodiment of the present application;

10 is a schematic structural diagram of another communication device in an embodiment of the present application;

FIG. 11 is a schematic structural diagram of another communication device in an embodiment of the present application.

detailed description

In order to make the objectives, technical solutions and advantages of the embodiments of the present application more clear, the embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

Hereinafter, some terms in the embodiments of the present application will be explained, so as to facilitate the understanding of those skilled in the art.

(1) Terminal devices, including devices that provide voice and/or data connectivity to users, may include, for example, handheld devices with wireless connectivity, or processing devices connected to wireless modems. The terminal equipment may communicate with the core network via a radio access network (RAN), and exchange voice and/or data with the RAN. The terminal equipment may include user equipment (UE), terminal, wireless terminal equipment, mobile terminal equipment, device-to-device (D2D) terminal equipment, vehicle-to-everything (vehicle-to-everything, V2X) terminal equipment, machine-to-machine/machine-type communications (M2M/MTC) terminal equipment, Internet of things (IoT) terminal equipment, subscriber unit (subscriber unit), Subscriber station (subscriber station), mobile station (mobile station), remote station (remote station), access point (access point, AP), remote terminal (remote terminal), access terminal (access terminal), user terminal (user terminal), user agent, or user device, etc. For example, these may include mobile telephones (or "cellular" telephones), computers with mobile terminal equipment, portable, pocket-sized, hand-held, computer-embedded mobile devices, and the like. For example, personal communication service (PCS) phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, personal digital assistants (personal digital assistants), PDA), etc. Also includes constrained devices, such as devices with lower power consumption, or devices with limited storage capacity, or devices with limited computing power, etc. For example, it includes information sensing devices such as barcodes, radio frequency identification (RFID), sensors, global positioning system (GPS), and laser scanners.

As an example and not a limitation, in this embodiment of the present application, the terminal device may also be a wearable device. Wearable devices can also be called wearable smart devices or smart wearable devices, etc. It is a general term for the application of wearable technology to intelligently design daily wear and develop wearable devices, such as glasses, gloves, watches, clothing and shoes. Wait. A wearable device is a portable device that is worn directly on the body or integrated into the user's clothing or accessories. Wearable device is not only a hardware device, but also realizes powerful functions through software support, data interaction, and cloud interaction. In a broad sense, wearable smart devices include full-featured, large-scale, complete or partial functions without relying on smart phones, such as smart watches or smart glasses, and only focus on a certain type of application function, which needs to cooperate with other devices such as smart phones. Use, such as all kinds of smart bracelets, smart helmets, smart jewelry, etc. for physical sign monitoring.

The various terminal devices described above, if they are located on the vehicle (for example, placed in the vehicle or installed in the vehicle), can be considered as on-board terminal equipment. For example, the on-board terminal equipment is also called on-board unit (OBU). ).

(2) Network equipment, including, for example, an access network (AN) equipment, such as a base station (eg, an access point), may refer to a device that communicates with wireless terminal equipment through one or more cells over the air interface in the access network. The device, or, for example, an access network device in a V2X technology is a road side unit (RSU). The base station may be used to convert received air frames to and from Internet Protocol (IP) packets and act as a router between the terminal device and the rest of the access network, which may include the IP network. The RSU can be a fixed infrastructure entity supporting V2X applications and can exchange messages with other entities supporting V2X applications. The access network equipment can also coordinate the attribute management of the air interface. For example, the access network equipment may include a long term evolution (long term evolution, LTE) system or an evolved base station (NodeB or eNB or e-NodeB, evolutional Node B) in long term evolution-advanced (LTE-A) ), or may also include the next generation node B (gNB) and the next generation evolved base station ( next generation evolutional Node B, ng-eNB), or may also include a centralized unit (central unit, CU) and a distributed unit (distributed unit, DU) in a separate access network system, which is not limited in the embodiment of the present application.

Of course, network equipment can also include core network equipment, which can be an access and mobility management function (AMF), which is mainly responsible for functions such as access control, mobility management, attachment and detachment, and gateway selection. The core network device may also be a network data analytics function (NWDAF), which is mainly responsible for functions such as data collection and analysis. The core network device may also be other devices.

(3) AI refers to the technology of presenting human intelligence through computer programs. It is a theory that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. , methods, techniques and application systems. In other words, AI is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

(4) ML is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. It specializes in how computers simulate or realize human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications are in all fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning and other technologies.

The essence of machine learning is to let the computer simulate the above process, let the computer "learn", so as to obtain a certain kind of cognition, through this cognition to judge new things, this kind of cognition can be reflected through the "machine learning model", expressed in mathematics It can be understood as a function.

(5) A machine learning model. In the embodiments of this application, no distinction is made between artificial intelligence and machine learning, and the machine learning model may be represented as an ML model or an AI model. The machine learning model in the embodiments of the present application generally refers to AI models and ML models in the AI field and ML field. For example, the machine learning model includes, for example, linear regression, logistic regression, and decision tree. (decision tree), naive bayes, k-nearest neighbors, support vector machines, deep neural network, random forest, etc.

(6) Federated learning (FL) is an emerging artificial intelligence basic technology and an encrypted distributed ML technology. Its design goals are to ensure information security during big data exchange, protect terminal data and On the premise of personal data privacy and legal compliance, efficient machine learning is carried out between multiple parties or computing nodes. Among them, the machine learning algorithms that can be used in federated learning are not limited to neural networks, but also include algorithms such as random forests. Federated learning is expected to be the basis for the next generation of AI collaborative algorithms and collaborative networks.

Federated learning is a machine learning framework designed on the premise of meeting data privacy, security and regulatory requirements, allowing artificial intelligence systems to use their own data more efficiently and accurately, to meet users' privacy protection and data security. Features of federated learning include:

Data isolation, data will not be leaked to the outside, to meet the needs of user privacy protection and data security;

It can ensure that the quality of the model is not damaged, and there will be no negative migration, and that the federated model is better than the split independent model;

All participants have equal status and can achieve fair cooperation;

It can ensure that each participant can carry out encrypted exchange of information and model parameters while maintaining independence, and grow at the same time.

(7) "At least one" means one or more, and "plurality" means two or more. "And/or", which describes the association relationship of the associated objects, means that there can be three kinds of relationships, for example, A and/or B, it can mean that A exists alone, A and B exist at the same time, and B exists alone, where A, B can be singular or plural. The character "/" generally indicates that the associated objects are an "or" relationship.

And, unless stated to the contrary, the ordinal numbers such as “first” and “second” mentioned in the embodiments of the present application are used to distinguish multiple objects, and are not used to limit the order, sequence, priority or priority of multiple objects. Importance. For example, the first information and the second information are only for distinguishing different signaling, and do not indicate the difference in content, priority, transmission order, or importance of the two kinds of information.

Some concepts involved in the embodiments of the present application are described above, and the technical features of the embodiments of the present application are described below.

As mentioned above, for machine learning models in related technologies, in order to avoid leakage of user privacy data, network equipment first distributes the initial machine learning model to each terminal device, and each terminal device generally uses its own data to pair locally. The training of the machine learning model distributed by the network device, and then the model update parameters obtained after training are sent to the network device, and then the network device aggregates the model update parameters sent by each participant (that is, each of the aforementioned terminal devices), and then The local machine learning model of the network device is directly updated to obtain an updated machine learning model. For example, FL can be used to update the machine learning model. The main feature of FL is that the data of each participant is kept locally and does not need to be uploaded to the network device, thus not revealing data privacy and reducing massive data. Network overhead required for uploading.

In the related art, the process of using FL to update a machine learning model (such as an ML model) in a wireless communication network is shown in Figure 1:

S11. The network device sends the initial ML model to the participant terminal devices 1-N. For the convenience of description, the ML model is represented by y=W ⁰ x+b, where W ⁰ is the initial parameter of the ML model.

S12. The terminal devices 1-N train the ML model based on their respective local training data sets, that is, update the ML model to obtain the updated model parameters W ₁ ⁰ ˜W _N ⁰ , or obtain the updated model parameter difference The values g ₁ =W ₁ ⁰ -W ⁰ , . . . , g _N =W _N ⁰ -W ⁰ , the model parameter differences g ₁ , . . . , g _N are also called gradients.

S13. The terminal devices 1 _- _N send the updated model update parameters or gradients to the network device, for example, send W10 _- ^WN0 to the network device, or send g1 _- ^gN to the network device.

S14. After receiving the model update parameters sent by all the participants (ie, terminal devices 1 to N), the network device performs a weighted average on the model update parameters of each participant to obtain the aggregated ML model update parameters, and then aggregates the model update parameters. The latter ML model update parameter updates the local ML model of the network device. For example, the network device calculates the updated ML model update parameter as W=(W ₁ ⁰ *p ₁ +W ₂ ⁰ *p ₂ +...+W _N ⁰ *p _N )/N, and replaces the original W ⁰ with W . Or, the network device calculates the updated ML model update parameter difference as g=(g ₁ *p ₁ +g ₂ *p ₂ +...+g _N *p _N )/N, so that the updated ML model parameters are obtained as W=g+W ⁰ , and the original W ⁰ is replaced by W. Among them, p ₁ , p ₂ , ..., p _N represent weights, for example, p ₁ refers to the weight of W ₁ ⁰ or g ₁ , p ₂ refers to the weight of W ₂ ⁰ or g ₂ , ..., p _N refers to The weight of W _N ⁰ or g _N , the sum of all weights is 1, ie p ₁ +p ₂ +...+p _N =1.

In the above process shown in FIG. 1 , since the data used locally for ML model training by each of the terminal devices 1 to N is different, the ability of each terminal device to perform ML model training is generally different, and in the training process The configuration information used by each terminal device for training is configured by each terminal device itself, or manually configured by the user of each terminal device, that is, the configuration information used by each terminal device for training is independent of each other. This results in the time for each terminal device to complete the ML model training is generally different, and there is a big difference. In order to facilitate the network device to update the ML model as soon as possible, each terminal device will generally report in time after obtaining the model update parameters. For the network device, due to the time difference between each terminal device reporting its own model update parameters, and the network device needs to obtain the model update parameters reported by all participants before the model update can be performed, so it needs to wait until the last reported model update parameter. When the model is updated, it may have been a long time since the time when the first model update parameters were received, which undoubtedly increases the time it takes for the network device to update the model, and causes the network device to update the machine learning model. The slower the convergence speed, the lower the update efficiency of the machine learning model.

By analyzing related technologies, the inventor found that the main reason for the slow efficiency of local updating of machine learning models by network devices is that the configuration information used by terminal devices for training does not take into account the differences between terminal devices, such as Due to differences in device capabilities and training data, each terminal device is completely isolated from each other when configuring the configuration information for local training. Therefore, the time for each terminal device to complete local training varies greatly. In view of this, an embodiment of the present application provides a method for updating a machine learning model. In the method, the network side uniformly configures configuration information for training (referred to as model training configuration information in the embodiment of the present application) for each terminal device, Specifically, the network device allocates corresponding model training configuration information to each terminal device according to the computing capability of each terminal device, so that each terminal device uses the model training configuration used when training the local machine learning model pre-distributed by the network device The information is matched with its own computing power, which can minimize the time difference caused by the difference in the capabilities of each terminal device during model training, thereby ensuring that each terminal device can complete the model training in the same time as possible, so that each terminal device can complete the model training within the same time. The time for the devices to report their respective model update parameters is roughly the same, thereby reducing the difference in the time when each terminal device reports the model update parameters, so that the network device can update the model based on the model update parameters reported by each terminal device in as short a time as possible. It can be completed in time to improve the convergence speed of model update, thereby improving the update efficiency of the machine learning model.

The technical solutions provided in the embodiments of this application can be applied to the 4G system of the fourth generation mobile communication technology (the 4th generation, 4G), such as the LTE system, or can be applied to the 5G system, such as the NR system, or can also be applied to the next generation mobile communication technology. The communication system or other similar communication systems are not specifically limited.

The machine learning model to be trained and updated in this embodiment of the present application may be some general models in the AI field, such as the aforementioned linear regression, logistic regression, decision tree, naive Bayes, K-nearest neighbor, support vector machine, and deep neural network , random forest, etc., which are not limited in the embodiments of the present application. In addition, the machine learning models for training in each terminal device participating in the model training (terminal devices 1 to N in FIG. 1 ) are uniformly distributed in advance by the network device, that is to say, each terminal device performs model training. The machine learning model is of the same type, and the machine learning model in which the network device uses the model update parameters reported by each terminal device to update the model is of the same type as the machine learning model in each terminal device for model training.

The following introduces a network architecture applied in the embodiment of the present application, and please refer to FIG. 2 .

FIG. 2 is a schematic diagram of a system architecture provided by an embodiment of the present application. As shown in FIG. 2 , the communication system includes a core network device, a first access network device, a second access network device, and a terminal device. The first access network device or the second access network device can communicate with the core network device; the terminal device can communicate with the first access network device or the second access network device, and the terminal device can also communicate with the first access network device or the second access network device. The access network device and the second access network device communicate at the same time, that is, multi-radio dual connectivity (MR-DC). In the MR-DC scenario, the first access network device may be the primary access network device, the second access network device may be the secondary access network device, or the second access network device may be the primary access network device , the first access network device may be a secondary access network device, wherein the first access network device and the second access network device may be access network devices of different communication modes, or may be access network devices of the same communication mode network equipment.

It can be understood that the foregoing communication system shown in FIG. 2 is only for illustrating the technical solutions of the embodiments of the present application more clearly, and does not constitute a limitation on the technical solutions provided by the embodiments of the present application. For example, the communication system may also include other devices, such as network control devices. The network control device may be an operation management and maintenance (operation administration and maintenance, OAM) system, also called a network management system. The network control device may manage the aforementioned first access network device, second access network device, and core network device.

In addition, those skilled in the art know that, with the evolution of the network architecture and the emergence of new service scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

The core network device in FIG. 2 may be AMF or NWDAF, but is not limited to AMF and NWDAF. The access network device in Figure 2, also known as a radio access network (RAN) device, is a device that connects a terminal device to a wireless network and can provide wireless resource management and services for the terminal device. Features such as quality management, data encryption and compression. Exemplarily, the access network equipment may include the following equipment:

gNB: Provide NR control plane and/or user plane protocols and functions for terminal equipment, and access to 5G core network (5th generation core, 5GC);

ng-eNB: Provides the protocols and functions of the control plane and/or user plane of the evolved universal terrestrial radio access (E-UTRA) for terminal equipment, and accesses the 5G core network;

CU: It mainly includes the radio resource control (RRC) layer of gNB, the service data adaptation protocol (SDAP) layer and the packet data donvergence protocol (PDCP) layer, or ng- RRC layer and PDCP layer of eNB;

DU: It mainly includes the radio link control (RLC) layer, the media access control (MAC) layer and the physical layer of the gNB or ng-eNB;

Central unit-control plane (CU-CP): The control plane of the CU, mainly including the RRC layer in the gNB-CU or ng-eNB-CU, and the control plane in the PDCP layer;

Central unit-user plane (CU-UP): the user plane of the CU, mainly including the SDAP layer in the gNB-CU or ng-eNB-CU, and the user plane in the PDCP layer;

Data analysis and management (DAM): mainly responsible for data collection, ML model training, ML model generation, ML model update, ML model distribution and other functions.

FIG. 3 is a schematic structural diagram of a separate access network device. The access network equipment is divided into one CU and one or more DUs according to functions, wherein the CU and the DU are connected through the F1 interface. Further, one CU may include one CU-CP and one or more CU-UPs. CU-CP and CU-UP can be connected through E1 interface, CU-CP and DU can be connected through F1 control plane interface (F1-C), CU-UP and DU can be connected through F1 user interface interface (F1-U) to connect. Further, the CU, DU or CU-CP can be connected to the DAM through the G1 interface, respectively. Optionally, the DAM can be used as an internal function of the CU, DU, or CU-CP, respectively. At this time, there is no G1 interface (or the G1 interface is an internal interface, which is invisible to the outside world).

In order to further illustrate the technical solutions provided by the embodiments of the present application, the following detailed descriptions are given in conjunction with the accompanying drawings and specific embodiments. Although the embodiments of the present application provide method operation steps as shown in the following embodiments or the accompanying drawings, more or less operation steps may be included in the method based on routine or without creative work. In steps that logically do not have a necessary causal relationship, the execution order of these steps is not limited to the execution order provided by the embodiments of the present application. In the actual processing process or when the method is executed by the device, the method may be executed sequentially or in parallel according to the methods shown in the embodiments or the accompanying drawings.

The technical solutions provided by the embodiments of the present application are described below with reference to the accompanying drawings.

An embodiment of the present application provides a method for updating a machine learning model. Please refer to FIG. 4 , which is a flowchart of the method. In the following introduction process, taking the method applied to the network architecture shown in FIG. 2 as an example, the network device in the following introduction process may be the aforementioned access network device, or core network device, or network control device. It should be noted that FIG. 4 uses a terminal device as an example to illustrate the technical solution of the present application. In the specific implementation process, other participants as model training can be understood according to the process shown in FIG. 4 .

S41. The network device acquires the computing capability of the terminal device.

Among them, the computing power of the terminal device, also known as the computing power, can be understood as the ability used to indicate or evaluate the speed at which the terminal device processes data, such as the output speed when the terminal device calculates the hash function. For example, you can use It is represented by the number of floating point operations per second (FLOPS). The computing power of a terminal device is positively related to the speed of processing data. For example, the greater the computing power, the faster the data processing speed, and the faster the model training speed. The computing power of the terminal device is related to the hardware configuration performance of the terminal device itself, the smoothness of the operating system and other factors.

In a specific implementation process, the network device may acquire the computing capability of the terminal device in any of the following manners.

way 1

The terminal device actively reports its computing capability to the network device. In this manner, the terminal device may send first computing power indication information for indicating the computing capability of the terminal device to the network device, and the first computing power indication information may also include the identifier of the terminal device. After the computing power indication information is received, the computing capability corresponding to the terminal device can be determined according to the first computing power indication information. In a possible implementation manner, the terminal device may report the computing capability of the terminal device to the network device through a UE assistance information (UE assistance information) message.

The terminal device can actively report its own computing capability to the network device when it is registered to the network device, or it can also actively report its own computing capability information when it receives the initial machine learning model distributed by the network device, or it can still be used. It is to actively report its own computing capability at other times, which is not limited in this embodiment of the present application. Through the way that each terminal device actively reports its own computing power, the network device can obtain the computing power of each terminal device in advance, so that the corresponding model training configuration information can be allocated to each terminal device in a timely manner, and the allocation efficiency can be improved.

way 2

The terminal device reports its computing capability to the network device according to the request of the network device. In this method, when the computing capability of the terminal device needs to be acquired, the network device can send a computing capability acquisition request to the terminal device, and the computing capability acquisition request is used to instruct the terminal device to report its own computing capability to the network device. After the computing capability acquisition request sent by the network device, the terminal device may send the second computing power indication information used to indicate the computing capability of the terminal device to the network device. After the network device receives the second computing power indication information sent by the terminal device , the computing capability of the terminal device can be obtained.

In a specific implementation, the network device sends a UE capability enquiry message to the terminal device, which is used to request to obtain the computing capability of the terminal device. Further, the terminal device sends UE capability information (UE capability information) to the network device. information) message, which contains the computing capability of the terminal device.

In this method, the network device only requests the computing capability of the terminal device, so that the computing capability of each terminal device can be stored locally without the need for advance, the storage consumption can be reduced to a certain extent, and the terminal device can be effectively used computing power.

way 3

The network device obtains the computing capability of the terminal device from other network devices. In this manner, other network devices can actively send the computing capability of the terminal device to the network device, or the network device can also send a request to other network devices first, and other network devices return the terminal to the network device based on the request of the network device. The computing capability of the device, for example, other network devices indicate the computing capability of the terminal device to the network device through the third computing capability indication information.

For example, if the network device is an access network device, the computing capability of the terminal device can be obtained from other access network devices, core network devices, or network control devices. For example, if the network device is a core network device, it can be obtained from the access network device or the network control device. The computing capability of the terminal device is obtained at the network control device. Of course, the premise of implementing this method is that other network devices themselves have the computing capability of the terminal device, or other network devices can acquire the computing capability of the terminal device.

In this embodiment, three ways of acquiring the computing capability of the terminal device are provided, so that the flexibility of the way of acquiring the computing capability of the terminal device can be improved.

S42. The network device determines model training configuration information corresponding to the terminal device according to the computing capability of the terminal device.

The model training configuration information is the configuration information required by the terminal device to train the local machine learning model. In other words, the model training configuration information is used for the terminal device to perform model training on the local machine learning model.

As mentioned above, the speed at which the terminal device processes data can be evaluated based on the computing capability of the terminal device. In order to reduce the time difference in training the local machine learning model by each terminal device, The computing power of the terminal device is used to allocate the corresponding model training configuration information for each terminal device. Based on this allocation mechanism, it can be ensured that each terminal device can complete the model training within approximately the same time period. For example, for terminal devices with poor computing power , model training configuration information with lower requirements can be configured for it. For terminal devices with better computing capabilities, model training configuration information with relatively high requirements can be configured for them. In this way, terminal devices with poor computing capabilities and computing The time for the terminal devices with better capability to complete the model training is roughly the same, thereby reducing the difference between the time for each terminal device to complete the model training.

S43. The network device sends the determined model training configuration information to the terminal device.

After the corresponding model training configuration information is determined according to the computing capability of the terminal device, the network device sends it to the corresponding terminal device, so that the terminal device can receive the model training configuration information sent by the network device.

S44. The terminal device trains the first machine learning model in the terminal device according to the model training configuration information to obtain model update parameters.

As mentioned above, the machine learning model for local training in each terminal device is pre-distributed by the network device. In this embodiment of the present application, the local machine learning model of the terminal device is called the first machine learning model, and the local The machine learning model is called the second machine learning model, so the first machine learning model is distributed by the network device to the terminal device. The first machine learning model and the second machine learning model are of the same type of machine learning model, or the first machine learning model and the second machine learning model are different types of machine learning models. The model update parameter information is used for local model update, and the first machine learning model and the second machine learning model are the same type of machine learning model.

After receiving the model training configuration information sent by the network device, the terminal device can locally train the first machine learning model in the terminal device according to the model training configuration information, and after the model training is completed, the corresponding model update parameters can be obtained , so the model update parameter here is the update parameter obtained after training the first machine learning model by using the model training configuration information. The model update parameters are model parameters of the trained first machine learning model.

In another possible implementation manner, in addition to sending the model training configuration information to the terminal device, the network device may also indicate to the terminal device the time to start local training by using the model training configuration information, for example, instructing the terminal device to start at a specific moment Local training, or instruct the terminal device to start local training after a predetermined period of time after receiving the model training configuration information, or can also instruct to start local training at other times (for example, 15:00:00), so that the network device can be stricter It controls the start time of each terminal device for local training, which can ensure that the time when each terminal device starts local training is as consistent as possible, and further reduces the time difference between each terminal device completing local training.

For the specific manner of model training, various training manners in the related art may be adopted, which is not limited in the embodiment of the present application.

S45, the terminal device sends the obtained model update parameters to the network device.

It should be noted that the model update parameters in the embodiments of the present application include the model update parameters themselves and their corresponding parameter values. For example, there are three model update parameters a, b, and c. The three model update parameters of the device sender are to send the three model update parameters a, b and c and the parameter values of various model update parameters to the network device. For example, the parameter value of a is 1.5, the parameter value of b is 2.6, The parameter value of c is 2.4.

Each terminal device uses the local training data to perform local training using the model training configuration configured by the network device according to its own computing capability. The types of model update parameters obtained by training are generally the same. For example, for terminal device 1, terminal device 2, and terminal device 3, after local training is performed using the model training configuration information configured by the network device respectively, the model update parameter a, the model update parameter b, and the model update parameter c are obtained. The parameter values of the three model update parameters obtained by terminal device 1 are 1.3, 1.8, and 2.4, respectively, and the parameter values of the three model update parameters obtained by terminal device 2 are 1.6, 1.4, 2.8, and those obtained by terminal device 3, respectively. The parameter values of the three model update parameters are 1.9, 1.3, and 2.7 respectively. It can be seen that the three terminal devices have obtained the same type of model update parameters after local training, but the parameter values corresponding to each model update parameter are different.

After obtaining the model update parameter corresponding to the model training configuration information, the terminal device sends it to the network device, and then the network device can receive the model update parameter sent by the terminal device.

In this embodiment of the present application, the model update parameter sent by the terminal device to the network device may include multiple types, for example, it may be a specific updated model parameter W _i ⁰ , or it may be an updated model parameter difference _gi =W _i ⁰ -W ⁰ , where i is any value from 1 to N, and the value of N is the total number of terminal devices for which the network device configures the model training configuration information, so W _i ⁰ indicates that the network configures it for Model update parameters obtained by the ith terminal device of the model training configuration information, _gi represents the model update parameter difference obtained by the ith terminal device for which the network device configures the model training configuration information.

In the first embodiment, the terminal device actively sends the model update parameters to the network device immediately after obtaining the model update parameters, so that the network device can obtain the model update parameters fed back by the terminal device as soon as possible. In this way, the network device can use the model training configuration information to try to make the time spent by each terminal device for local training to be consistent, so that after the local training of each terminal device is completed and the model update parameters are obtained, all terminal devices can be updated in time. It is reported to the network device, thereby reducing the time difference for the network device to obtain the model update parameters sent by each terminal device.

In the second embodiment, the terminal device does not actively send the model update parameters to the network device immediately after obtaining the model update parameters, but only sends the model update parameters to the network device when a specific trigger condition is met, which will be described below with an example.

Scenario 1: The terminal device sends the obtained model update parameters to the network device only after receiving the acquisition request sent by the network device to instruct the terminal device to send the model update parameters to the network device. That is to say, a possibility The trigger condition is that the terminal device receives the acquisition request sent by the network device, and under this trigger condition, the terminal device only sends the model update parameters to the network device when it actively requests. In this embodiment, the network device may first determine the time point for requesting the model update parameters from the terminal device, that is, it may first determine the time point for sending the aforementioned acquisition request to the terminal device, and then send the acquisition request to the terminal device at the determined time point. request to actively request the terminal device for the model update parameters obtained by the terminal device. In this way, the network device can explicitly control the time for requesting the model update parameters from each terminal device, and reduce the time for each terminal device to complete local training through the model training configuration information. On the basis of the difference, the time difference between each terminal device reporting the model update parameters can be further reduced, thereby reducing the time difference between the network device actually acquiring the model update parameters sent by each terminal device.

Scenario 2: The network device may directly indicate to the terminal device the time point for reporting the model update parameters. Specifically, the network device may send the terminal device the reporting time information used to instruct the terminal device to send the model update parameter to the network device. After the terminal device receives the report time information, it can send the model update parameters to the network device at the time point indicated by the report time information. That is to say, another possible trigger condition is that the network device sends a The time indicated by the reporting time information. Under this trigger condition, the terminal device regularly reports the model update parameters to the network device according to the instruction of the network device. In this embodiment, the network device can explicitly control the specific time at which each terminal device reports the model update parameters. On the basis of reducing the time difference between each terminal device completing local training through the model training configuration information, the reporting by each terminal device can be further reduced. The time difference between the model update parameters is reduced, thereby reducing the time difference between the network device actually obtaining the model update parameters sent by each terminal device.

Through the model training configuration information corresponding to each terminal device, the network device can determine the time point at which each terminal device completes its own local training, and on this basis, can directly determine the time point in the above scenario 1 and the reporting time in the scenario 2. The time point indicated by the information, for example, if the network device determines that each terminal device completes local training at about 16:05:00, it can send an acquisition request to each terminal device at 16:06:00, and can instruct each terminal device to 16:06:30 Send the model update parameters to the network device, which not only reduces the time difference between each terminal device reporting their own model update parameters, but also obtains the model update parameters as soon as possible, which is the basis for improving the efficiency of local model update. , you can try to update the model in advance to improve the timeliness of the model update. For the above Scenario 1 and Scenario 2, in another implementation manner, the network device may first determine the transmission duration for each terminal device in most or all of the terminal devices participating in the local training to send their respective model update parameters to the network device, Then, according to each transmission duration corresponding to each terminal device, the acquisition time for acquiring the model update parameters of the terminal device is determined, that is, the aforementioned time point for sending the acquisition request is determined.

Among them, the transmission duration can be understood as the interval between sending the model update parameter from the terminal device to the network device receiving the model update parameter, which is related to the quality of the communication link between each terminal device and the network device. In a possible implementation manner, the network device may obtain the uplink transmission rate of the terminal device according to the channel quality indicator (CQI) sent by the terminal device, and then determine the terminal device according to the data volume and the uplink transmission rate of the model update parameter. The transmission duration corresponding to the device, for example, w represents the uplink transmission rate of the terminal device, and q represents the data volume of the model update parameter of the terminal device, then the corresponding transmission duration of the terminal device T=q/w. According to this method, the transmission duration corresponding to each terminal device can be determined, and then the time point at which the acquisition request is sent to the terminal device can be determined according to the transmission duration corresponding to most terminal devices (for example, 80%) or all terminal devices, or each terminal device can be determined. The point in time when the device sends the model update parameters to the network device.

For the above-mentioned calculation formula of the transmission duration T=q/w, q represents the data amount of the model update parameter sent by the corresponding terminal device to the network device. Since the machine learning model for local training in each terminal device is uniformly distributed in advance by the network device, the model update parameters obtained after the training of each terminal device is known to the network device, and the network device can know each Data volume of model update parameters. In one embodiment, the terminal device transmits all model update parameters to the network device, then the network device can estimate the total data of all model update parameters on the basis of knowing the data volume of each model update parameter In another embodiment, the network device can request the terminal device for a specified type of model update parameter, and on the basis of knowing the data amount of each specified model update parameter, the network device can estimate all The total data volume of the specified model update parameters, obtain the aforementioned q.

That is to say, the network device can actively request model update parameters from each terminal device, and at the time when it matches each terminal device, it sends an acquisition request for requesting model update parameters to each corresponding terminal device or indicates that the terminal device is matching The model update parameters are reported to the network device at the same time point. In this way, the time required for the transmission of model update parameters of most or even all participants can be comprehensively considered, and the time for each terminal device to report its own model update parameters can be more accurately controlled. , which reduces the time difference between each terminal device reporting the model update parameters to the network device, thereby reducing the time difference between the network device acquiring the model update parameters sent by each terminal device, thereby improving the convergence speed of the local model update and improving the model update efficiency.

For example, the transmission duration corresponding to terminal device 1 is 10 minutes, the transmission duration corresponding to terminal device 2 is 15 minutes and hours, the transmission duration corresponding to terminal device 3 is 22 minutes, and the transmission duration corresponding to terminal device 4 is 28 minutes, then the network device You can send an acquisition request to terminal device 4 at 10:00, send an acquisition request to terminal device 3 at 10:06, send an acquisition request to terminal device 2 at 10:13, and send an acquisition request to terminal device 2 at 10:18. Terminal device 1 sends an acquisition request. That is to say, the acquisition request can be sent to the terminal device with the longer transmission duration earlier, and the acquisition request can be sent to the terminal device with the shorter transmission duration later. In this way, the terminal device with the longer transmission duration can receive the acquisition request earlier. Start sending the model update parameters to the network device as soon as possible after the request. Similarly, the terminal device with longer transmission duration can be instructed to send the model update parameter earlier, and the terminal device with shorter transmission duration can be instructed to send the model update parameter later. way, by starting to send model update parameters in advance to make up for the longer transmission time, reducing the time difference of each terminal device's transmission model update parameters, so that the network device can receive the data of each terminal device within the same (or as much as possible the same) time. Model update parameters, thereby reducing the time difference for the network device to receive the model update parameters of each terminal device.

In the embodiment of the present application, generally speaking, when each terminal device performs local model training by means of FL, the types of model update parameters obtained by each terminal device are the same, and the network device is also aware of it, because each terminal device The initial model of the locally trained machine learning model is distributed to each end device by the network device. The model update parameter obtained by the terminal device performing local training through the model training configuration information sent by the network device, the terminal device can send parameter availability indication information to the network device, and the parameter availability indication information can indicate the availability of the model update parameter in the terminal device, After the terminal device completes the local training of the machine learning model, the model update parameters of the terminal device are all available. Therefore, the parameter availability indication information can also indicate that the terminal device has completed the training of the local machine learning model. In other words, the terminal The device can inform the network device that it has completed the local training of the machine learning model through the parameter availability indication information. For example, the parameter availability indication information may be carried in the RRC reestablishment complete (RRC reestablishment complete) message, the RRC reconfiguration complete (RRC reconfiguration complete) message, the RRC resume complete (RRC resume complete) message, and the RRC establishment complete message sent by the terminal device to the network device. Completion (RRC setup complete) message, UE information response (UE information response) message, non-access stratum (non-access stratum, NAS) message and other messages, that is, the terminal device can pass any of the aforementioned A message informs the network device of the availability of model update parameters in the terminal device.

For the above-mentioned first embodiment, that is, for the way in which the terminal device actively reports the model update parameters, through the indication of the parameter availability indication information, the network device can specify all the model update parameters available in the terminal device, and then report the model update parameters reported by the terminal device. By comparing the update parameters, it can be determined whether the terminal device has missed some model update parameters or whether the model update parameters are incompletely obtained due to abnormal transmission during the transmission process, thereby improving the integrity and accuracy of acquiring model update parameters.

For the above-mentioned second embodiment, according to the indication of the parameter availability indication information, when the network device requests model update parameters from the terminal device through the acquisition request, it can obtain the specified type of model update parameters according to the actual needs of the network device, so the optional , the aforementioned acquisition request can also be used to indicate that the network device needs to acquire the specified model update parameters. For example, the model update parameters a, b, c, and d in the terminal device are all available, and the network device only requests the model update parameter a. b, c, in this way, the model update parameters transmitted by the terminal device can be reduced, the time spent by the terminal device in transmitting the model update parameters can be reduced, invalid transmission can be reduced, network transmission resources can be saved, and the overhead of air interface resources can be reduced. In another implementation manner, the specified model update parameters required by the network device may be indicated to the terminal device in other manners, for example, the specified model update parameters may be indicated by a message other than the acquisition request.

S46. The network device updates the machine learning model in the network device according to the model update parameter.

As described in Figure 1 above, the network device is configured with corresponding model training configuration information to multiple terminal devices, so the network device can receive the model update parameters sent by the above-mentioned terminal devices, as well as other terminal devices. other model update parameters, and by configuring the model training configuration information for each terminal device through the computing power, the difference in time required for the network device to receive model update parameters fed back by multiple terminal devices can be minimized.

For example, the network device configures the first model training configuration parameters and the second model training configuration parameters for the terminal device 1, the terminal device 2, and the terminal device 3 according to the respective computing capabilities of the terminal device 1, the terminal device 2, and the terminal device 3, respectively. , the third model training configuration parameter, the terminal device 1 obtains the first model update parameter after performing local training on the machine learning model in the terminal device 1 according to the first model training configuration information, and the terminal device 2 according to the second model training configuration information to the terminal The machine learning model in device 2 obtains second model update parameters after local training, and terminal device 3 obtains third model update parameters after locally training the machine learning model in terminal device 3 according to the third model training configuration information. Since the model training configuration information for each terminal device to perform local training is allocated according to the network device according to their respective computing capabilities, the time for the terminal device 1, the terminal device 2, and the terminal device 3 to complete the local training can be roughly the same, and further , each terminal device sends the model update parameters obtained by each terminal device to the network device after the training is completed, and the network device can receive the model update parameters sent by each terminal device in roughly the same time, thereby reducing the number of network devices receiving multiple Variation in the time it takes for the model to update the parameters of the end device.

Further, after obtaining the model update parameters fed back by each participant (that is, each terminal device to which the network device sends the model training configuration information), the network device can update all models according to the method described in the corresponding embodiment of FIG. 1 . The parameters are aggregated, and then the local machine learning model of the network device is updated with the model update parameters after the aggregation process, that is, the local machine learning model in the network device is updated locally. Because the method of configuring model training parameters through computing power can reduce the difference in the time it takes for network devices to receive model update parameters from multiple terminal devices, network devices can quickly converge when updating the local machine learning model, improving model updates. effectiveness.

It should be noted that the network device uses the obtained model update parameters to update the parameters of the local machine learning model, the training data is local to each terminal device, and each terminal device does not need to transmit training data to the network device, so the FL method can be used to update the parameters of the local machine learning model. Ensure user data security and avoid leakage of user privacy. In addition, the amount of data for updating parameters of the transmission model is generally much smaller than the training data, which can reduce network transmission overhead to a large extent, thereby saving network transmission resources.

As mentioned above, the model training configuration information is the information used by the terminal device to train the local machine learning model. It can be understood that the model training configuration information is the information instructing the terminal device how to train the local model. The model training configuration information in this embodiment of the present application may include one type of information or a combination of multiple types of information. For example, the model training configuration information includes one of hyperparameters, accuracy, and training time information. information or a combination of information. In the embodiment of the present application, the model training configuration information configured by the network device for the terminal device is routinely used by the terminal device for model training, which generally meets the configuration requirements of most terminal devices for local model training, and has good versatility.

For ease of understanding, the following describes the case where the model training configuration information includes different information.

first case

Model training configuration information is a hyperparameter. That is, the network device selects appropriate hyperparameters for the terminal device according to the computing capability of the terminal device, and then sends the selected hyperparameters to the terminal device, and the terminal device performs local training on the local machine learning model according to the hyperparameters, and then sends The obtained model update parameters are fed back to the network device.

Machine learning models involve two basic concepts, one is parameters and the other is hyperparameters. Among them, parameters are variables obtained by the model through learning, such as weight w and bias b; hyperparameters are set based on experience and affect the size of model parameters (such as weight w and bias b). A parameter is a parameter that sets a value before starting the learning process, not the parameter data obtained through model training. In layman's terms, a hyperparameter is also a parameter. It has the characteristics of a parameter, but it is not obtained through learning. For example, it can be obtained by The user specifies its value based on existing or existing experience. That is to say, hyperparameters are parameters that can affect model parameters. Therefore, the value set by hyperparameters can directly affect the effect of model training. Therefore, setting appropriate hyperparameters for the terminal device according to the computing power of the terminal device can control the terminal device as much as possible. Time spent training locally.

The hyperparameters in the embodiments of the present application may include at least one of a learning rate (learning rate), a batch size (batch size), an iteration (iteration), and a training epoch (epoch), that is, may include the aforementioned specific one or more of the hyperparameters. in:

Learning rate: When the machine learning model is updated, for example, many random decision trees will be generated. The weight of each decision tree is different. The learning rate determines the magnitude of the weight update. If the learning rate is too small, the convergence rate of the machine learning model will be slow and it will take longer to train. For example, the network device sets a certain threshold (for example, called the first threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the first threshold, select a smaller threshold for the terminal device. Learning rate, such as 0.0001; if the computing capability of the terminal device is less than the first threshold, a larger learning rate, such as 0.01, is selected for the terminal device.

Batch size: refers to the number of samples fed into the machine learning model during each training, that is, the number of samples required for one training. For example, if 100 samples are used in one training, the batch size is 100. The network device sets a certain threshold (for example, called the second threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the second threshold, select a larger batch size for the terminal device. , such as 128; if the computing capability of the terminal device is less than the second threshold, select a smaller batch size for the terminal device, such as 16.

The number of iterations: refers to the number of times the entire training set is input to the machine learning model for training. Implementation 1: The network device sets a certain threshold (for example, called the third threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the third threshold, select a higher threshold for the terminal device. A large iteration, such as 10000; if the computing capability of the terminal device is less than the third threshold, a smaller iteration, such as 1000, is selected for the terminal device. Implementation mode 2: The network device calculates the iteration of the terminal device according to the computing capability required by the machine learning model and the computing capability of the terminal device. For example, the computing capability required for one iteration of the machine learning model is M, and the computing capability of the terminal device is M. is P, then the iteration of the terminal device is N=P/M.

Number of training rounds: refers to the number of rounds in which the entire training set is input to the machine learning model for training. Implementation 1: The network device sets a certain threshold (for example, called the first threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the fourth threshold, select a higher threshold for the terminal device. A large epoch, such as 10; if the computing capability of the terminal device is less than the fourth threshold, a smaller epoch, such as 5, is selected for the terminal device. Implementation mode 2: The network device calculates the epoch of the terminal device according to the computing capability required by the machine learning model and the computing capability of the terminal device. For example, the computing capability required for one iteration of the machine learning model is M, and the computing capability of the terminal device is P, then the iteration of the terminal device is N=P/M.

For example, to illustrate the relationship between batch size, number of iterations, and number of training rounds: Assuming that the entire training set has 1000 samples, the number of training rounds is 10, and the batch size is 20, then the number of iterations = 10* (1000/20)=500.

Due to the different computing capabilities of each terminal device, hyperparameters are the basic training requirements for model training, and through different values of different types of hyperparameters, they can be quantified into the corresponding training time, for example, for 1000 The number of samples, the number of training rounds is 10, and the batch size is 20 for local training. According to the computing power of terminal device 1, it takes about 2 minutes to complete the training, and according to the computing power of terminal device 2, it only takes about 1.5 minutes to complete the training. Completing the training, that is, for the aforementioned hyperparameters with specific values, the training time of 2 minutes can be quantified for terminal device 1, and the training time of 1.5 minutes can be quantified for terminal device 2. Therefore, combined with the terminal device itself By configuring hyperparameters, the computing power of the terminal device can be more clearly defined when the terminal device completes the local training, so this can better control the training time of each terminal device participating in the local training, thereby reducing the time difference for each terminal device to complete the local training. .

second case

The model training configuration information is the accuracy required to train the machine learning model. That is, the network device selects the appropriate accuracy for the terminal device according to the computing capability of the terminal device, and then informs the terminal device of the selected accuracy. The terminal device performs local training on the local machine learning model according to the accuracy required by the network device, and then obtains The model update parameters are fed back to the network device. Among them, the accuracy required when training the machine learning model refers to the difference between the actual predicted output of the machine learning model and the actual output of the sample.

The precision in this embodiment of the present application may include at least one of an error rate, a correct rate, a deviation rate, and a recall rate. in:

Error rate: refers to the ratio of the number of samples with classification errors (or prediction errors) to the total number of samples based on the updated machine learning model (that is, the trained machine learning model). The network device sets a certain threshold (for example, called the fifth threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the fifth threshold, a smaller error rate is selected for the terminal device. ; if the computing capability of the terminal device is less than the fifth threshold, select a larger error rate for the terminal device.

Correct rate: refers to the proportion of the number of correctly classified (or correctly predicted) samples to the total number of samples based on the updated machine learning model (that is, the trained machine learning model). The network device sets a certain threshold (for example, called the sixth threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the sixth threshold, select a higher correct rate for the terminal device. ; if the computing capability of the terminal device is less than the sixth threshold, select a smaller correct rate for the terminal device. For example, for a neural network, the machine learning model produces a probability prediction for each test sample. In this case, the accuracy rate can be the Top-1 accuracy rate, that is, the accuracy rate of the category with the first probability prediction being consistent with the actual result; or, The accuracy rate can be the Top-5 accuracy rate, that is, the accuracy rate of the probability prediction that the top five categories contain the actual results.

Precision: It refers to how many of the samples predicted to be positive based on the updated machine learning model are true positive samples. The network device sets a certain threshold (for example, referred to as the seventh threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the seventh threshold, select a larger threshold for the terminal device. rate; if the computing capability of the terminal device is less than the seventh threshold, a smaller precision rate is selected for the terminal device.

Recall rate: refers to how many positive examples in the sample are predicted correctly based on the updated machine learning model. The network device sets a certain threshold (for example, referred to as the eighth threshold) according to the computing power required by the machine learning model. If the computing power of the terminal device is greater than or equal to the eighth threshold, the lower retrieval value is selected for the terminal device. rate; if the computing capability of the terminal device is less than the eighth threshold, a higher recall rate is selected for the terminal device.

The third situation

The model training configuration information is the training time information required when training the machine learning model, and the training time information is used to indicate the time used for training. The training duration, either indicates the training end time and the training duration, or only indicates the training duration (eg, 5 minutes or 10 minutes, etc.). That is, the network device selects appropriate training time information for the terminal device according to the computing capability of the terminal device, and then informs the terminal device of the selected training time information, and the terminal device performs the local machine learning model according to the training time information required by the network device. Local training, and then feedback the obtained model update parameters to the network device.

In the specific implementation process, for multiple or all terminal devices that need to participate in local training, the network device can calculate the time required for each terminal device to perform local training according to the computing capability of each terminal device, and then use most (for example, 90% of the time) ) or the maximum value of the training time required by all terminal devices as the training time of each participant. In this way, a longer training time can be configured for each terminal device, so that each terminal device can complete the local training as much as possible within the specified training time, and the time it takes for most terminal devices to complete the local training is approximately the same, so It can reduce the time difference of each terminal device completing local training, thereby reducing the time difference of the network device acquiring the model update parameters fed back by each terminal device.

fourth situation

The model training configuration information is the hyperparameters and accuracy required to train the machine learning model. In the specific implementation process, for the hyperparameters and accuracy set by the network device, the terminal device may not be able to absolutely meet these two requirements at the same time during local training. For example, the hyperparameter configured by the network device for the terminal device is the batch size. 50 and the number of training rounds is 10, the accuracy of the configuration is 96%, so when the terminal device is training, when 10 rounds of training are performed according to the batch size of 50, the accuracy may not reach 96%. In this case, one approach is that the terminal device can configure other hyperparameters by itself to achieve the 96% accuracy requirement. For example, in addition to the aforementioned batch size of 50 and the number of training rounds configured by the network device, it can also be configured by itself. A larger learning rate and a larger number of iterations are used for local training to try to meet the accuracy requirements of network devices; another approach is to find a balance in the training configuration information of the two models configured on the network side, such as in On the basis of the hyperparameters and precision configured by the network equipment, some values of the hyperparameters can be appropriately increased and the precision can be reduced, but the adjustment range should not be too large, so as to meet the training requirements of the network equipment as much as possible.

That is to say, when the terminal device performs local training, it may not absolutely meet the requirements of the model training configuration information configured by the network device. In a possible implementation manner, the model training configuration information configured by the network device may be Appropriate adjustments are made in a small range, so as to meet the requirements of multiple model training configuration information at the same time, so as not to affect the network equipment through the model training configuration information to limit the time difference for each terminal device to complete local training, and at the same time to obtain better results. training effect.

Fifth situation

The model training configuration information is the hyperparameter, accuracy and training time information required when training the machine learning model.

That is to say, the network device can simultaneously configure three (or more) types of model training configuration information for the terminal device. In this case, when the terminal device performs local training, it can refer to the methods listed in the fourth situation to To perform local training, that is, you can perform local training not strictly according to the configuration of the network equipment, but to adjust the training configuration information of one or more models in an appropriate and small range, in order to achieve a better training effect and at the same time reduce the training cost. Changes in time should be as small as possible.

In another embodiment, considering that the purpose of training configuration information for various models configured by the network device according to the computing capability of the terminal device is to reduce the time difference between the completion of local training by each terminal device participating in the local training, the network The training time information configured by the device for each terminal device is the most direct requirement to reduce the time difference. For this reason, when training configuration information of various types of models including training time information, priority can be given to ensure that the training time information remains unchanged. The training configuration information of other types of models is slightly adjusted in order to achieve a better training effect, or it is better to ensure that the training time information remains unchanged, and perform local training strictly according to the training configuration information of other types of models. That is to say, when training time information is included in the training configuration information of various models, the priority of the training time information is the highest, and the corresponding time requirements can be kept unchanged, so as to satisfy the training time of each terminal device by the network device as much as possible. requirements, so as to better reduce the time difference of local training performed by each terminal device.

The foregoing describes an embodiment in which the network device allocates corresponding model training configuration information to each terminal device participating in the local training according to the computing capability of the terminal device. On the basis of the above embodiment, the network device and the terminal device can also More interaction, in order to achieve better training effect, obtain more accurate model update parameters, and at the same time, it can better reduce the time difference of each terminal device completing local training, thereby reducing the model update that the network device obtains feedback from each terminal device. The time difference of the parameters facilitates the rapid convergence of the network device when updating the local machine learning model, and improves the update efficiency of the network device updating the machine learning model.

In one embodiment, on the basis that the network device configures the model training configuration information to the terminal device, the network device may also send training feature information to the terminal device, where the training feature information is used to instruct the terminal device to perform local training. For example, the training feature information includes: channel quality indicator (CQI), channel state information reference signal (channel interference information reference signals, CSI-RS) measurement result, synchronization signal and physical broadcast channel block (synchronization One or more of information such as signal and physical broadcast channel block, SSB) measurement results and packet delay. After receiving the training feature information, the terminal device can use the samples in the corresponding training features to locally train the local machine learning model, for example, use the samples in the SSB measurement result to locally train the machine learning model. The network device sends the training feature information to the terminal device, so that each participant can use the same training feature information for local training, thereby reducing the difference in the time spent by each terminal device for local training based on different training feature information.

In an embodiment, after the network device configures the model training configuration information to the terminal device, the network device may also send accuracy evaluation information to the terminal device, where the accuracy evaluation information is used by the terminal device for the locally trained machine learning model The accuracy is evaluated, and the accuracy evaluation information includes at least one of a method for evaluating the accuracy or a test sample for evaluating the accuracy. The method for evaluating the accuracy may be any one of hold-out, cross validation, bootstrapping or other methods.

The set aside method is to divide the samples into two mutually exclusive sets, one of which is used as the training sample of the machine learning model, and the other set is used as the test sample of the machine learning model. sample for testing.

The cross-validation method is to divide the sample into k mutually exclusive subsets of similar size, and then use the union of k-1 subsets as the training sample each time, and the remaining subset as the test sample, so as to train k times as soon as possible. Test, and finally return the mean of k test results.

The bootstrap method is to give a data set D of m samples, randomly select a sample from D each time, copy it to E, and then put the sample back into the initial data set D. This process is repeated m times to obtain a data set E containing m samples. The samples in data set E are used as training samples, and the samples in data set D that are different from data set E are used as test samples.

By specifying the accuracy evaluation information to the terminal device, each participant can use the same accuracy evaluation information to evaluate the accuracy of the locally trained machine learning model. Since the same accuracy evaluation method is used, each terminal device can be The specified accuracy requirements are met under the same accuracy evaluation standard, thereby reducing the difference in the time spent by each terminal device for local training.

In this embodiment of the present application, after completing the local training, the terminal device needs to feed back the obtained model update parameters to the network device. On this basis, the terminal device may also send accuracy indication information to the network device, where the accuracy indication information is used to indicate The accuracy achieved by the terminal device using the model training configuration information configured by the network device to perform the local training of the machine learning model. That is, the terminal device can not only feed back the model update parameters to the network device, but also the accuracy of the corresponding model training. Feedback to the network device, so that the network device can know the training effect of the terminal device, and can be used as a reference when the network device configures the model training configuration information for the terminal device in the future. For example, the accuracy indication information sent by the terminal device indicates the training accuracy. If it is poor, then when the subsequent network device selects the model training configuration information for the terminal device, it can perform directional adjustment on the basis of the previous model training configuration information.

Further, the role of the model update parameters fed back by the terminal device in the local model update can be determined according to the judgment of its training effect. For example, the accuracy indication information fed back by the terminal device 1 indicates that the accuracy of its local training is 97%, while The accuracy indication information fed back by the terminal device 2 indicates that the accuracy of its local training is 85%. It can be seen that the accuracy of the local training of the terminal device 1 is higher than that of the local training of the terminal device 2. In other words, the training effect of the terminal device 1 should be better than that of the terminal device. device 2. Further, when using the model update parameters fed back by each terminal device to perform local update, the network device may give a larger weight to the model update parameter fed back by the terminal device 1, while giving the model update parameter fed back by the terminal device 2 a relatively small weight. . In this way, the effectiveness and accuracy of the local update of the model by the network device can be improved.

The method in which the network device configures the model training parameter information for each terminal device to perform local training according to the computing capability of the terminal device has been introduced above. Each terminal device can try to complete the training of the machine learning model in the mainland within the same (or approximately the same) time, reducing the time difference between each terminal device sending model update parameters to the network device, thereby reducing the network device receiving the data sent by each terminal device. Due to the time difference of the model update parameters, the network device can try to use the model update parameters fed back by each terminal device to locally update the local machine learning model in a short period of time, thereby improving the convergence speed of the local update and improving the update of the machine learning model. effectiveness.

In the embodiment of the present application, another method for updating a machine learning model is also provided. In this method, the network device actively requests each terminal device for the model update parameters in each terminal device, and the time point of the request is the network device according to each terminal device. The amount of data that the terminal device actually needs to transmit and the condition of the transmission chain (the quality of the transmission link) are determined. Specifically, the network device selects the time point for requesting the model update parameters from each terminal device according to the time required for each terminal device to send the respective model update parameters to the network device (for example, the duration is referred to as the transmission duration). The difference between the transmission time and the different time points to request the respective model update parameters from each terminal device, which can minimize the difference between the network device receiving the model update parameters sent by each terminal device due to the difference in transmission time. The time difference between each terminal device allows the model update parameters sent by each terminal device to reach the network device at the same time (or approximately the same short time), thereby reducing the time difference between the network device acquiring the model update parameters of each terminal device. In this way, the network device can locally update the local machine learning model according to the model update parameters of each terminal device in a short time, thereby improving the convergence speed of the local update, thereby improving the updating efficiency of the machine learning model.

For ease of understanding, another method for updating a machine learning model provided by an embodiment of the present application is described below with reference to FIG. 5 . In the introduction of FIG. 5 , a first terminal device is used as an example for description, wherein the first terminal device is a participant. Any one of the multiple terminal devices of FL.

S51: The first terminal device sends parameter availability indication information to the network device.

As mentioned above, the parameter availability indication information can be used to indicate the availability of the model update parameters in the first terminal device, because after the terminal device completes the local training of the machine learning model, the model update parameters of the terminal device are all available, Therefore, the parameter availability indication information can also indicate that the terminal device has completed the training of the local machine learning model. In other words, the first terminal device can inform the network device that it has completed the local training of the machine learning model through the parameter availability indication information. event. In a specific implementation process, the parameter availability indication information may be carried in the RRC re-establishment complete message, the RRC reconfiguration complete message, the RRC recovery complete message, the RRC establishment complete message, the UE information response message, the NAS In any of the messages such as messages, that is, the first terminal device may notify the network device of the availability of the model update parameters in the first terminal device through any of the aforementioned messages.

In the specific implementation process, S51 is not a necessary step, so S51 is represented by a dotted line in FIG. 5 , that is to say, the first terminal device can send parameter availability indication information to the network device, or it does not need to send the parameter availability indication to the network device. information, which is not limited in the embodiments of the present application.

S52: The network device determines the transmission duration for each terminal device in the plurality of terminal devices to send their respective model update parameters to the network device.

The plurality of terminal devices may include the first terminal device, or may not include the first terminal device. Each of the multiple terminal devices is a terminal device for which the network device pre-distributed the initial machine learning model, and the multiple terminal devices participate in the local training of the machine learning model distributed by the network device. Most (eg 80%) or the vast majority (eg 95%) of all terminal devices of

S53: The network device selects a time point for acquiring the model update parameters of the first terminal device according to the transmission duration of each terminal device sending the respective model update parameters to the network device.

In the embodiment of the present application, before the network device requests each terminal device (including the first terminal device) for the model update parameters corresponding to each terminal device, each terminal device has completed the training of its own local machine learning model and obtained the corresponding corresponding In this way, the network device can determine the time point for acquiring the model update parameters in each terminal device according to the transmission duration of each terminal device sending the respective model update parameters to the network device. For example, the selected time point is called acquisition. time, the acquisition time may be the time when the network device sends the acquisition request for requesting the acquisition of model update parameters to the first terminal device, that is, the network device may send the acquisition request to the first terminal device at the acquisition time point; or, the acquisition time It may also be the time indicated by the network device to the first terminal device to send the model update parameters to the network device, that is, the first terminal device may send its local model update parameters to the network device at the acquisition time.

The transmission duration can be understood as the interval from each terminal device sending its own model update parameter to the network device receiving the model update parameter, which is related to the quality of the communication link between each terminal device and the network device , so in a possible implementation manner, the network device can obtain the uplink transmission rate of the terminal device according to the CQI sent by the terminal device, and then determine the transmission duration corresponding to the terminal device according to the data volume of the model update parameter and the uplink transmission rate, For example, w represents the uplink transmission rate of the terminal device, and q represents the data volume of the model update parameter of the terminal device, then corresponding to the transmission duration of the terminal device T=q/w. According to this method, the transmission duration corresponding to each terminal device can be determined, and then the time point for acquiring the model update parameters in the terminal device is determined according to the transmission duration corresponding to most terminal devices (eg, 80%) or all terminal devices.

For example, the transmission duration corresponding to terminal device 1 is 10 minutes, the transmission duration corresponding to terminal device 2 is 15 minutes, the transmission duration corresponding to terminal device 3 is 22 minutes, and the transmission duration corresponding to terminal device 4 is 28 minutes, then the network device can Instruct the terminal device 4 to send the model update parameters to the network device at 13:02, instruct the terminal device 3 to send the model update parameters to the network device at 13:08, and instruct the terminal device 2 to send the network device at 13:15. At 13:08, instruct the terminal device 1 to send the model update parameters to the network device at 13:20. That is to say, the terminal device with the longer transmission duration can be instructed to send the model update parameters in advance, and the terminal device with the shorter transmission duration can be instructed to send the model update later. request, and send the acquisition request to the terminal device with the shorter transmission time later. In this way, the long transmission time is compensated for by starting to send the model update parameters in advance, and the time difference of the transmission model update parameters of each terminal device is reduced, so that the The network device can receive the model update parameters of each terminal device within the same (or as much as possible the same) time, thereby reducing the time difference for the network device to receive the model update parameters of each terminal device.

In this way, the time required for the transmission model update parameters of most or even all participants can be comprehensively considered, and the time for each terminal device to report its own model update parameters can be more accurately controlled, thereby reducing the network device's acquisition of the data sent by each terminal device. The time difference of the model update parameters, thereby improving the convergence speed of the local model update and improving the model update efficiency.

S54: At the selected time point, the network device sends an acquisition request to the first terminal device.

According to the methods introduced in S52 to S53, the network device determines the time point for acquiring the model update parameters of the first terminal device, and may send an acquisition request to the first terminal device at the determined time point, where the acquisition request is used to instruct the first terminal device The model update parameters in the first terminal device are sent to the network device.

In the specific implementation process, the network device does not explicitly indicate to the first terminal device which model update parameters need to be acquired, then according to the default agreement between the network device and the first terminal device, the first terminal device can obtain all the model update parameters obtained by the first terminal device. are sent to the network device. In another possible implementation manner, the network device may select the required partial model update parameters from the model update parameters available in the first terminal device, then in this manner, the acquisition request may also be used to indicate the network device specified Model update parameters, which means that the network device only needs to obtain the model update parameters indicated by the request, so that the first terminal device only needs to feed back the specified model update parameters requested by the network device to the network device, which can reduce the amount of data transmission, Reduce network transmission overhead.

S55: The first terminal device determines, according to the acquisition request, model update parameters that need to be sent to the network device.

In the above two cases, the model update parameters determined by the first terminal device according to the acquisition request may be all model update parameters or part of the model update parameters in the first terminal device.

S56: The first terminal device sends the determined model update parameters to the network device.

The above S51 to S56 (or S51 may not be included) show that the network device determines a time point according to the transmission duration of each terminal device, and sends an acquisition request to the first terminal device at the determined time point to request model update from the first terminal device Examples of parameters. In this embodiment, the network device can explicitly control the time for requesting model update parameters from each terminal device, and on the basis of reducing the time difference between each terminal device completing local training through the model training configuration information, it can further reduce the reporting of each terminal device. The time difference between the model update parameters is reduced, thereby reducing the time difference between the network device actually obtaining the model update parameters sent by each terminal device.

S57: The network device sends reporting time information to the first terminal device.

The reporting time information is used to instruct the first terminal device to send the model update parameters in the first terminal device to the network device at the acquisition time determined by the network device, that is, the network device can explicitly indicate each terminal device to each terminal device The specific time for reporting model update parameters to the network device.

S58: The first terminal device sends the model update parameter to the network device at the acquisition time indicated by the reporting time information.

According to the indication of the reporting time information, when the acquisition time indicated by the reporting time information arrives, the first terminal device sends the model update parameters obtained through local training to the network device.

The above S51, S52, S53, S57, S58 (may not include S51) show that the network device determines a time point according to the transmission duration of each terminal device, and instructs the first terminal device to report the model update parameter to the network device at this time point. example. In this embodiment, the network device can explicitly control the specific time at which each terminal device reports the model update parameters. On the basis of reducing the time difference between each terminal device completing local training through the model training configuration information, the reporting by each terminal device can be further reduced. The time difference between the model update parameters is reduced, thereby reducing the time difference between the network device actually obtaining the model update parameters sent by each terminal device. It should be noted that, in the specific implementation process, the processes shown in the above S51 to S56 and the processes shown in S51, S52, S53, S57, and S58 may be alternatively implemented, which is not limited in the embodiment of the present application. In FIG. 5 , the process corresponding to S51 to S56 is implemented as an example, so in FIG. 5 , the steps corresponding to S57 and S58 are represented by dotted lines, indicating that the steps may not be executed.

S59: The network device updates the local machine learning model according to the model update parameter sent by the first terminal device.

The above only takes the first terminal device as an example to introduce the implementation manner in which the network device obtains the model update parameters in one terminal device. According to the method described above, the network device can obtain the model update parameters in other terminal devices, and because it is based on each The transmission duration of the terminal device transmission model update parameters is differentiated and the acquisition request is sent to each terminal device, so the network device can receive the respective model update parameters sent by each terminal device at almost the same time, reducing the network device receiving each terminal device. Time variance of model update parameters sent by the device. Further, the network device can use the model update parameters of all terminal devices to locally update the local machine learning model in a short period of time, so that the update can be quickly converged, thereby improving the update efficiency of the machine learning model.

In all the above-mentioned embodiments of the method for updating a machine learning model, for example, whether in the introduction of the embodiment of FIG. 4 or the introduction of the embodiment of FIG. 5, the terminal device and the network device may send messages based on the existing protocol stack. For the related information, for example, the related information is sent between the terminal device and the access network device based on the RRC message, or the related information is sent between the terminal device and the core network device based on the NAS message.

In addition, when the network device is an access network device and under the CU-DU architecture, the information exchanged between the terminal device and the CU can be forwarded through the DU, that is, the terminal device first sends the information that needs to be transmitted to the CU to the DU, and then the DU The information is then forwarded to the CU based on the F1 interface with the CU. And, the information exchanged between the terminal device and the DU can be directly sent to each other, that is, the information that the terminal device needs to send to the DU can be directly sent to the DU, and the information that the DU needs to send to the terminal device can also be directly sent to the terminal device.

Taking Figure 4 as an example, assuming that the network device in Figure 4 is a CU, then S41, S42, and S46 are executed by the CU, S44 is executed by the terminal device, and the model training configuration information in S43 is first sent by the CU to the DU. The DU is then forwarded to the terminal device, and the model update parameters in S45 are first sent by the terminal device to the DU and then forwarded by the DU to the CU; assuming that the network device in Figure 4 is a DU, then S41, S42, S46 Performed by the DU, S44 is performed by the terminal device, the model training configuration information in S43 is directly sent by the DU to the terminal device, and the model update parameters in S45 are directly sent by the terminal device to the DU.

Taking Fig. 5 as an example again, assuming that the network device in Fig. 5 is a CU, then S52, S53, and S59 are executed by the CU, S55 is executed by the terminal device, the parameter availability indication information in S51, and the model update that needs to be sent in S56. The parameters and the model update parameters sent at the time point indicated by the reporting time information in S58 are first sent by the terminal device to the DU and then forwarded by the DU to the CU. The acquisition request in S54 and the reporting time in S57 are sent by the CU. It is first sent to the DU and then forwarded to the terminal device by the DU; then suppose that the network device in Figure 5 is a DU, then S52, S53, and S59 are performed by the DU, and S55 is performed by the terminal device. The parameter availability indication information in S51, The model update parameters that need to be sent in S56 and the model update parameters sent at the time point indicated by the reporting time information in S58 are directly sent by the terminal device to the DU, and the acquisition request in S54 and the reporting time in S57 are sent by the DU sent directly to the end device.

When the network devices in FIG. 4 and FIG. 5 are respectively CU or DU, the specific embodiments of the steps performed by the CU and the DU respectively can refer to the description of the embodiments in the aforementioned FIG. 4 and FIG. 5 , and the description will not be repeated here. .

Based on the same inventive concept, an embodiment of the present application provides a communication device, where the communication device may be a network device or a chip provided inside the network device. The communication apparatus has the function of implementing the network equipment in the embodiments shown in FIG. 4 to FIG. 5 . For example, the communication apparatus includes the functions corresponding to the steps performed by the network equipment in the embodiments shown in FIG. 4 to FIG. 5 . Modules or units or means, the functions or units or means can be implemented by software, or by hardware, or by executing corresponding software by hardware. For example, as shown in FIG. 6 , the communication apparatus in this embodiment of the present application includes a processing unit 601 and a communication unit 602, wherein:

A processing unit 601, configured to determine model training configuration information corresponding to the terminal device according to the computing capability of the terminal device;

The communication unit 602 is configured to send the model training configuration information to the terminal device, and receive the model update parameter sent by the terminal device, wherein the model update parameter is that the terminal device trains the first machine learning model according to the model training configuration information parameters updated later;

The processing unit 601 is further configured to update the second machine learning model according to the model update parameter.

In a possible implementation manner, the communication unit 602 is also used for:

In a possible implementation, the model training configuration information includes at least one of hyperparameters, precision, and training time information.

In a possible implementation manner, the communication unit 602 is further configured to send training feature information to the terminal device, where the training feature information is used to indicate the training feature set used by the terminal device to train the first machine learning model.

In a possible implementation manner, the communication unit 602 is further configured to send accuracy evaluation information to the terminal device, where the accuracy evaluation information includes at least one of a method for evaluating the accuracy or a test sample for evaluating the accuracy.

In a possible implementation manner, the communication unit 602 is further configured to receive accuracy indication information from the terminal device, where the accuracy indication information is used to instruct the terminal device to use the model training configuration information to train the first machine learning model. precision.

In a possible implementation manner, the processing unit 601 is further configured to determine a time point for acquiring the model update parameters of the terminal device; correspondingly, the communication unit 602 is further configured to:

Send an acquisition request to the terminal device at the aforementioned time point, where the acquisition request is used to instruct the terminal device to send the model update parameters of the terminal device to the network device; or,

Send reporting time information to the terminal device, where the reporting time information is used to indicate that the model update parameters are sent to the network device at the aforementioned time point.

In a possible implementation manner, the processing unit 601 is specifically configured to determine the transmission duration for each terminal device in the multiple terminal devices to send their respective model update parameters to the network device; and determine the aforementioned transmission duration according to each transmission duration corresponding to each terminal device point in time.

In a possible implementation manner, the acquisition request is further used to indicate that the specified model update parameters need to be acquired.

In a possible implementation manner, the communication unit 602 is further configured to receive parameter availability indication information from the terminal device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

Wherein, all relevant contents of the steps involved in the above method embodiments can be cited in the functional descriptions of the corresponding functional modules, which will not be repeated here.

Based on the same inventive concept, an embodiment of the present application provides a communication device, where the communication device may be a terminal device or a chip provided inside the terminal device. The communication device has the function of implementing the terminal device in the embodiment shown in FIG. 4, or the communication device has the function of implementing the first terminal device in the embodiment shown in FIG. 5. For example, the communication device includes executing the above-mentioned FIG. 4 ~ the modules or units or means corresponding to the steps performed by the terminal device or the first terminal device in the embodiment shown in FIG. 5 , the functions, units or means may be implemented by software, or by hardware, or by hardware Execute the corresponding software implementation. For example, as shown in FIG. 7 , the communication apparatus in this embodiment of the present application includes a communication unit 701 and a processing unit 702, wherein:

A communication unit 701, configured to receive model training configuration information sent by a network device, where the model training configuration information is determined according to the computing capability of the terminal device;

A processing unit 702, configured to train the first machine learning model according to the model training configuration information to obtain model update parameters;

The communication unit 701 is further configured to send the model update parameter to the network device, where the model update parameter is used by the network device to update the second machine learning model.

In a possible implementation manner, the communication unit 701 is further configured to receive a computing capability acquisition request sent by the network device; and send second computing power indication information to the network device according to the computing capability acquisition request, where the second computing power indication information Used to indicate the computing capability of the terminal device.

In a possible implementation manner, the communication unit 701 is further configured to receive training feature information from the network device, where the training feature information is used to indicate the training feature set used by the terminal device to train the first machine learning model; then the corresponding , the processing unit 702 is further configured to train the first machine learning model according to the model training configuration information and the training feature information.

In a possible implementation manner, the communication unit 701 is further configured to receive accuracy evaluation information from the network device, where the accuracy evaluation information includes at least one of a method for evaluating the accuracy or a test sample for evaluating the accuracy; then the corresponding The processing unit 702 is further configured to determine the accuracy achieved by the trained first machine learning model according to the accuracy evaluation information.

In a possible implementation manner, the communication unit 701 is configured to send accuracy indication information to the network device, where the accuracy indication information is used to indicate the accuracy achieved by the terminal device after training the first machine learning model by using the model training configuration information.

In a possible implementation manner, the communication unit 701 is further configured to receive an acquisition request from the network device, and send model update parameters to the network device according to the acquisition request, where the acquisition request is used to instruct the terminal device to send the network device the information of the terminal device to the network device. Model update parameters.

In a possible implementation manner, the communication unit 701 is further configured to receive reporting time information from the network device, and send model update parameters to the network device at the time point indicated by the reporting time information.

In a possible implementation manner, the communication unit 701 is further configured to send parameter availability indication information to the network device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.

Based on the same inventive concept, an embodiment of the present application provides a communication device, where the communication device may be a network device or a chip provided inside the network device. The communication apparatus has the function of implementing the network equipment in the embodiments shown in FIG. 4 to FIG. 5 . For example, the communication apparatus includes the functions corresponding to the steps performed by the network equipment in the embodiments shown in FIG. 4 to FIG. 5 . Modules or units or means, the functions or units or means can be implemented by software, or by hardware, or by executing corresponding software by hardware. For example, as shown in FIG. 8 , the communication apparatus in this embodiment of the present application includes a processing unit 801 and a communication unit 802, wherein:

A processing unit 801, configured to select a time point for acquiring the update parameters acquired by the first terminal device according to the transmission duration of each terminal device in the plurality of terminal devices sending their respective model update parameters to the network device;

A communication unit 802, configured to send an acquisition request to the first terminal device at the above-mentioned time point, and receive the model update parameter sent by the first terminal device, where the acquisition request is used to request the first terminal device to send the model update parameter to the network device; or is used to send reporting time information to the first terminal device and receive model update parameters sent by the first terminal device to the network device, where the reporting time information is used to instruct the first terminal device to send the model update parameters to the network device at the above-mentioned time point;

The processing unit 801 is further configured to update the second machine learning model according to the model update parameter.

In a possible implementation manner, the communication unit 802 is further configured to receive parameter availability indication information from the first terminal device.

In a possible implementation manner, the communication unit 802 is further configured to receive indication information from the network device for indicating the specified model update parameter.

Based on the same inventive concept, an embodiment of the present application provides a communication device, where the communication device may be a terminal device or a chip provided inside the terminal device. The communication device has the function of implementing the terminal device in the embodiment shown in FIG. 4, or the communication device has the function of implementing the first terminal device in the embodiment shown in FIG. 5. For example, the communication device includes executing the above-mentioned FIG. 4 ~ the modules or units or means corresponding to the steps performed by the terminal device or the first terminal device in the embodiment shown in FIG. 5 , the functions, units or means may be implemented by software, or by hardware, or by hardware Execute the corresponding software implementation. For example, as shown in FIG. 9 , the communication apparatus in this embodiment of the present application includes a communication unit 901 and a processing unit 902, wherein:

The communication unit 901 is configured to receive an acquisition request sent by a network device, wherein the time point at which the acquisition request is sent is determined by the network device according to the transmission duration of each terminal device in the plurality of terminal devices sending their respective model update parameters to the network device ; Or used to receive the reporting time information sent by the network device, the reporting time information is used to indicate that the model update parameters are sent to the network device at a time point;

a processing unit 902, configured to determine model update parameters to be sent according to the acquisition request;

The communication unit 901 is further configured to send the determined model update parameter to the network device, or to send the model update parameter to the network device at the time point indicated by the reporting time information, where the model update parameter is used by the network device to learn about the second machine The model is updated.

In a possible implementation manner, the communication unit 901 is further configured to send parameter availability indication information to the network device.

In a possible implementation manner, the communication unit 901 is further configured to receive indication information from the network device for indicating the specified model update parameter.

Based on the same inventive concept, referring to FIG. 10 , an embodiment of the present application further provides a communication device, including:

At least one processor 1001; and a communication interface 1003 communicatively connected to the at least one processor 1001; the at least one processor 1001 executes the instructions stored in the memory 1002 by executing the instructions stored in the memory 1002, so that the communication device executes the above-mentioned operations shown in FIG. 4 to FIG. 5 through the communication interface 1003 Method steps performed by a network device in an embodiment.

Optionally, the memory 1002 is located outside the communication device.

Optionally, the communication apparatus includes a memory 1002 , the memory 1002 is connected to the at least one processor 1001 , and the memory 1002 stores instructions that can be executed by the at least one processor 1001 . The memory 1002 is optional to the communication device as indicated by dashed lines in FIG. 10 .

Wherein, at least one of the processor 1001 and the memory 1002 may be coupled through an interface circuit, or may be integrated together, which is not limited here.

The specific connection medium between the processor 1001 , the memory 1002 , and the communication interface 1003 is not limited in the embodiments of the present application. In this embodiment of the present application, the processor 1001, the memory 1002, and the communication interface 1003 are connected through a bus 1004 in FIG. 10. The bus is represented by a thick line in FIG. 10, and the connection between other components is only for schematic illustration. , is not limited. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in FIG. 10, but it does not mean that there is only one bus or one type of bus.

Based on the same inventive concept, referring to FIG. 11 , an embodiment of the present application further provides a communication device, including:

At least one processor 1101; and a communication interface 1103 communicatively connected to at least one processor 1101; at least one processor 1101 executes the instructions stored in the memory 1102 by executing the instructions stored in the memory 1102, so that the communication device executes the above-mentioned embodiment shown in FIG. 4 through the communication interface 1103. The steps of the method executed by the terminal device shown in FIG. 5 may be executed, or the steps of the method executed by the first terminal device in the above-mentioned embodiment shown in FIG. 5 are executed.

Optionally, the memory 1102 is located outside the communication device.

Optionally, the communication apparatus includes a memory 1102 , the memory 1102 is connected to the at least one processor 1101 , and the memory 1102 stores instructions that can be executed by the at least one processor 1101 . Figure 11 shows in dashed lines that the memory 1102 is optional to the communication device.

The processor 1101 and the memory 1102 may be coupled through an interface circuit, or may be integrated together, which is not limited here.

The specific connection medium between the processor 1101 , the memory 1102 , and the communication interface 1103 is not limited in the embodiments of the present application. In the embodiment of the present application, the processor 1101, the memory 1102, and the communication interface 1103 are connected by a bus 1104 in FIG. 11. The bus is represented by a thick line in FIG. 11, and the connection between other components is only for schematic illustration. , is not limited. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in FIG. 11, but it does not mean that there is only one bus or one type of bus.

It should be understood that the processor mentioned in the embodiments of the present application may be implemented by hardware or software. When implemented in hardware, the processor may be a logic circuit, an integrated circuit, or the like. When implemented in software, the processor may be a general-purpose processor implemented by reading software codes stored in memory.

Exemplarily, the processor may be a central processing unit (central processing unit, CPU), or other general-purpose processors, digital signal processors (digital signal processors, DSP), application specific integrated circuits (application specific integrated circuit, ASIC) , off-the-shelf programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

It should be understood that the memory mentioned in the embodiments of the present application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Wherein, the non-volatile memory may be a read-only memory (Read-Only Memory, ROM), a programmable read-only memory (programmable ROM, PROM), an erasable programmable read-only memory (Erasable PROM, EPROM), an electrically programmable Erase programmable read-only memory (electrically EPROM, EEPROM) or flash memory. Volatile memory may be random access memory (RAM), which acts as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Eate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (Synchlink DRAM, SLDRAM) ) and direct memory bus random access memory (Direct Rambus RAM, DR RAM).

It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, the memory (storage module) can be integrated in the processor.

It should be noted that the memory described herein is intended to include, but not be limited to, these and any other suitable types of memory.

Based on the same inventive concept, an embodiment of the present application further provides a communication system, the communication system includes the communication device in FIG. 6 and the communication device in FIG. 7 , or includes the communication device in FIG. 8 and the communication device in FIG. 9 , Or include the communication device in FIG. 10 and the communication device in FIG. 11 .

Based on the same inventive concept, an embodiment of the present application further provides a computer-readable storage medium, including a program or an instruction. When the program or instruction is run on a computer, the network in the embodiments shown in FIG. 4 to FIG. 5 can be The method executed by the device is executed.

Based on the same inventive concept, an embodiment of the present application further provides a computer-readable storage medium, including a program or an instruction. When the program or instruction is run on a computer, the terminal in the embodiments shown in FIG. 4 to FIG. 5 can be The method performed by the device or the first terminal device is performed.

Based on the same inventive concept, an embodiment of the present application further provides a chip, which is coupled to a memory and used to read and execute program instructions stored in the memory, so that the The method performed by the network device is performed.

Based on the same inventive concept, an embodiment of the present application further provides a chip, which is coupled to a memory and used to read and execute program instructions stored in the memory, so that the The method performed by the terminal device or the first terminal device is performed.

Based on the same inventive concept, an embodiment of the present application further provides a computer program product, including instructions, which, when running on a computer, cause the methods performed by the network devices in the embodiments shown in FIG. 4 to FIG. 5 to be executed.

Based on the same inventive concept, an embodiment of the present application also provides a computer program product, which includes instructions, which when run on a computer, cause the terminal device or the first terminal device in the embodiments shown in FIG. 4 to FIG. 5 to execute the program. method is executed.

Since the communication devices shown in FIG. 6 to FIG. 11 provided in the embodiments of the present application can be used to execute the methods provided by the corresponding embodiments in the embodiments shown in FIG. 4 to FIG. 5 , the technical effects that can be obtained may refer to The foregoing method embodiments are not repeated here.

The embodiments of the present application are described with reference to flowcharts and/or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

In the above-mentioned embodiments, it may be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented in software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, from a website site, computer, server or data center via Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, digital subscriber line, DSL) or wireless (eg, infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes an integration of one or more available media. The available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, digital versatile disc (DVD)), or semiconductor media (eg, solid state disk (SSD) ))Wait.

Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

A method for updating a machine learning model, wherein the method comprises:

The network device determines model training configuration information corresponding to the terminal device according to the computing capability of the terminal device;

The network device sends the model training configuration information to the terminal device;

receiving, by the network device, a model update parameter sent by the terminal device, wherein the model update parameter is a model parameter updated by the terminal device after training the first machine learning model according to the model training configuration information;

The network device updates the second machine learning model according to the model update parameter.
The method of claim 1, wherein the model training configuration information includes at least one of the following:

hyperparameters;

precision;

Training time information.
The method of claim 1 or 2, wherein the method further comprises:

The network device sends training feature information to the terminal device, where the training feature information is used to indicate a training feature set used by the terminal device to train the first machine learning model.
The method of claim 1 or 2, wherein the method further comprises:

The network device sends accuracy evaluation information to the terminal device, where the accuracy evaluation information includes at least one of a method for evaluating accuracy or a test sample for evaluating accuracy.
The method of claim 1 or 2, wherein the method further comprises:

The network device receives accuracy indication information from the terminal device, where the accuracy indication information is used to indicate the accuracy achieved by the terminal device after training the first machine learning model by using the model training configuration information.
The method of claim 1 or 2, wherein the method further comprises:

determining, by the network device, a time point for acquiring the model update parameter of the terminal device;

Send an acquisition request to the terminal device at the time point, where the acquisition request is used to instruct the terminal device to send the model update parameter of the terminal device to the network device; or,

Sending reporting time information to the terminal device, where the reporting time information is used to instruct to send model update parameters to the network device at the time point.
The method according to claim 6, wherein determining, by the network device, a time point for acquiring the model update parameters of the terminal device comprises:

The network device determines the transmission duration for each terminal device in the plurality of terminal devices to send the respective model update parameters to the network device;

The network device determines the time point according to each transmission duration corresponding to each terminal device.
The method of claim 6, wherein the obtaining request is further used to indicate that the specified model update parameters need to be obtained.
The method of claim 6, wherein the method further comprises:

The network device receives parameter availability indication information from the terminal device, where the parameter availability indication information is used to indicate the availability of model update parameters in the terminal device.
A method for updating a machine learning model, the method comprising:

The terminal device receives the model training configuration information sent by the network device, and the model training configuration information is determined according to the computing capability of the terminal device;

The terminal device trains the first machine learning model according to the model training configuration information to obtain model update parameters;

The terminal device sends the model update parameter to the network device, where the model update parameter is used by the network device to update the second machine learning model.
The method of claim 10, wherein the method further comprises:

receiving, by the terminal device, training feature information from the network device, where the training feature information is used to indicate a training feature set used by the terminal device to train the first machine learning model;

The terminal device trains the first machine learning model according to the model training configuration information, including:

The terminal device trains the first machine learning model according to the model training configuration information and the training feature information.
The method of claim 10, wherein the method further comprises:

The terminal device receives accuracy evaluation information from the network device, the accuracy evaluation information including at least one of a method for evaluating accuracy or a test sample for evaluating accuracy;

The terminal device determines the accuracy achieved by the trained first machine learning model according to the accuracy evaluation information.
The method according to any one of claims 10-12, wherein the method further comprises:

The terminal device sends accuracy indication information to the network device, where the accuracy indication information is used to indicate the accuracy achieved by the terminal device after training the first machine learning model by using the model training configuration information.
The method according to any one of claims 10-12, wherein the method further comprises:

The terminal device receives an acquisition request from the network device, and sends the model update parameter to the network device according to the acquisition request, where the acquisition request is used to instruct the terminal device to send the data to the network device. the model update parameters of the terminal device described above; or,

The terminal device receives the report time information from the network device, and sends the model update parameter to the network device at the time point indicated by the report time information.
The method according to any one of claims 10-12, wherein the method further comprises:

The terminal device sends parameter availability indication information to the network device, where the parameter availability indication information is used to indicate the availability of a model update parameter in the terminal device.
A communication device, comprising:

a processing unit, configured to determine model training configuration information corresponding to the terminal device according to the computing capability of the terminal device;

a communication unit, configured to send the model training configuration information to the terminal device, and receive a model update parameter sent by the terminal device, wherein the model update parameter is the terminal device according to the model training configuration information Model parameters updated after training the first machine learning model;

The processing unit is further configured to update the second machine learning model according to the model update parameter.
The apparatus of claim 16, wherein the model training configuration information includes at least one of the following:

hyperparameters;

precision;

Training time information.
The device according to claim 16 or 17, wherein the communication unit is further used for:

Sending training feature information to the terminal device, where the training feature information is used to indicate a training feature set used by the terminal device to train the first machine learning model.
The device according to claim 16 or 17, wherein the communication unit is further used for:

Sending accuracy evaluation information to the terminal device, the accuracy evaluation information including at least one of a method for evaluating accuracy or a test sample for evaluating accuracy.
The device according to claim 16 or 17, wherein the communication unit is further used for:

Accuracy indication information from the terminal device is received, where the accuracy indication information is used to indicate the accuracy achieved by the terminal device after training the first machine learning model by using the model training configuration information.
The device of claim 16 or 17, wherein:

The processing unit is also used to:

determining a time point for acquiring the model update parameters of the terminal device;

Then, the communication unit is also used for:

Send an acquisition request to the terminal device at the time point, where the acquisition request is used to instruct the terminal device to send the model update parameter of the terminal device to the network device; or,

Sending reporting time information to the terminal device, where the reporting time information is used to instruct to send model update parameters to the network device at the time point.
The apparatus of claim 21, wherein the processing unit is specifically configured to:

determining the transmission duration for each terminal device in the plurality of terminal devices to send the respective model update parameters to the network device;

The time point is determined according to each transmission duration corresponding to each terminal device.
The apparatus according to claim 21, wherein the obtaining request is further used to indicate that the specified model update parameter needs to be obtained.
The apparatus of claim 21, wherein the communication unit is further configured to:

Parameter availability indication information is received from the terminal device, where the parameter availability indication information is used to indicate the availability of model update parameters in the terminal device.
A communication device, comprising:

a communication unit, configured to receive model training configuration information sent by the network device, where the model training configuration information is determined according to the computing capability of the terminal device;

a processing unit, configured to train the first machine learning model according to the model training configuration information to obtain model update parameters;

The communication unit is further configured to send the model update parameter to the network device, where the model update parameter is used by the network device to update the second machine learning model.
The apparatus of claim 25, wherein the communication unit is further configured to:

receiving training feature information from the network device, where the training feature information is used to instruct the terminal device to use a training feature set for training the first machine learning model;

Then, the processing unit is also used for:

The first machine learning model is trained according to the model training configuration information and the training feature information.
The apparatus of claim 25, wherein the communication unit is further configured to:

receiving accuracy evaluation information from the network device, the accuracy evaluation information including at least one of a method for evaluating accuracy or a test sample for evaluating accuracy;

Then, the processing unit is also used for:

The accuracy achieved by the trained first machine learning model is determined according to the accuracy evaluation information.
The apparatus according to any one of claims 25-27, wherein the communication unit is used for:

Sending accuracy indication information to the network device, where the accuracy indication information is used to indicate the accuracy achieved by the terminal device after training the first machine learning model by using the model training configuration information.
The apparatus according to any one of claims 25-27, wherein the communication unit is further configured to:

Receive an acquisition request from the network device, and send the model update parameter to the network device according to the acquisition request, where the acquisition request is used to instruct the terminal device to send the terminal device's information to the network device. model update parameters; or,

Receive reporting time information from the network device, and send the model update parameter to the network device at a time point indicated by the reporting time information.
The apparatus according to any one of claims 25-27, wherein the communication unit is further configured to:

Sending parameter availability indication information to the network device, where the parameter availability indication information is used to indicate the availability of the model update parameter in the terminal device.
A communication device, characterized in that it includes:

at least one processor; and a memory, a communication interface communicatively coupled to the at least one processor;

Wherein, the memory stores instructions executable by the at least one processor, and the at least one processor executes the instructions stored in the memory to cause the apparatus to perform the method described in claim 1-9 or 10-15. The method of any one.
A computer-readable storage medium, characterized in that it includes a program or an instruction, which, when the program or instruction is run on a computer, causes the method according to any one of claims 1-9 or 10-15 to be executed .