Model training method and communication device

Abstract

The present application provides a model training method and a communication device, which relate to the field of communications. The model training method includes the steps of receiving capability information from multiple child node devices representing the capability of the child node devices to train a model, transmitting to each of the multiple child node devices a model and first instruction information indicating first model parameters to be updated when the child node device trains the model, receiving the updated first model parameters from the multiple child node devices, fusing global model parameters of the multiple child node devices using a predetermined fusion algorithm to obtain fused parameters, and transmitting the fused parameters to the multiple child node devices. This solves the problem of low efficiency in model training using existing federated learning architectures.

Description

[Technical Field]

[0001] This application claims priority to Chinese Patent Application No. 202210872876.X, entitled "Model Training Method and Communication Apparatus," filed with the State Intellectual Property Office of the People's Republic of China on July 21, 2022, which is incorporated herein by reference in its entirety.

[0002] The present application relates to the field of communications, and in particular to a model training method and a communications device. [Background technology]

[0003] Based on the conventional federated learning architecture, each child node device needs to download a unified model from the central node device for local training. Therefore, if the child node devices have different capabilities, the model training efficiency will be limited by the child node device with the lowest capability, resulting in low model training efficiency. Summary of the Invention [Means for solving the problem]

[0004] The present application provides a model training method and a communication device to solve the existing problem of low efficiency in model training using federated learning.

[0005] To achieve the aforementioned objectives, the present application uses the following technical solutions:

[0006] According to a first aspect, the present application provides a model training method, the method being applied to a central node device, the method comprising: receiving capability information from a plurality of child node devices (child node devices participating in federated learning of the model) representing a capability of training the model by the child node devices; transmitting to each of the plurality of child node devices the model and first instruction information instructing first model parameters (determined based on the capability information of the child node devices) to be updated when the child node device trains the model; receiving updated first model parameters from the plurality of child node devices; fusing global model parameters of the plurality of child node devices (including the updated first model parameters and second model parameters other than the first model parameters in the model) using a predetermined fusion algorithm to obtain fused parameters; and transmitting the fused parameters to the plurality of child node devices. Includes.

[0007] In a first aspect, first, multiple child node devices transmit capability information to a central node device, and then the central node device transmits first instruction information and a model to each of the multiple child node devices based on the capability information. Then, the multiple child node devices train a model based on a training set and the model parameters to be updated to obtain updated first model parameters. Additionally, the central node device uses a predefined fusion algorithm to fuse global model parameters of the multiple child node devices, including the first model parameters, to obtain fused parameters. Then, the multiple child node devices train a model based on the fused parameters. In this way, when the child node devices train a model, some of the parameters are frozen and do not participate in the training. Therefore, during model training, the progress of the model training is not affected by the central node device, which has low computing power, and model training efficiency is high.

[0008] In one possible implementation, the method comprises: separately predicting model training periods for a plurality of child node devices based on the capability information; and determining first model parameters for the child node devices based on the model training periods; Further includes:

[0009] In this implementation, the model training periods of multiple child node devices are first predicted separately based on capability information, and then the first model parameters of the child node devices are determined based on the model training periods, so that the determined first model parameters to be trained can match the corresponding child node devices and ensure the progress of model training.

[0010] In one possible implementation, the step of determining first model parameters of the child node device based on the model training period includes: Selecting some of the parameters of the model as the first model parameters if the model training period does not satisfy the predetermined period condition, or selecting all of the parameters of the model as the first model parameters if the model training period satisfies the predetermined period condition. Includes.

[0011] In this implementation, based on different comparison results between the model training period and the predetermined period conditions, two policies are used accordingly: a policy of selecting part of the model's parameters as the first model parameters, and a policy of selecting all of the model's parameters as the first model parameters, so that the progress of model training is guaranteed and the model performance is guaranteed at the same time.

[0012] In one possible implementation, the step of selecting a portion of the parameters of the model as the first model parameters comprises: Randomly selecting a portion of the parameters of the model as first model parameters. Includes.

[0013] In this implementation, some of the model parameters are randomly selected as the first model parameters, so that the determined first model parameters to be trained can match the corresponding child node devices, and the progress of model training is guaranteed.

[0014] In one possible implementation, the model includes a first network layer (the model application scenario indicates the first network layer) and a second network layer, the first network layer and the second network layer are different, and the step of selecting some of the parameters of the model as the first model parameters includes: Selecting the parameters of the first network layer and a portion of the parameters of the second network layer as first model parameters. Includes.

[0015] In this implementation, the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, so that the requirements of the model application scenario for the network layers are taken into consideration and the model training performance is guaranteed.

[0016] In one possible implementation, before the predetermined first network layer parameters in the first network layer are selected as the first model parameters, the method includes: determining the first network layer required to build the model based on the model application scenario; Further includes:

[0017] In this implementation, the central node device determines the first network layer required to build a model based on the model application scenario, so that the requirements of the model application scenario for the network layer are taken into consideration and model training performance is guaranteed.

[0018] In one possible implementation, the capability information further indicates a type of model that can be trained by the child node device, and the method further comprises: determining a child node device whose capability information includes a model type of the model as a target device to which the first instruction information and the model are to be transmitted; Further includes:

[0019] In this implementation, a child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, thereby ensuring that all child node devices participating in the training can train the model and ensuring progress in model training.

[0020] In one possible implementation, the step of fusing global model parameters of multiple child node devices using a predefined fusion algorithm to obtain fused parameters includes: For model parameters having the same function of different child node devices, setting the weight of the second model parameter to a first default weight value and setting the weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter; or, for model parameters having the same function of different child node devices, setting both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter. Includes.

[0021] In this implementation, the second model parameters are designed to be unable to affect the default fusion algorithm of the fused parameters and unable to participate in the calculation of the fused parameters, so that the second model parameters that are not updated cannot affect the model training, and model training performance is guaranteed.The second model parameters are further designed to be able to affect the default fusion algorithm of the fused parameters and able to participate in the calculation of the fused parameters, so that the determined fused parameters can participate in the model training, and model training progress is guaranteed.

[0022] According to a second aspect, the present application provides a beam information selection method, the method being applied to a child node device, comprising: receiving beam information; inputting the beam information into a model (obtained through training according to the method of the first aspect) and outputting target beam information; Includes.

[0023] In the second embodiment, a model is used to select the target beam, which provides a high efficiency in selecting the target beam.

[0024] According to a third aspect, the present application provides a model training method, the method being applied to a child node device, comprising: transmitting capability information representing a capability of training the model by the child node device to the central node device; receiving from the central node device the model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters of the child node device to be updated are determined based on the capability information of the child node device; training the model based on the model parameters to be updated and a training set including beam information collected by the child node device to obtain updated first model parameters; transmitting the updated first model parameters to the central node device; receiving from the central node device fused parameters calculated based on the updated first model parameters; and training the model based on the fused parameters. Includes.

[0025] According to a fourth aspect, the present application provides a communication device, which may be a chip or a system-on-chip of a central node device, a receiving module configured to receive capability information representing a capability of training a model by a child node device from a plurality of child node devices (child node devices participating in federated learning of the model); a transmitting module configured to transmit to each of the plurality of child node devices the model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters to be updated of the child node device are determined based on the capability information of the child node device, and the receiving module is further configured to receive the updated first model parameters from the plurality of child node devices; and a processing module configured to fuse global model parameters of the plurality of child node devices using a predetermined fusion algorithm to obtain fused parameters, wherein the global model parameters include the updated first model parameters and second model parameters other than the first model parameters in the model, and the transmitting module is configured to transmit the fused parameters to the plurality of child node devices; Includes.

[0026] In one possible implementation, the processing module: Separately predicting model training periods for the plurality of child node devices based on the capability information, and determining first model parameters for the child node devices based on the model training periods. It is specifically structured as follows.

[0027] In one possible implementation, the processing module: Selecting some of the parameters of the model as the first model parameters when the model training period does not satisfy a predetermined period condition, or selecting all of the parameters of the model as the first model parameters when the model training period satisfies a predetermined period condition. It is specifically structured as follows.

[0028] In one possible implementation, the processing module: Randomly select some of the model's parameters as the first model parameters It is specifically structured as follows.

[0029] In one possible implementation, the model includes a first network layer and a second network layer, the model application scenario indicates the first network layer, the first network layer and the second network layer are different, and the processing module is specifically configured to select parameters of the first network layer and some of the parameters of the second network layer as first model parameters.

[0030] In one possible implementation, the processing module: Before selecting a predetermined first network layer parameter in the first network layer as a first model parameter, a first network layer required for constructing a model is determined based on a model application scenario. It is further configured as follows.

[0031] In one possible implementation, the capability information further indicates a type of model that can be trained by the child node device, and the processing module: A child node device whose capability information includes a model type of the model is determined as a target device to which the first instruction information and the model are to be transmitted. It is further configured as follows.

[0032] In one possible implementation, the processing module: For model parameters having the same function of different child node devices, set the weight of the second model parameter to a first default weight value, set the weight of the first model parameter to a second default weight value, and perform weighted averaging to obtain fused parameters; or for model parameters having the same function of different child node devices, set both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and perform weighted averaging to obtain fused parameters. It is specifically structured as follows.

[0033] According to a fifth aspect, the present application provides a communication device, which may be a child node device chip or a system-on-chip, a receiving module configured to receive beam information; and a processing module configured to input the beam information into a model and output target beam information, the model being obtained through training by a communication device according to the fourth aspect; Includes.

[0034] According to a sixth aspect, the present application provides a communication device, which may be a child node device chip or a system-on-chip, The system includes: a transmitting module configured to transmit capability information representing a capability of training a model by a child node device to a central node device; a receiving module configured to receive from the central node device a model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters of the child node device to be updated are determined based on the capability information of the child node device; and a processing module configured to train the model based on a training set (including beam information collected by the child node device) and the model parameters to be updated to obtain updated first model parameters, wherein the transmitting module is configured to transmit the updated first model parameters to the central node device, and the receiving module is configured to receive from the central node device fused parameters calculated based on the updated first model parameters, and the processing module is configured to train the model based on the fused parameters.

[0035] According to a seventh aspect, the present application provides a communications device. The communications device includes a processor and a transceiver, the processor and the transceiver configured to support the communications device in performing a method according to the first, second, or third aspect. The communications device may further include a memory. The memory stores computer instructions, and the processor is operable to execute the computer instructions to perform the method according to the first, second, or third aspect.

[0036] According to an eighth aspect, the present application provides a computer-readable storage medium having stored thereon computer instructions that, when executed, perform a method according to the first, second, or third aspect.

[0037] According to a ninth aspect, the present application provides a computer program product comprising instructions which, when executed on a computer, enable the computer to perform a method according to the first, second or third aspect.

[0038] According to a tenth aspect, the present application provides a chip, the chip including a processor and a transceiver, the processor and the transceiver configured to support a communications device in performing a method according to the first aspect, the second aspect, or the third aspect.

[0039] For the beneficial effects described in the fourth to tenth aspects of the present application, please refer to the analysis of the beneficial effects in the first to third aspects accordingly, and the details will not be described again here. [Brief explanation of the drawings]

[0040] [Figure 1] FIG. 1 is a diagram of the structure of a neuron according to an embodiment of the present application. [Figure 2] FIG. 1 is a diagram of a neural network structure of a DNN according to an embodiment of the present application. [Figure 3]FIG. 1 is a diagram of a system architecture for conventional federated learning according to an embodiment of the present application. [Figure 4-a] 1 is a schematic flowchart of federated learning according to an embodiment of the present application; [Figure 4-b] 1 is another schematic flowchart of federated learning according to an embodiment of the present application; [Figure 5] 1 is another schematic flowchart of federated learning according to an embodiment of the present application; [Figure 6] FIG. 1 is a diagram of a differentiated network parameter update principle according to an embodiment of the present application; [Figure 7] 1 is a diagram of a communication system according to an embodiment of the present application; [Figure 8] 1 is a schematic flowchart of a model training method according to an embodiment of the present application; [Figure 9] FIG. 1 is a diagram of a network structure of a model according to an embodiment of the present application. [Figure 10] 1 is a schematic flowchart of another model training method according to an embodiment of the present application. [Figure 11] 1 is a diagram of the structure of a communication device according to an embodiment of the present application; [Figure 12] FIG. 10 is a diagram of the structure of another communication device according to an embodiment of the present application; [Figure 13] FIG. 10 is a diagram of the structure of another communication device according to an embodiment of the present application; [Figure 14] 1 is a diagram of the structure of a communication system according to an embodiment of the present application; DETAILED DESCRIPTION OF THE INVENTION

[0041] The following describes the technical solutions of the embodiments of the present application with reference to the accompanying drawings of the embodiments of the present application. Unless otherwise specified, " / " in the description of the present application represents an "or" relationship between related objects. For example, A / B can represent A or B. In the present application, "and / or" is merely a relationship for describing related objects and indicates that three relationships may exist. For example, A and / or B can represent the following three cases: when only A exists, when both A and B exist, and when only B exists, and A and B may be singular or plural. In addition, in the description of the present application, "plurality" means two or more unless otherwise specified. "At least one of the following items (moieties)" or similar expressions refers to any combination of these items, including any combination of a single item (moiety) or multiple items (moieties). For example, at least one of a, b, or c can refer to a, b, c, a and b, a and c, b and c, or a, b, and c, where a, b, and c may be singular or plural. Additionally, to clearly describe the technical solutions of the embodiments of the present application, terms such as "first" and "second" are used in the embodiments of the present application to distinguish between identical or similar items that provide essentially the same functions and purposes. Those skilled in the art can understand that terms such as "first" and "second" do not limit the quantity or execution order, and terms such as "first" and "second" do not indicate clear distinctions. Additionally, terms such as "example" or "for example" are used in the embodiments of the present application to represent providing an example, illustration, or explanation. Any embodiment or design scheme described using "example" or "for example" in the embodiments of the present application should not be described as being preferred over another embodiment or design scheme, or as having more advantages than another embodiment or design scheme. Strictly speaking, the use of terms such as "example" or "for example" is intended to present related concepts in a specific way for ease of understanding.

[0042] In addition, the network architectures and service scenarios described in the embodiments of the present application are intended to more clearly explain the technical solutions of the embodiments of the present application, and do not constitute limitations on the technical solutions provided in the embodiments of the present application. As those skilled in the art can understand, with the evolution of network architectures and the emergence of new service scenarios, the technical solutions provided in the embodiments of the present application can also be applied to similar technical problems.

[0043] Before describing the embodiments of the present application, some terms of the embodiments of the present application will be explained.

[0044] Artificial intelligence is the ability of machines to possess human intelligence and to mimic some intelligent human behavior using computer software and hardware. Artificial intelligence includes machine learning and many other methods.

[0045] Machine learning is the learning of models or rules from raw data. There are many different machine learning methods, including those that use neural networks, decision trees, and support vector machines.

[0046] The AI model here is a function model that maps inputs of a specific dimension to outputs of a specific dimension. The model parameters of the AI model are obtained through machine learning training. For example, f(x)=ax^2+b is a quadratic function model and can be considered an AI model. a and b are the model parameters, which can be obtained through machine learning training.

[0047] The neural network here is an artificial neural network, a mathematical model that performs distributed parallel information processing by imitating the behavioral characteristics of animal neural circuits, and is a special form of AI model.

[0048] A dataset includes data used for model training, validation, and testing in machine learning. The quantity and quality of data affect the effectiveness of machine learning.

[0049] Model training involves selecting an appropriate loss function and training the model parameters using an optimization algorithm to minimize the loss function value.

[0050] The loss function is used to measure the difference between the model's predicted values and the actual values.

[0051] Model testing is the evaluation of model performance after training using test data.

[0052] Model application is the use of a trained model to solve a real-world problem.

[0053] Building a new generation of information infrastructure that adapts to the Internet of Everything in an intelligent society and ensuring its security are crucial for promoting the integration of information technology and the real economy and expanding the digital economy. Currently, there is a global wave of artificial intelligence (AI) applications. In 5G networks, multiple service modes coexist, and the application environment is complex and dynamic. Traditional approaches—static configuration, manual maintenance, and single-service capabilities—cannot meet the demands of 5G. Introducing AI into 5G and future communication systems can provide data-driven perception, prediction, and management and control capabilities for information infrastructure, including networks, computing, and applications, and promote the integration and collaboration of these infrastructures. In increasingly complex scenarios, AI can make better decisions than humans. This will undoubtedly open up a new vision for intelligent network construction, bring new network development opportunities, and provide efficient and rapid solutions to the many difficulties and challenges that arise in the reconstruction and transformation of telecommunications networks.

[0054] Machine learning (ML) is an important technological means for realizing artificial intelligence. Machine learning can be divided into supervised learning, unsupervised learning, and reinforcement learning.

[0055] In supervised learning, a machine learning algorithm is used to learn a mapping relationship between sample values and sample labels based on collected sample values and sample labels, and the learned mapping relationship is represented using a machine learning model. The process of training a machine learning model is a process of learning a mapping relationship. For example, during signal detection, a received signal containing noise is a sample, and the actual signal point corresponding to the signal is a label. Machine learning expects to learn the mapping relationship between samples and labels through training, in other words, enable the machine learning model to train a signal detector. During training, model parameters are optimized by calculating the error between the model's predicted value and the actual label. Once the mapping relationship is learned, the learned mapping can be used to predict the sample label of each new sample. The mapping relationship learned in supervised learning may include linear mapping and nonlinear mapping. Learning tasks may be classified into classification tasks and regression tasks based on the type of label.

[0056] In unsupervised learning, internal patterns of samples are autonomously explored by using an algorithm based only on collected sample values. In a specialized unsupervised learning algorithm, the samples are used as supervised signals, in other words, the model learns the mapping relationship between the samples, which is called self-supervised learning. During training, the model parameters are optimized by calculating the error between the model's predicted values and the samples. Self-supervised learning can be used for signal compression and decompression. Common algorithms include autoencoders and generative adversarial networks.

[0057] Unlike supervised learning, reinforcement learning is an algorithm that learns a policy to solve a problem by interacting with the environment. Unlike supervised and unsupervised learning, reinforcement learning does not have a clear "correct" action label. The algorithm must interact with the environment to obtain a reward signal fed back by the environment and adjust its decision action to obtain a larger reward signal value. For example, in downlink power control, a reinforcement learning model adjusts each user's downlink transmission power based on the total system throughput fed back by the wireless network in the hope of achieving higher system throughput. The goal of reinforcement learning is to learn the mapping relationship between the environmental state and the optimal decision action. However, the "correct action" label cannot be obtained in advance. Therefore, it is not possible to calculate the error between the action and the "correct action" and optimize the network. Reinforcement learning training is carried out through repeated interactions with the environment.

[0058] Deep neural networks (DNNs) are a specialized implementation of machine learning that typically optimizes model parameters using supervised or unsupervised learning policies. According to the universal approximation theorem, neural networks can theoretically approximate any continuous function, and therefore have the ability to learn any mapping. Traditional communication systems require extensive expert knowledge to design communication modules. However, DNN-based deep learning communication systems can automatically discover implicit pattern structures from large datasets, establish mapping relationships between data, and achieve better performance than traditional modeling methods.

[0059] The idea behind DNNs comes from the neuron structure of the brain. As shown in the neuron structure in Figure 1, each neuron performs a weighted sum operation on its inputs, and the weighted sum is passed through a nonlinear function to generate an output. Specifically, if the inputs to a neuron are x=[x0,…,xn ], and the weight values corresponding to the input are w=[w0,…,w n ] and the weighted sum bias is b. The nonlinear function can take various forms. We use the maximum function max{0,x} as an example, and the corresponding execution effect of the neuron is

number

[0060] DNNs typically have two or more hidden layers, which directly affect their information extraction and function fitting capabilities. Increasing the number of hidden layers or the width of each layer can improve the function fitting capabilities of DNNs. The weights in each neuron are the parameters of the DNN model. The model parameters are optimized during the training process to ensure that the DNN has the ability to extract data features and represent mapping relationships.

[0061] Based on the network construction method, DNNs can be classified into feedforward neural networks (FNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Figure 2 shows an FNN, which is characterized by the fact that neurons in adjacent layers are fully connected. As a result, FNNs usually require a large amount of storage space and have high computational complexity.

[0062] CNN is a neural network specialized for processing data with a similar grid structure. For example, both time series data (timeline discrete sampling) and image data (two-dimensional discrete sampling) can be considered as data with a similar grid structure. CNN does not perform calculations using all input information at once, but instead performs convolution calculations by incorporating partial information through a fixed-size window, which significantly reduces the amount of calculation required for model parameters. In addition, based on the different types of information incorporated through the window (for example, people and objects in the same image represent different types of information), different convolution kernel operations can be used for each window, allowing CNN to better extract features from the input data.

[0063] RNN is a DNN that uses feedback time-series information. The input of the RNN includes the new input value at the current time and the output value of the RNN at the previous time. RNN is suitable for obtaining time-correlated sequence features, and is particularly suitable for applications such as speech recognition and channel coding and decoding.

[0064] The aforementioned FNN, CNN, and RNN are common neural network structures, and these network structures are built based on neurons. As mentioned above, each neuron performs a weighted sum operation on the neuron's input values, and the weighted sum result is passed through a nonlinear function to generate an output. In this case, the weight values of the weighted sum operation of the neurons in a neural network and the nonlinear function are called the parameters of the neural network. As an example, we will use a neuron that uses the nonlinear function max{0,x}.

number

[0065] DNNs can be trained using federated learning (FL). Federated learning is a machine learning framework designed to effectively support multiple organizations in using data to build machine learning models while meeting user privacy and data security requirements. In a federated learning framework, data is not transferred between nodes. Instead, intermediate results obtained through training, such as model parameters and gradients, are transferred between nodes. Figure 3 illustrates the system architecture of a conventional federated learning system. As shown in Figure 3, the system architecture of a conventional federated learning system may include a central node 301 and multiple distributed nodes (distributed nodes 302a-302c). The distributed nodes 302a-302c are all connected to the central node 301 and belong to different organizations or companies. Each of the distributed nodes 302a-302c contains a distributed dataset related to the application environment of the distributed node. As a distributed machine learning paradigm, federated learning can effectively solve the problem of data silos. Federated learning allows participants to perform joint modeling without sharing data, thus technically breaking down data silos and realizing AI collaboration. Based on the distribution of participants' different data sources, federated learning can be divided into three types: horizontal federated learning, vertical federated learning, and federated transfer learning.

[0066] Horizontal associative learning means that when two datasets have high overlap in user features but low overlap in users, the datasets are split horizontally (i.e., in the user dimension) and portions of data with the same user features but not the exact same user are extracted from the datasets for training. Vertical associative learning means that when two datasets have high overlap in users but low overlap in user features, the datasets are split vertically (i.e., in the feature dimension) and portions of data with the same user but not the exact same user features are extracted from the datasets for training. Associative transfer learning means that when two datasets have low overlap in both users and user features, the data are not split and transfer learning can be used to overcome data or label shortages.

[0067] As an example, we use the high-frequency beam management problem. When a base station performs beam sweeping on codebook-based synchronization signals and physical broadcast channel (PBCH) blocks, i.e., synchronization signal blocks (SSBs) or channel state information-reference signals (CSI-RSs), the channels between users at different locations and the base station are different. The user terminal (or user for short) needs to measure the parameters of the received SSB or CSI-RS beams, such as the physical layer reference signal received power (L1-RSRP), select the beam corresponding to the maximum RSRP value as the target beam, and feed the target beam back to the base station.

[0068] A user terminal can train a model for beam selection using AI / ML. For example, the user terminal uses (some or all of) multiple received SSB / CSI-RS signals as input or (some or all of) multiple received SSB / CSI-RS strength (RSRP) signals as input to estimate a target beam ID and feed the target beam ID back to the base station. Each user can collect its received beam / channel information and the corresponding target beam ID as samples (i.e., local samples) for training an AI / ML model. However, the number of samples each user can collect is limited, and the target beam IDs in the training samples are only a subset of the actual SSB / CSI-RS codebook. As a result, the performance of a model obtained by training a user using only local data is limited.

[0069] If users send their local data to a server and the server summarizes the users' data for model training, model performance can be improved, but there is a risk of leaking users' privacy information. For example, information such as the user's current location may be inferred based on the channel. To solve this problem, federated learning can be used. In the federated learning framework, users correspond to child nodes, and network devices correspond to a central node. The central node distributes a global model to each user participating in federated learning. Each user trains the global model using their local data to obtain a local model and sends parameter information of the local model, such as gradients and weight values (after encryption), to the network device. The network device performs model aggression (MA) to update the global model and then sends the updated global model to each user. The users continue to update their local models, output parameter information of the updated local models, and send the parameter information to the central node. In this way, multiple iterations are performed until the model converges.

[0070] In traditional federated learning architectures, AI is applied to wireless networks, and training data, such as label information, must first be collected. However, some local user data may pose privacy issues. For example, in beam management and positioning scenarios, user equipment (UE) may collect data including user channel, target beam ID, and geographic location. Because the data collected by a single UE may not be universal (e.g., as mentioned above, due to the user's geographic location, the target beam ID fed back by the user may only be a subset of the entire codebook), it is not possible to traverse all possible distributions of actual data, and the amount of data received by a single UE may be small. Therefore, training using only data from a single UE is difficult to achieve sufficient neural network performance. However, if all UEs transmit their original data to a central node for unified training, it will incur very high air interface overhead and may pose privacy issues.

[0071] In the above wireless communication scenario, when different UEs need to perform distributed training on the same type of data, horizontal federated learning is suitable for use. The main steps of horizontal federated learning are as follows: Step 1: A central node (e.g., a base station or an AI network element) distributes a global model to child nodes (e.g., UEs). Step 2: Each child node trains a model using its local data and uploads its gradients, weights, or local model (after encryption) to the central node, which aggregates the gradients or local models of its child nodes to update the parameters of the global model.

[0072] The two steps are repeated until the model converges.

[0073] In the federated learning procedure shown in Figure 4-a, three child nodes (users 1, 2, and 3) are used as an example. First, the central node sends an initial model (i.e., the downloaded global model M0 in the figure) to the child nodes. Users 1, 2, and 3 train (global) models based on local data 1, 2, and 3, respectively, to obtain models M1, M2, and M3, and then send these models to the central node. The central node performs model fusion on the received child node models to obtain an updated global model. For example, the fusion algorithm FedAvg (Federated averaging) is used, i.e.

number

[0074] In a traditional federated learning architecture, each user needs to download a unified model from a central node for local training. Taking into account the different capabilities of the user devices participating in the training (the difference in capabilities may refer to fluctuations in the computing power of the same user device at different times caused by power changes or the execution of different authorized programs), computing power can be understood as the computing capability or computing speed of the terminal device, and an indicator that can measure the computing power of the terminal device is floating point operations per second (FLOPS). In a federated learning architecture, a child node (i.e., a user participating in a particular round of federated learning) must take on the role of training the network. However, model complexity usually varies from model to model. Therefore, a common indicator for measuring model complexity is floating point operations (FLOPs). Specifically, different model complexities can be reflected in the following: different network types, such as convolutional networks (CNN), fully connected networks (FC), and recurrent neural networks (RNN); the same network type with different network depths / layers, such as multi-layered CNN or FC; and different numbers of neurons, such as the number of neurons in the hidden layer in an FC network. Due to differences in computing power, the ability to train the same model varies depending on the terminal device, and it can be seen that different computing power is required to train different models even on the same terminal device. The following three cases can occur:

[0075] Case 1: All users participating in training report their computational power (training power, memory, etc.) to the central node, and the central node distributes a model that can be trained by all users, as shown in Figure 4-b. In this case, the complexity of training the model is limited by the users with low computational power. Since the central node adjusts the global model based on the computational power reported by all users, even user 1 with low computational power can train the model in this case. However, the performance of the global model may be limited by the user with the lowest computational power.

[0076] Case 2: As shown in Figure 5, the unified model is downloaded, but some users are unable to train the model and ultimately cannot use the global model, which may result in losses. After user 1 downloads M0, user 1 is unable to train the global model due to limited computing power. As a result, the features of user 1's local data cannot be learned by the global model, which may ultimately result in a decrease in the performance of predicting user 1's beam using the global model.

[0077] Case 3: The central node performs differential model updates for different user terminals. Specifically, the base station broadcasts multiple global models (grouped based on the computational power reported by the child nodes) or, optionally, both multiple global models and a classification neural network. The child node determines the group of the global model or classification neural network to which it belongs based on predetermined conditions, updates the local model corresponding to that group, reports the parameters, and reports the index of the child node's updated model. The base station determines the global model corresponding to the parameters reported by the child node based on the index and updates the global model. The base station then broadcasts the multiple global models again in the second round of federated learning. This corresponds to implementing multiple (parallel) federated learning architectures. The central node is responsible for unifying the management of multiple global models, with each model corresponding to a group of computational power users. Although this method solves the problem of being unable to train a unified model due to the difference in computing power between user equipment, it has the following drawbacks: multiple models need to be exchanged (broadcast) over the air interface, which increases the air interface overhead; the amount of data required to train one model is reduced; multiple models are distributed in groups, which reduces the total amount of data required to train one model; and it is difficult to guarantee model performance.

[0078] In short, the efficiency of model training in existing federated learning frameworks is low.

[0079] To solve the aforementioned technical problems, the embodiments of the present application provide a model training method. The method provided in the embodiments of the present application will be described below with reference to the accompanying drawings.

[0080] The embodiments of the present application provide a method for supporting a central node in designing and distributing differential AI / ML network parameter update or freeze signaling based on the computational power state reported by each child node in a federated learning architecture to instruct the child nodes to train all or part of a downloaded unified model. A child node computational power feedback procedure and a procedure for the central node to perform differential AI / ML network parameter update or freeze for instructions are added to the original federated learning.

[0081] Figure 6 illustrates a specific differentiation network parameter update or freezing principle. It shows an update / freeze instruction diagram based on AI / ML network layers. All users (child nodes) participating in federated learning download the AI / ML network shown in Figure 6. This network is an AI / ML model with four network layers. In traditional federated learning, each child node must use local data to train four complete network layers (all network parameters) to obtain a model and then feed the model back to the central node. However, in an embodiment of the present application, based on different computing power, the central node instructs users with insufficient computing power to update some of the parameters in the network, while users with sufficient computing power update all parameters. As shown on the right side of Figure 6, user A updates only the network layer parameters corresponding to neurons in columns 1, 4, and 6; user B updates the network layer parameters corresponding to neurons in columns 1, 2, 4, and 6; and user C updates the network layer parameters corresponding to neurons in all columns 1, 2, 3, 4, 5, and 6. The central node is responsible for sending differentiation update instructions to each corresponding child node. In addition to being directed on a network layer basis, parameter updates can be directed on a neuron basis: instructions can indicate the network layer at which a child node should perform a freeze, or can indicate the network layer at which a child node should perform an update.

[0082] Each child node user refers to the update / freeze instructions and trains a model using its local data as a training set. After training is complete, the gradients / weights / networks (or only the parts responsible for updating) are uploaded to the central node (after encryption). The central node then fuses the gradients / weights / networks uploaded by multiple child nodes. For example, if the FedAvg algorithm is still used for fusion, and parameter update instructions based on the network layer are used, and users A, B, and C correspond to models 1, 2, and 3, then network layer g is the frozen layer for user A and the updated layer for users B and C. During model fusion, the central node does not consider the parameters fed back by user A when updating the parameters of network layer g. That is, the fusion process can be expressed using the following equation:

number

[0083] M 0,g represents the fused model parameters, and M 1,g , M 2,g , and M 3,g represents the gth network layer of the user's model, and β k,g ∈{0,1} reflects the update instruction corresponding to the gth network layer of the model of the kth child node, specifically, β k,g = 1 indicates that the parameters of the gth network layer of the model are updated, and β k,g = 0 indicates that the parameters of the gth network layer of the model are frozen, and only two users are responsible for updating the parameters of the network layer, so

number

[0084] The communication method provided in the embodiments of the present application may be applied to various communication systems, such as a long term evolution (LTE) system, a fifth generation (5G) mobile communication system, a wireless fidelity (Wi-Fi) system, a future communication system, or a system integrating multiple communication systems. This is not limited to the embodiments of the present application. 5G is also sometimes called new radio (NR).

[0085] The communication methods provided in the embodiments of the present application may be applicable to various communication scenarios, for example, one or more of the following communication scenarios: enhanced mobile broadband (eMBB), ultra reliable low latency communication (URLLC), machine type communication (MTC), massive machine type communication (mMTC), device to device (D2D), vehicle to everything (V2X), vehicle to vehicle (V2V), and internet of things (IoT).

[0086] In the following, the communication system shown in FIG. 7 is used as an example to describe the communication method provided in the embodiment of the present application.

[0087] FIG. 7 is a diagram of a communication system according to an embodiment of the present application. As shown in FIG. 7, the communication system may include a network device and multiple terminals. The network device corresponds to a central node in the federated learning architecture, and the terminals correspond to child nodes in the federated learning architecture. The network device distributes a global model to each terminal participating in the federated learning. Each user trains the global model using local data to obtain a local model, and transmits parameter information of the local model, such as gradients and weight values (after encryption), to the network device. The network device performs model aggression (MA) to update the global model and then transmits the updated global model to each user. The users continue to update their local models, output parameter information of the updated local models, and transmit the parameter information to the network device. In this manner, multiple iterations are performed until the model converges.

[0088] It should be noted that Figure 7 is merely an exemplary framework diagram, and the number of network devices and terminals included in Figure 7 is not limited. In addition to the functional nodes shown in Figure 7, other nodes such as core network devices, gateway devices, and application servers may also be included. This is not limited. The network devices and core network devices communicate with each other in a wired or wireless manner, for example, through a next generation (NG) interface.

[0089] The terminal device in the embodiments of the present application may be referred to as a terminal and may be a device having a wireless transceiver function. The terminal may be deployed on land, including an indoor device, an outdoor device, a handheld device, and / or an in-vehicle device, on water (e.g., on a ship), or in the air (e.g., on an airplane, a balloon, or a satellite). The terminal device may be user equipment. The terminal device may include a handheld device, an in-vehicle device, a wearable device, or a computing device having a wireless communication function. For example, the terminal device may be a mobile phone, a tablet computer, or a computer having a wireless transceiver function. Alternatively, the terminal device may be a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, an industrial control wireless terminal, an autonomous driving wireless terminal, a telemedicine wireless terminal, a smart grid wireless terminal, a smart city wireless terminal, a smart home wireless terminal, etc.

[0090] In the embodiments of the present application, an apparatus configured to perform the functions of a terminal device may be a terminal device, or may be an apparatus capable of supporting a terminal device in performing the functions, such as a chip system. The apparatus may be mounted on a terminal device or used together with a terminal device. In the embodiments of the present application, the chip system may include a chip, or may include a chip and other discrete components. In the following embodiments, the technical solutions provided in the embodiments of the present application are described using an example in which an apparatus configured to perform the functions of a terminal device is a terminal device, and the terminal device is a UE.

[0091] The network device in the embodiment of the present application may include a base station (BS) and may be a device deployed in a radio access network (RAN) and capable of communicating with terminal devices. Optionally, the radio access network may be abbreviated as an access network. The network device may also be referred to as an access network device. The base station may take multiple forms, for example, a macro base station, a micro base station, a relay station, or an access point. The base station in the embodiment of the present application may be a base station of a 5G system, a base station of a Long Term Evolution (LTE) system, or a base station of another system. This is not limited thereto. The base station in the 5G system may also be referred to as a Transmission Reception Point (TRP) or a Next Generation NodeB (gNB). The base station in the embodiment of the present application may be an integrated base station or a base station including a central unit (CU) and distributed units (DUs). A base station including a CU and a DU may also be referred to as a base station in which the CU and the DU are separated. For example, a base station includes a gNB-CU and a gNB-DU. The CU may be further separated into a CU Control Plane (CU-CP) and a CU User Plane (CU-UP). For example, a base station includes a gNB-CU-CP, a gNB-CU-UP, and a gNB-DU.

[0092] In the embodiments of the present application, the apparatus configured to implement the functions of a network device may be a network device, or may be an apparatus capable of supporting a network device in implementing functions, such as a chip system. The apparatus may be installed in a network device or used together with a network device. In the following embodiments, an example in which the apparatus configured to implement the functions of a network device is a network device and the network device is a base station is used to describe the technical solutions provided in the embodiments of the present application.

[0093] To support machine learning functions in a wireless network, a dedicated AI network element or module may be further introduced into the network. When an AI network element is introduced, the AI network element corresponds to an independent network element. When an AI module is introduced, the AI module may be located within a specific network element, and the corresponding network element may be a terminal device, a network device, etc.

[0094] In the technical solutions provided in the embodiments of the present application, the apparatus configured to perform the functions of a network device is a base station, and the apparatus configured to perform the functions of a terminal device is a terminal. Using an example, the methods provided in the embodiments of the present application are described below with reference to a beam selection scenario.

[0095] 8 is a schematic flowchart of a model training method according to an embodiment of the present application. As shown in FIG. 8, the method may include the following steps:

[0096] S801: A plurality of child node devices transmit capability information to a central node device, and the central node device receives capability information from the plurality of child node devices accordingly.

[0097] The capability information of the multiple child node devices represents the capability of training a model by the multiple child node devices, and the multiple child node devices are multiple child node devices that participate in federated learning of the model.

[0098] Multiple child node devices are communicatively connected to a central node device. In this embodiment of the present application, the child node device's transmission of information to the central node device is referred to as reporting, and the central node device's transmission of information to the child node device is referred to as distributing. The central node configures downlink resources to be used by the central node to distribute the first indication information and the model to the child nodes. For example, in a beam selection scenario, the downlink resources may be control channel resources, such as physical downlink control channel (PDCCH) resources, or data channel resources, such as physical downlink shared channel (PDSCH) resources. Specifically, the downlink resources include parameters such as frequency domain resource block numbers and starting positions, subband numbers, subband bandwidths, frequency hopping parameters, and modulation and coding schemes (MCSs).

[0099] The model may be distributed by the central node in broadcast or multicast mode. An example is shown below: in a single-cell federated learning architecture where the central node is a base station and the child nodes are UEs, the model may be distributed in broadcast mode. Due to the broadcast characteristics, child nodes that do not participate in federated learning can also receive broadcast information. In a multi-cell federated learning architecture where a base station with federated learning management functionality is used as the central node and other base stations are used as child nodes, the central node can also distribute the model to each child node base station in broadcast mode. Similarly, other child nodes that do not participate in federated learning can also receive broadcast information. Alternatively, multicast mode may be used for child nodes participating in federated learning, where child nodes associated with the same central node are classified into one group, have the same group number, and are configured with the same downlink resources. In multicast mode, child nodes that do not participate in federated learning do not receive multicast information.

[0100] The central node may further configure, to the child node, uplink resources to be used by the child node to report the updated first model parameters. Alternatively, a separate federated learning management node may configure, to the central node and the child node, uplink resources to be used by the child node to report the updated first model parameters and necessary signaling. Similar to the downlink resource configuration, the uplink resources may be control channel resources, for example, physical uplink control channel (PUCCH) resources, or data channel resources, for example, physical uplink shared channel (PUSCH) resources.

[0101] In this embodiment of the present application, a child node needs to send capability information to a central node device on an access network. The capability information needs to include at least the size of memory space available for storing AI / ML models by the child node device, the child node's computing power information (the computing power for executing the AI / ML model, e.g., the FLOPS computing performance and power status mentioned above; current power is also one of the factors affecting the computing power of a terminal device), and collection information related to local data volume (which can help the central node predict the duration required for model training). Other optional capability information may include whether the child node device supports the execution of AI / ML models and the supported AI / ML model types (e.g., CNN, RNN, fully connected model, or random forest model). The central node allows the child node to participate in federated learning and transmits the model to the child node only if the child node supports the AI / ML model. Other optional capability information may include the child node's hardware information, such as antenna configuration (number of antennas, polarization direction, etc.), number of radio frequency channels, sensor types (position sensor / GPS, motion sensor, etc.), and child node parameters. The model is trained in a federated learning framework, and child nodes perform training using their local data, so child nodes do not need to report information related to the beam information actually collected or privacy-related information.

[0102] S802: The central node device sends first indication information and a model to each of the plurality of child node devices, and the plurality of child node devices receive the first indication information and the model from the central node device accordingly.

[0103] The first instruction information indicates first model parameters to be updated when the multiple child node devices train the model, and the first model parameters to be updated of the multiple child node devices are determined based on capability information of the multiple child node devices.

[0104] Model transmission includes the transmission of model structure and model parameters. The model structure is described by, for example, the number of convolutional layers in a convolutional neural network (CNN), the number of channels and the size of the convolution kernel in each convolutional layer, the number of fully connected layers and the number of neurons, the number of layers in a recurrent neural network (RNN), and the calculation method for each state in each layer of the structure.

[0105] The first indication information may be distributed by a signaling method. Specifically, the following three methods may be selected: 1. The first indication information is transmitted to a designated child node in a unicast mode; 2. If different child nodes correspond to the same first indication information, the same first indication information may be transmitted to a group of child nodes in a multicast mode; 3. The first indication information is distributed in a broadcast mode. Each signaling includes a child node index corresponding to the signaling. After receiving the broadcast signal, each child node uses the signaling corresponding to the child node index.

[0106] In addition, the central node can further distribute information related to the child nodes' model training and reporting, such as the number of iterations required by the child nodes in each round, the learning rate, the loss function, the batch size, and the type of parameters to be reported (model parameters or gradients). To reduce the air interface overhead for distributing the model, the model may be compressed before being distributed. Model compression methods include, but are not limited to, model pruning, model distillation, model quantization, etc.

[0107] S803: To obtain updated first model parameters, multiple child node devices train a model based on the training set and the model parameters to be updated.

[0108] The training set includes beam information collected by multiple child node devices. In a beam management scenario, in federated learning, the child nodes may be base stations or user terminal devices, and the central node may be an independent federated learning management node or a base station functioning as the central node. The global model to be trained is assumed to be an AI / ML model that uses estimated channel measurements or received signals as input and an optimal beam index (ID) as output. In this case, the child node is responsible for collecting the training set as model input during the data collection phase. The training set may include the channel measurements or received signals and a label (i.e., the optimal beam ID) used to train the model. The label can be understood as follows: the base station transmits all possible beams (codebook-based SSB or CSI-RS beams) to the UE one by one, and the UE selects the beam with the best performance and uses the index of that beam as the label (the beam with the best performance may be the beam with the largest L1-RSRP or SNR measurement among the SSB / CSI-RS beams).

[0109] During model training, if a child node device has frozen parameters, gradient-free (required_grad=False) processing may be performed on the frozen parameters in the corresponding AI / ML network optimizer to prevent the frozen parameters from participating in training.

[0110] The loss function used for training is related to the application scenario and the type of model used. For example, in this embodiment of the present application, the Cross-Entropy (CE) function may be used in the corresponding beam management classification problem (to predict the optimal beam ID), the Binary Cross-Entropy (BCE) function may be used in the receiver performance / data demodulation performance problem, and the Mean Squared Error (MSE) function or the Mean Absolute Error (MAE) function may be used in the regression problem such as channel estimation. The child node may select the loss function based on the application scenario, or may select the loss function according to instructions distributed by the central node. Specifically, the child node can use the received global model as the initial model of this round for training. The loss function is mainly related to the local dataset. For example, the MSE loss function of the kth child node is

number

number

[0111] S804: A plurality of child node devices send updated first model parameters to the central node device, and the central node device receives updated first model parameters from the plurality of child node devices accordingly.

[0112] The child nodes report the updated first model parameters using the uplink resources described in step S801. Each child node may report all parameters of the beam-trained model, or may report only the first model parameters that participate in the update. For the case where all parameters of the beam-trained model are reported, for example, if the local model has a total of G parameters and the kth child node reports M k =[m k,1 ,m k,2 ,…,m k,G ], and m k,g represents the gth parameter of the kth child node. All parameters may be combined and sent as a vector. Alternatively, a set of differences between the current model parameters and those of the previous round is reported, i.e., the parameter gradient ΔM k =[Δm k,1 ,Δm k,2 ,…,Δm k,G ] is reported. The order in which the updated first model parameters are reported matches the order in which the first model parameters are delivered. This reporting method is applicable to both parameter reporting at the network layer granularity and parameter reporting at the single parameter granularity, and there is no need to configure different reporting methods for the two parameter modes.

[0113] In consideration of saving air interface overhead, if only the first model parameter participating in the update is reported, the child node reports the first model parameter of the child node in the local model using the uplink resource configured in S801. For example, assume that the model has a total of G parameters, and G' parameters are the first model parameters of the kth child node. In this case, the kth child node reports the first model parameter of the local model M. k =[m k,1 ,m k,2 ,…,m k,G ] and m k,grepresents the gth parameter of the kth child node. All parameters may be combined and sent as a vector (or a set of differences between the current model parameters and the model parameters of the previous round may be reported, i.e., the parameter gradient ΔM k =[Δm k,1 ,Δm k,2 ,…,Δm k,G ] will be reported).

[0114] In addition, the child node is assigned a number n of samples in the child node's local dataset. k It can also report auxiliary information such as the model type indicator.

[0115] S805: The central node device fuses global model parameters of multiple child node devices using a predefined fusion algorithm to obtain fused parameters.

[0116] The global model parameters include the updated first model parameters and second model parameters other than the first model parameters in the model. A default fusion algorithm may be used to fuse the global model parameters of the multiple child node devices to obtain fused parameters. For example, the default fusion algorithm may be selected to obtain a mean value, a median value, a mode value, etc.

[0117] S806: The central node device sends the fused parameters to the multiple child node devices, and the multiple child node devices receive the fused parameters accordingly.

[0118] S807: The multiple child node devices train a model based on the fused parameters.

[0119] The multiple child node devices train a model based on the fused parameters until the model trained based on the fused parameters converges. If not, S802 to S807 are executed after returning to S802 to adjust the first instruction information until the model trained based on the fused parameters converges. Whether the model has converged may be determined using a loss function. See the corresponding description of S803, the loss function may be a cross-entropy (CE) function.

[0120] In this embodiment of the present application, first, multiple child node devices send capability information to the central node device, and then the central node device sends first instruction information and a model to each of the multiple child node devices based on the capability information. Then, the multiple child node devices train a model based on the training set and the model parameters to be updated to obtain updated first model parameters. In addition, the central node device uses a predetermined fusion algorithm to fuse global model parameters of the multiple child node devices, including the first model parameters, to obtain fused parameters, and then the multiple child node devices train a model based on the fused parameters. In this way, when the child node devices train a model, some parameters are frozen and do not participate in the training. Therefore, during model training, the progress of the model training is not affected by the central node device with low computing power, and model training efficiency is high.

[0121] In one embodiment, the method further comprises the following steps:

[0122] S808: The central node device separately predicts model training periods for the multiple child node devices based on the capability information.

[0123] The central node device needs to separately predict the model training periods of multiple child node devices based on the capability information to determine whether the child node devices can train the model. The concept of a time window may be used, specifically, the child node k is required to be able to complete the model training within a given time window range (the time window may be a time window having a length T0 from time t after the model is broadcast to time t+T0, where t≧0), and the child node needs to feed back the parameters / gradients of the local model within this time window range. The time T used by the child node device k to train the model is k Assume that can be expressed as follows:

number

[0124] α k contains all hyperparameters that can affect the training duration (which may be pre-configured by the central node in the first round of training or provided by child nodes after the first round). y=f() means that the hyperparameters are mapped, which is equivalent to training y models. FLOPS represents the complexity of the model and is determined by the model type of the model. FLOPs k represents the computing power performance of the child node device k (determined based on the capability information). The model training period of multiple child node devices can be predicted using the above formula.

[0125] S809: The central node device determines first model parameters of the child node device based on the model training period.

[0126] Referring to the formula in the description of S808, after the model training period of the child node device is predicted, the first model parameter of the child node device can be determined based on the model training period. kIf ≧T0 (where k=0,...,K), specifically, if the prediction model training period of child node device k exceeds the predetermined time window range T0, this means that child node device k cannot train the model, or that child node device k can train the model but the training period exceeds the time window range T0.

[0127] In order to allow all child nodes to train models and complete training tasks within a given time window, in this embodiment of the present application, for child node devices with insufficient capabilities, model training is performed according to a policy of training part of the model parameters instead of all the model parameters. Correspondingly, the central node needs to determine the first model parameters of the child node devices based on the reported model training period, that is, the model parameters that will participate in training are called update parameters, and the remaining model parameters that will not participate in training are called frozen parameters.

[0128] For example, if the first model parameters are selected with a single parameter granularity, then the model M0 will have a total of G parameters, and M 0,g Let denote the gth parameter. In this case, the set of all parameters corresponding to the model of the kth child node is φ Mk ={M 1,1 ,M 1,2 ,M 1,3 ,…,M 1,G}, and M 1,g denotes the gth parameter in the parameter set. The possible sets that indicate whether a parameter participates in training can be expressed as follows:

number

[0129] Any element β k is a 1-by-G vector, and the elements of the vector are {βk,1 ,β k,2 ,…,β k,G}. Any element β k,g ={0,1} indicates whether the gth parameter of the model corresponding to the kth child node needs to be updated / frozen. k,g = 0 indicates that the parameter is a frozen parameter for the child node, and β k,g =1 indicates that the parameter is an update parameter for the child node. T represents the transpose operation. Similarly, if the freezing / updating is done based on each network layer, the index g above represents all the parameters of the gth network layer of the model, and β k,g ={0,1} indicates whether all parameters in the gth network layer of the model corresponding to the kth child node need to be updated / frozen.

[0130] Specifically, we use a layer-based freezing / updating model as an example to explain how parameters are classified as frozen parameters and updated parameters. First, freezing more variables can save more computational power, but at the cost of reduced model performance. In addition, model performance will be affected if update layers are unevenly allocated to all child nodes participating in federated learning. A possible classification rule is as follows: The number of layers that need to be frozen is gradually increased in ascending order based on the capabilities reported by the terminal until the terminal's memory constraints are met and the local model training and upload can be completed within the specified time window. In addition, to increase terminal differentiation and reduce the probability of uneven update layer allocation among child nodes, the number of terminals participating in one round of federated learning is increased as much as possible. Rule design is primarily implemented in the base station. Better optimization / rule design algorithms can effectively improve model performance.

[0131] In this embodiment of the present application, the model training periods of multiple child node devices are first predicted separately based on capability information, and then the first model parameters of the child node devices are determined based on the model training periods, so that the determined first model parameters to be trained can match the corresponding child node devices and ensure the progress of model training.

[0132] In one embodiment, S808 specifically includes the step of the central node device separately predicting the model training periods of multiple child node devices based on the capability information.

[0133] S8081: If the model training period does not satisfy the predetermined period condition, select some of the model parameters as the first model parameters.

[0134] S8082: If the model training period meets the predetermined period condition, select all parameters of the model as the first model parameters.

[0135] In this embodiment of the present application, based on different comparison results between the model training period and the predetermined period conditions, two policies are used accordingly, namely, a policy of selecting part of the parameters of the model as the first model parameters and a policy of selecting all the parameters of the model as the first model parameters, so that the progress of model training is guaranteed and at the same time the model performance is guaranteed.

[0136] In one embodiment, in S8081, specifically, the step of selecting a part of the parameters of the model as the first model parameters includes: The method may include randomly selecting some of the parameters of the model as first model parameters.

[0137] Randomly selecting some of the model parameters as the first model parameters is equivalent to selecting a parameter with the unity of a single parameter. For details, see the description of selecting a first model parameter with the granularity of a single parameter in S809. The details will not be described again.

[0138] In this embodiment of the present application, a portion of the model parameters is randomly selected as the first model parameters, so that the determined first model parameters to be trained can match with the corresponding child node devices, and the progress of model training is guaranteed.

[0139] In one embodiment, the model includes a first network layer and a second network layer, the model training scenario indicates the first network layer, and the first network layer and the second network layer are different. S8081 of selecting a part of the parameters of the model as the first model parameters includes: The method may include selecting parameters of the first network layer and some of the parameters of the second network layer as first model parameters.

[0140] In this embodiment of the present application, selecting the first model parameters based on the network layer corresponds to parameter selection at the granularity of the network layer. For example, the first network layer is a fully connected layer. In a network with multiple fully connected layers, if each child node user updates parameters corresponding to a portion of the neurons in each layer in a distributed manner, it may be difficult to guarantee the final training effect of the network. Therefore, in this embodiment of the present application, all parameters of a specific first network layer are selected as the first model parameters to guarantee model performance.

[0141] Considering that the first network layer may include multiple network layers, when selecting parameters of the first network layer, the central node device can determine the first network layer corresponding to the child node device based on the capability information of the child node device. To ensure model training accuracy, when determining the first network layers corresponding to the multiple child node devices, one set of first network layers corresponding to all the child node devices covers all the first network layers.

[0142] In addition, the second network layer (e.g., a non-fully connected layer or a fully connected layer that does not require updates directed by child node devices) may be selected in a random selection manner. Because the channel conditions within the same cell are fixed, to reduce the overhead of computational power, a fully connected layer trained by some users and responsible for finer granularity may also be used by other users with lower computational power in the implementation phase.

[0143] In addition, in practice, the proportion of parameters occupied by some network layers can be much higher than that of other types of network layers. For example, as shown in Figure 9, an AI / ML model is formed by serially connecting convolutional layers (Conv1, Conv2) with feature extraction functions to fully connected layers (FC1 16 Neural, FC2 64 Neural) with classification functions. In this model, the parameters of the fully connected layers account for a large proportion of all the model's parameters. In addition, the same task may have different classification granularity. For example, in an AI-assisted beam management task, if the base station initiates SSB wide beam sweeping based on a 16-codeword codebook, some user equipment's computing power may support predicting the optimal beam ID for 16 wide beams, while some other user equipment's computing power may support direct prediction of the optimal beam ID for 64 narrow beams. Therefore, at least two fully connected layers with different numbers of neurons (16 neurons and 64 neurons) are required for two types of users to complete the task.

[0144] Correspondingly, in a parameter selection scenario at the network layer granularity, the model parameters can be expressed as:

number

[0145] β nFC represents the parameter set corresponding to the neural network parameters of the layer before all fully connected layers, and any element in the set

number

number

number

[0146] This means that there are three fully connected layers in total, and the kth child node is responsible for updating all the network parameters of the first two fully connected layers. Correspondingly, the first model parameters corresponding to the kth child node can be expressed as follows:

number

[0147] Random() represents a partial selection function.

[0148] Referring to Figure 9, during loss function calculation, each child node calculates the loss function using the output of the corresponding first network layer. Continuing to use the beam management task as an example, the fully connected layer corresponds to beam ID labels with three resolutions: {16, 64, 128}. Suppose the kth child node corresponds to a 64-beam prediction task. Instead of determining the cross-entropy based on the final network output (the cross-entropy between the label and the output of the 128-neuron fully connected layer), the kth child node's loss function determines the cross-entropy between the label and the output obtained after a softmax operation is performed on the output of the 64-neuron fully connected layer (CE Loss). Similarly, this applies to the child nodes corresponding to the 16-beam and 128-beam prediction tasks.

[0149] In this embodiment of the present application, the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, so that the requirements of the model application scenario for the network layer are taken into consideration and the model training performance is guaranteed.

[0150] In one embodiment, before the selected parameters of the default fully connected layer in the fully connected layer are selected as the first model parameters, the method may further include the following steps:

[0151] S810: The central node device determines, based on a model application scenario, a first network layer required to construct a model.

[0152] Depending on the various model application scenarios, the model may also need to have specific network layers. For example, in a beam management task, assuming that a central node device has 64 antennas in a high-frequency millimeter-wave downlink system, each child node device can complete the following tasks: predicting the optimal beam among 16 wide beams based on a 16-DFT codebook; directly predicting the ID of the optimal beam among 64 narrow beams based on a 16-DFT wide beam codebook; and more accurate beam ID prediction based on a steering vector without a codebook (e.g., default prediction with an accuracy of 128 or higher without an orthogonal codebook). In this case, the last fully connected layer in the distributed global model must have 16, 64, or 128 / 256 neurons. In other words, the model application scenario requires a fully connected layer with 16, 64, or 128 / 256 neurons, and the fully connected layer with 16, 64, or 128 / 256 neurons is the first network layer required to build the model.

[0153] In this embodiment of the present application, at least one fully connected layer required to build a model is determined as a default fully connected layer corresponding to a child node device, so that the requirements of the model application scenario for the network layer are taken into consideration and the model training performance is guaranteed.

[0154] In one embodiment, the capability information further represents a type of model that can be trained by the child node device, and the method further comprises: The method further includes a step in which the central node device determines a child node device whose capability information includes a model type of the model as a target device to which the first instruction information and the model are to be transmitted.

[0155] In this embodiment of the present application, a child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, thereby ensuring that all child node devices participating in the training can train the model and ensuring the progress of model training.

[0156] In one embodiment, S805 specifically includes the step of the central node device fusing global model parameters of multiple child node devices using a default fusion algorithm to obtain fused parameters, which may include the following steps: S8051: For model parameters having the same function of different child node devices, the central node device sets the weight of the second model parameter to a first default weight value, sets the weight of the first model parameter to a second default weight value, and performs weighted averaging to obtain fused parameters.

[0157] The function of the first default weight value is to reduce the influence of the second model parameter on the fused parameter as much as possible, and the function of the second default weight value is to allow the parameter assigned with the weight value to influence the fused parameter. For example, the first default weight value may be selected as 0, and the second default weight value may be selected as 1. Correspondingly, the calculation process of the gth fused parameter can be expressed using the following formula:

number

[0158] U∈Z + where U≦K, and U represents the number of first model parameters. See the explanation of S809, and β k,g = 0 means that the parameter is a frozen parameter for the child node, and β k,g = 1 means that the parameter is an update parameter for the child node. Both the first and second default weight values function as β k,g This is achieved by:

[0159] In this embodiment of the present application, the second model parameters are designed to be unable to affect the default fusion algorithm of the fused parameters and unable to participate in the calculation of the fused parameters, so that the second model parameters that are not updated cannot affect the model training, and the model training performance is guaranteed.

[0160] In one embodiment, S805, specifically, the step of the central node device fusing global model parameters of multiple child node devices using a predefined fusion algorithm to obtain fused parameters, may instead include the following steps: S8052: For model parameters having the same function of different child node devices, the central node device sets both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and performs weighted averaging to obtain fused parameters.

[0161] Referring to the description of the second default weight value in S8051, in this embodiment of the present application, the weight of the first model parameter and the weight of the second model parameter are both set to the second default weight value, so that both the first model parameter and the second model parameter affect the fused parameter. For example, the second default weight value may be set to 1. Correspondingly, the calculation process of the gth fused parameter can be expressed using the following formula:

number

[0162] The above describes the model training method provided in this embodiment of the present application from the perspective of exchanging model parameters between the central node device and the child node device. When the central node device and the child node device exchange model gradients, the fused parameters can be calculated by adding the model gradients to the model parameters of the previous round, which can be expressed using the following formula:

number

[0163] ΔM 0,g represents the fused model gradient, which corresponds to the fused parameters, and ΔM k,g represents the model gradient, which corresponds to the model parameters. To perform model training, we can also use the above formula to calculate the fused parameters.

[0164] In this embodiment of the present application, the second model parameters are further designed to be able to influence the default fusion algorithm of the fused parameters and participate in the calculation of the fused parameters, so that the determined fused parameters can participate in model training and the progress of model training is guaranteed.

[0165] The above describes a model training method provided in an embodiment of the present application from the perspective of model training. After a model is obtained through training by the above method, the trained model can be used for beam selection. The following describes a beam information selection method provided in an embodiment of the present application from the perspective of model application. The method may be performed by a child node device. As shown in FIG. 10, the method may include the following steps:

[0166] S101: Receive beam information.

[0167] The child node device can receive beam information from the network device.

[0168] S102: Input beam information into the model and output target beam information.

[0169] The model is obtained through training by the model training method described in the previous embodiment.

[0170] In this embodiment of the present application, a model is used to select the target beam, and since the training efficiency of the model is high, the efficiency of selecting the target beam is high.

[0171] The above mainly describes the solutions provided in the embodiments of the present application from the perspective of interactions between nodes. It should be understood that, to implement the aforementioned functions, nodes such as child node devices and central node devices include hardware structures and / or software modules corresponding to those functions. Those skilled in the art will easily realize that, in combination with the example algorithm steps described in the embodiments disclosed herein, the methods of the embodiments of the present application can be implemented by hardware, software, or a combination of hardware and computer software. Whether the functions are performed by hardware or by hardware driven by computer software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described functions using various methods for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0172] In the embodiment of the present application, the child node device and the central node device may be divided into functional modules based on the above-described exemplary method. For example, each functional module may be obtained by dividing the functional modules based on their corresponding functions, or two or more functions may be integrated into one processing module. The integrated module may be implemented in the form of hardware or a software functional module. It should be noted that the division into modules in the embodiment of the present application is merely an example and is merely a logical division of functions, and may be divided differently in actual implementation.

[0173] In specific implementation, network elements shown in the present application, such as a child node device or a central node device, may use the composition structure shown in FIG. 11 or may include the components shown in FIG. 11. FIG. 11 is a diagram of the structure of a communication device 1100 according to an embodiment of the present application. If the communication device 1100 has the functionality of a child node device described in the embodiments of the present application, the communication device 1100 may be a chip or a system-on-chip within the child node device. If the communication device 1100 has the functionality of a central node device described in the embodiments of the present application, the communication device 1100 may be a chip or a system-on-chip within the central node device.

[0174] 11, a communications device 1100 may include a processor 1101, a communications line 1102, and a transceiver 1103. The processor 1101, memory 1104, and the transceiver 1103 may be connected through the communications line 1102. In one example, the processor 1101 may include one or more CPUs, such as CPU 0 and CPU 1 of FIG.

[0175] In an optional implementation, the communications device 1100 includes multiple processors. For example, the communications device 1100 may further include a processor 1107 in addition to the processor 1101 of FIG.

[0176] The processor 1101 may be a central processing unit (CPU), a general-purpose processor, a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. Alternatively, the processor 1101 may be another device, such as a circuit, component, or software module, having processing capabilities.

[0177] The communication lines 1102 are configured to transmit information between components included in the communication device 1100 .

[0178] The transceiver 1103 is configured to communicate with another device or another communication network. The other communication network may be an Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc. The transceiver 1103 may be an interface circuit, a pin, a radio frequency module, a transceiver, or any device capable of implementing communications.

[0179] Additionally, the communications device 1100 may include a memory 1104. The memory 1104 is configured to store instructions. The instructions may be a computer program.

[0180] Memory 1104 may be read-only memory (ROM) or another type of static storage device capable of storing static information and / or instructions, or random access memory (RAM) or another type of dynamic storage device capable of storing information and / or instructions, or may be electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disk storage, optical disk storage, magnetic disk storage media or other magnetic storage device. Optical disk storage includes compressed optical disks, laser disks, optical disks, digital versatile disks, Blu-ray discs, etc.

[0181] It should be noted that the memory 1104 may exist independently of the processor 1101 or may be integrated with the processor 1101. The memory 1104 may be configured to store instructions, program codes, some data, etc. The memory 1104 may be located inside the communication device 1100 or outside the communication device 1100. This is not limited thereto. Executing the instructions stored in the memory 1104 enables the processor 1101 to perform the methods provided in the embodiments of the present application.

[0182] In an optional implementation, the communication apparatus 1100 further includes an output device 1105 and an input device 1106. For example, the input device 1106 is a keyboard, a mouse, a microphone, a joystick, or other device, and the output device 1105 is a display, a speaker, or other device.

[0183] It should be noted that the communications device 1100 may be a desktop computer, a portable computer, a network server, a mobile phone, a tablet computer, a wireless terminal, an embedded device, a chip system, or a device having a structure similar to that of Figure 11. Additionally, the compositional structure shown in Figure 11 does not constitute a limitation on the communications device. In addition to the components shown in Figure 11, the communications device may include more or fewer components than those shown in the figure, or some components may be combined, or a different component arrangement may be used.

[0184] In this embodiment of the present application, a chip system may include a chip, or may include a chip and other discrete components.

[0185] FIG. 12 is a diagram of the configuration of a communication device 120. This communication device is used in a central node device. The modules in the device shown in FIG. 12 have functions for implementing corresponding steps in FIG. 8 and can achieve corresponding technical effects. For beneficial effects corresponding to steps performed by the modules, please refer to the descriptions of the corresponding steps in FIG. 8. Details will not be described again. Those functions may be implemented by hardware, or may be implemented by hardware by executing corresponding software. The hardware or software includes one or more modules corresponding to those functions. The communication device may be a chip or a system-on-chip in a central node device. For example, the communication device may include: The federated learning system includes: a receiving module 121 configured to receive capability information representing a capability of training a model by a child node device from a plurality of child node devices (child node devices participating in the federated learning of the model); a transmitting module 122 configured to transmit to each of the plurality of child node devices the model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters to be updated of the child node device are determined based on the capability information of the child node device, and the receiving module 121 is further configured to receive the updated first model parameters from the plurality of child node devices; and a processing module 123 configured to fuse global model parameters of the plurality of child node devices using a predefined fusion algorithm to obtain fused parameters, wherein the global model parameters include the updated first model parameters and second model parameters other than the first model parameters in the model, and the transmitting module 122 is configured to transmit the fused parameters to the plurality of child node devices.

[0186] In one possible implementation, the processing module 123: Separately predicting model training periods for the plurality of child node devices based on the capability information, and determining first model parameters for the child node devices based on the model training periods. It is specifically structured as follows.

[0187] In one possible implementation, the processing module 123: Selecting some of the parameters of the model as the first model parameters when the model training period does not satisfy a predetermined period condition, or selecting all of the parameters of the model as the first model parameters when the model training period satisfies a predetermined period condition. It is specifically structured as follows.

[0188] In one possible implementation, the processing module 123: Randomly select some of the model's parameters as the first model parameters It is specifically structured as follows.

[0189] In one possible implementation, the model includes a first network layer and a second network layer, the model application scenario indicates the first network layer, the first network layer and the second network layer are different, and the processing module 123 is specifically configured to select parameters of the first network layer and some of the parameters of the second network layer as first model parameters.

[0190] In one possible implementation, the processing module 123: Before selecting the default fully connected layer parameters in the fully connected layer as the first model parameters, determine the first network layer required to build the model based on the model application scenario. It is further configured as follows.

[0191] In one possible implementation, the capability information further indicates the type of model that can be trained by the child node device, and the processing module 123: A child node device whose capability information includes a model type of the model is determined as a target device to which the first instruction information and the model are to be transmitted. It is further configured as follows.

[0192] In one possible implementation, the processing module 123: For model parameters having the same function of different child node devices, set the weight of the second model parameter to a first default weight value, set the weight of the first model parameter to a second default weight value, and perform weighted averaging to obtain a fused parameter; or, for model parameters having the same function of different child node devices, set both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and perform weighted averaging to obtain a fused parameter. It is specifically structured as follows.

[0193] In this embodiment of the present application, first, multiple child node devices send capability information to the central node device, and then the central node device sends first instruction information and a model to each of the multiple child node devices based on the capability information. Then, the multiple child node devices train a model based on the training set and the model parameters to be updated to obtain updated first model parameters. In addition, the central node device uses a predetermined fusion algorithm to fuse global model parameters of the multiple child node devices, including the first model parameters, to obtain fused parameters, and then the multiple child node devices train a model based on the fused parameters. In this way, when the child node devices train a model, some parameters are frozen and do not participate in the training. Therefore, during model training, the progress of the model training is not affected by the central node device with low computing power, and model training efficiency is high.

[0194] Furthermore, the model training periods of multiple child node devices are first predicted separately based on the capability information, and then the first model parameters of the child node devices are determined based on the model training periods, so that the determined first model parameters to be trained can match the corresponding child node devices, and the progress of the model training is guaranteed.

[0195] Furthermore, based on different comparison results between the model training period and the predetermined period conditions, two policies, namely, a policy of selecting part of the model's parameters as the first model parameters and a policy of selecting all of the model's parameters as the first model parameters, are used accordingly, so that the progress of model training is guaranteed and the model performance is guaranteed at the same time.

[0196] Furthermore, since a portion of the model parameters is randomly selected as the first model parameters, the determined first model parameters to be trained can match with the corresponding child node devices, and the progress of model training is guaranteed.

[0197] Furthermore, since the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, the requirements of the model application scenario for the network layers are taken into consideration and the model training performance is guaranteed.

[0198] Furthermore, the central node device determines the first network layer required to build the model based on the model application scenario, so that the requirements of the model application scenario for the network layer are taken into consideration and model training performance is guaranteed.

[0199] Furthermore, since the child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, it is guaranteed that all child node devices participating in the training can train the model, and progress in model training is guaranteed.

[0200] Furthermore, the second model parameters are designed to be unable to affect the default fusion algorithm of the fused parameters and unable to participate in the calculation of the fused parameters, so that the second model parameters that are not updated cannot affect the model training, and the model training performance is guaranteed.The second model parameters are further designed to be able to affect the default fusion algorithm of the fused parameters and able to participate in the calculation of the fused parameters, so that the determined fused parameters can participate in the model training, and the progress of the model training is guaranteed.

[0201] FIG. 13 is a diagram of the configuration of a communication device 130. This communication device is used in a child node device. The modules in the device shown in FIG. 13 have functions to implement corresponding steps in FIG. 8 and can achieve corresponding technical effects. For beneficial effects corresponding to steps performed by the modules, please refer to the descriptions of the corresponding steps in FIG. 8. Details will not be described again. Those functions may be implemented by hardware, or may be implemented by hardware by executing corresponding software. The hardware or software includes one or more modules corresponding to those functions. The communication device may be a chip or a system-on-chip in a child node device. For example, the communication device may include: a sending module 131 configured to send capability information representing a capability of training a model by a child node device to a central node device; a receiving module 132 configured to receive, from the central node device, a model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters to be updated of the child node device are determined based on capability information of the child node device; a processing module 133 configured to train a model based on a training set (including beam information collected by the child node device) and the model parameters to be updated to obtain updated first model parameters; Including, the transmitting module 131 is configured to transmit the updated first model parameters to the central node device; the receiving module 132 is configured to receive, from the central node device, fused parameters calculated based on the updated first model parameters; The processing module 133 is configured to train a model based on the fused parameters; The receiving module 132 is configured to receive beam information; The processing module 133 is configured to input beam information into the model and output target beam information.

[0202] In this embodiment of the present application, first, multiple child node devices send capability information to the central node device, and then the central node device sends first instruction information and a model to each of the multiple child node devices based on the capability information. Then, the multiple child node devices train a model based on the training set and the model parameters to be updated to obtain updated first model parameters. In addition, the central node device uses a predetermined fusion algorithm to fuse global model parameters of the multiple child node devices, including the first model parameters, to obtain fused parameters, and then the multiple child node devices train a model based on the fused parameters. In this way, when the child node devices train a model, some parameters are frozen and do not participate in the training. Therefore, during model training, the progress of the model training is not affected by the central node device with low computing power, and model training efficiency is high.

[0203] Furthermore, the model training periods of multiple child node devices are first predicted separately based on the capability information, and then the first model parameters of the child node devices are determined based on the model training periods, so that the determined first model parameters to be trained can match the corresponding child node devices, and the progress of the model training is guaranteed.

[0204] Furthermore, based on different comparison results between the model training period and the predetermined period conditions, two policies, namely, a policy of selecting part of the model's parameters as the first model parameters and a policy of selecting all of the model's parameters as the first model parameters, are used accordingly, so that the progress of model training is guaranteed and the model performance is guaranteed at the same time.

[0205] Furthermore, since a portion of the model parameters is randomly selected as the first model parameters, the determined first model parameters to be trained can match with the corresponding child node devices, and the progress of model training is guaranteed.

[0206] Furthermore, since the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, the requirements of the model application scenario for the network layers are taken into consideration and the model training performance is guaranteed.

[0207] Furthermore, the central node device determines the first network layer required to build the model based on the model application scenario, so that the requirements of the model application scenario for the network layer are taken into consideration and model training performance is guaranteed.

[0208] Furthermore, since the child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, it is guaranteed that all child node devices participating in the training can train the model, and progress in model training is guaranteed.

[0209] Furthermore, the second model parameters are designed to be unable to affect the default fusion algorithm of the fused parameters and unable to participate in the calculation of the fused parameters, so that the second model parameters that are not updated cannot affect the model training, and the model training performance is guaranteed.The second model parameters are further designed to be able to affect the default fusion algorithm of the fused parameters and able to participate in the calculation of the fused parameters, so that the determined fused parameters can participate in the model training, and the progress of the model training is guaranteed.

[0210] 14 is a diagram of the structure of a communication system according to an embodiment of the present application. This communication system is a communication system corresponding to a model training scenario. As shown in FIG. 14, the communication system may include a central node device 140 and a child node device 141. The central node device 140 may have the function of the communication device 120, and the child node device 141 may have the function of the communication device 130.

[0211] Specifically, the central node device 140 performs the following steps: receiving capability information from a plurality of child node devices (child node devices participating in federated learning of the model) representing a capability of training a model by the child node devices; transmitting to each of the plurality of child node devices the model and first instruction information instructing first model parameters (determined based on the capability information of the child node devices) to be updated when the child node device trains the model; receiving updated first model parameters from the plurality of child node devices; fusing global model parameters of the plurality of child node devices (including the updated first model parameters and second model parameters other than the first model parameters in the model) using a predetermined fusion algorithm to obtain fused parameters; and transmitting the fused parameters to the plurality of child node devices; Execute.

[0212] In this embodiment of the present application, first, multiple child node devices send capability information to the central node device, and then the central node device sends first instruction information and a model to each of the multiple child node devices based on the capability information. Then, the multiple child node devices train a model based on the training set and the model parameters to be updated to obtain updated first model parameters. In addition, the central node device uses a predetermined fusion algorithm to fuse global model parameters of the multiple child node devices, including the first model parameters, to obtain fused parameters, and then the multiple child node devices train a model based on the fused parameters. In this way, when the child node devices train a model, some parameters are frozen and do not participate in the training. Therefore, during model training, the progress of the model training is not affected by the central node device with low computing power, and model training efficiency is high.

[0213] In one possible implementation, the central node device 140 performs the following steps: Separately predicting model training periods for a plurality of child node devices based on the capability information; and determining first model parameters for the child node devices based on the model training periods; Further implementation.

[0214] In this embodiment of the present application, the model training periods of multiple child node devices are first predicted separately based on capability information, and then the first model parameters of the child node devices are determined based on the model training periods, so that the determined first model parameters to be trained can match the corresponding child node devices and ensure the progress of model training.

[0215] In one possible implementation, the step of determining first model parameters of the child node device based on the model training period includes: selecting some of the parameters of the model as first model parameters if the model training period does not satisfy a predetermined period condition, or selecting all of the parameters of the model as first model parameters if the model training period satisfies the predetermined period condition; Includes.

[0216] In this embodiment of the present application, based on different comparison results between the model training period and the predetermined period conditions, two policies are used accordingly, namely, a policy of selecting part of the parameters of the model as the first model parameters and a policy of selecting all the parameters of the model as the first model parameters, so that the progress of model training is guaranteed and at the same time the model performance is guaranteed.

[0217] In one possible implementation, the step of selecting a portion of the parameters of the model as the first model parameters comprises: Randomly selecting a portion of the parameters of the model as first model parameters. Includes.

[0218] In this embodiment of the present application, a portion of the model parameters is randomly selected as the first model parameters, so that the determined first model parameters to be trained can match with the corresponding child node devices, and the progress of model training is guaranteed.

[0219] In one possible implementation, the model includes a first network layer (the model application scenario indicates the first network layer) and a second network layer, the first network layer and the second network layer are different, and the step of selecting some of the parameters of the model as the first model parameters includes: Selecting the parameters of the first network layer and a portion of the parameters of the second network layer as first model parameters. Includes.

[0220] In this embodiment of the present application, the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, so that the requirements of the model application scenario for the network layer are taken into consideration and the model training performance is guaranteed.

[0221] In one possible implementation, before selecting the predetermined first network layer parameters in the first network layer as the first model parameters, the central node device 140 performs the following steps: determining the first network layer required to build the model based on the model application scenario; Further implementation.

[0222] In this embodiment of the present application, the central node device determines the first network layer required to build a model based on a model application scenario, so that the requirements of the model application scenario for the network layer are taken into consideration and model training performance is guaranteed.

[0223] In one possible implementation, the capability information further indicates the type of model that can be trained by the child node device, and the central node device 140 performs the following steps: determining a child node device whose capability information includes a model type of the model as a target device to which the first instruction information and the model are to be transmitted; Further implementation.

[0224] In this embodiment of the present application, a child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, thereby ensuring that all child node devices participating in the training can train the model and ensuring the progress of model training.

[0225] In one possible implementation, the step of fusing global model parameters of multiple child node devices using a predefined fusion algorithm to obtain fused parameters includes: For model parameters having the same function of different child node devices, setting a weight of a second model parameter to a first default weight value, setting a weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter; or, for model parameters having the same function of different child node devices, setting both a weight of the second model parameter and a weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter; The parameters of the first network layer and some of the parameters of the second network layer are selected as first model parameters.

[0226] In this embodiment of the present application, the parameters of the first network layer and part of the parameters of the second network layer are specifically selected as the first model parameters, so that the requirements of the model application scenario for the network layer are taken into consideration and the model training performance is guaranteed.

[0227] In one possible implementation, before selecting the predetermined first network layer parameters in the first network layer as the first model parameters, the central node device 140 performs the following steps: Further performing the step of determining a first network layer required for constructing the model based on the model application scenario.

[0228] In this embodiment of the present application, the central node device determines the first network layer required to build a model based on a model application scenario, so that the requirements of the model application scenario for the network layer are taken into consideration and model training performance is guaranteed.

[0229] In one possible implementation, the capability information further indicates the type of model that can be trained by the child node device, and the central node device 140 performs the following steps: determining a child node device whose capability information includes a model type of the model as a target device to which the first instruction information and the model are to be transmitted; Further implementation.

[0230] In this embodiment of the present application, a child node device whose capability information includes the model type of the model is determined as the target device to which the first instruction information and the model are sent, thereby ensuring that all child node devices participating in the training can train the model and ensuring the progress of model training.

[0231] In one possible implementation, the step of fusing global model parameters of multiple child node devices using a predefined fusion algorithm to obtain fused parameters includes: For model parameters having the same function of different child node devices, setting a weight of a second model parameter to a first default weight value, setting a weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter; or, for model parameters having the same function of different child node devices, setting both a weight of the second model parameter and a weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain a fused parameter; Includes.

[0232] In this embodiment of the present application, the second model parameters are designed so that they cannot affect the default fusion algorithm of the fused parameters and cannot participate in the calculation of the fused parameters, so that the second model parameters that are not updated cannot affect the model training, and the model training performance is guaranteed.The second model parameters are further designed so that they can affect the default fusion algorithm of the fused parameters and can participate in the calculation of the fused parameters, so that the determined fused parameters can participate in the model training, and the progress of the model training is guaranteed.

[0233] The child node device 141 performs the following steps: transmitting capability information representing a capability of training the model by the child node device to the central node device; receiving from the central node device a model and first instruction information instructing first model parameters to be updated when the child node device trains the model, wherein the first model parameters of the child node device to be updated are determined based on the capability information of the child node device; training the model based on the model parameters to be updated and a training set including beam information collected by the child node device to obtain updated first model parameters; transmitting the updated first model parameters to the central node device; receiving from the central node device fused parameters calculated based on the updated first model parameters; training the model based on the fused parameters; receiving beam information; inputting the beam information into the model; and outputting target beam information. Execute.

[0234] In this embodiment of the present application, a model is used to select the target beam, which provides a high efficiency in selecting the target beam.

[0235] An embodiment of the present application further provides a computer-readable storage medium. All or part of the procedures of the aforementioned method embodiments may be completed by a computer program instructing associated hardware. The program may be stored in a computer-readable storage medium. When the program is executed, the procedures of the aforementioned method embodiments may be implemented. The computer-readable storage medium may be, for example, an internal storage unit of the terminal device in any one of the aforementioned embodiments, including the data transmitting side and / or the data receiving side. For example, the computer-readable storage medium may be a hard disk or memory of the terminal device. Alternatively, the computer-readable storage medium may be an external storage device of the terminal device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, or a flash card configured in the terminal device. Furthermore, the computer-readable storage medium may include both the internal storage unit and the external storage device of the terminal device. The computer-readable storage medium is configured to store computer programs and other programs and data required for the terminal device. The computer-readable storage medium may further be configured to temporarily store output data or data to be output.

[0236] An embodiment of the present application further provides computer instructions. All or part of the procedures of the above-mentioned method embodiments may be completed by computer instructions that instruct related hardware (e.g., computers, processors, network devices, and terminals). The program may be stored in the above-mentioned computer-readable storage medium.

[0237] An embodiment of the present application further provides a chip system. The chip system may include a chip, or may include a chip and another discrete component. This is not limited to this. The chip system includes a processor and a transceiver. All or part of the procedures of the above-mentioned method embodiments may be completed by the chip system. For example, the chip system may be configured to perform the functions performed by the central node device in the above-mentioned method embodiments, or to perform the functions performed by the child node device in the above-mentioned method embodiments.

[0238] In one possible design, the chip system further includes a memory configured to store program instructions and / or data. When the chip system is operational, the processor executes the program instructions stored in the memory to enable the chip system to perform the functions performed by the central node device in the aforementioned method embodiments or the functions performed by the child node device in the aforementioned method embodiments.

[0239] In the embodiments of the present application, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or perform the methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor, any conventional processor, etc. The steps of the methods disclosed with reference to the embodiments of the present application may be performed directly by a hardware processor, or may be performed using a combination of hardware and software modules in the processor.

[0240] In embodiments of the present application, the memory may be a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or a volatile memory, such as a random-access memory (RAM). The memory may be, but is not limited to, any other medium capable of carrying or storing expected program code in the form of instructions or data structures and accessible by a computer. Alternatively, the memory in embodiments of the present application may be a circuit or any other device capable of performing a storage function and configured to store instructions and / or data.

[0241] It should be noted that in the specification, claims, and accompanying drawings of this application, terms such as "first," "second," etc. are intended to distinguish between different objects and do not indicate a particular order. Additionally, the terms "comprise," "have," and any other variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally further include unlisted steps or units, or may optionally further include other specific steps or units of the process, method, product, or device.

[0242] In the embodiments of the present application, it should be understood that "at least one" means one or more, "multiple" means two or more, "at least two" means two, three, or more, and "and / or" is used to describe a relationship between related objects and indicates that there can be three relationships. For example, "A and / or B" can indicate the following three cases: when only A is present, when only B is present, and when both A and B are present, where A and B may be singular or plural. The character " / " generally indicates an "or" relationship between related objects. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of a singular item or multiple items. For example, at least one of a, b, or c can refer to a, b, c, a and b, a and c, b and c, or a, b, and c, where a, b, and c may be singular or plural. It should be understood that in the embodiments of the present application, "B corresponding to A" indicates that B is associated with A. For example, B may be determined based on A. It should be further understood that determining B based on A does not mean that B is determined based only on A. B may instead be determined based on A and / or other information. In addition, "connection" in the embodiments of the present application refers to various connection methods, such as direct connection or indirect connection, for implementing communication between devices. This is not limited in the embodiments of the present application.

[0243] Unless otherwise specified, "transmit" in the embodiments of the present application means bidirectional transmission, including sending and / or receiving. Specifically, "transmission" in the embodiments of the present application includes data transmission, data reception, or data transmission and reception. In other words, data transmission in this document includes uplink data transmission and / or downlink data transmission. Data may include channels and / or signals. Uplink data transmission is uplink channel transmission and / or uplink signal transmission, and downlink data transmission is downlink channel transmission and / or downlink signal transmission. In the embodiments of the present application, "network" and "system" represent the same concept, and a communication system is a communication network.

[0244] The above description of the implementation is for the purpose of simple explanation, and those skilled in the art can clearly understand that the division into functional modules is only used as an example for explanation. In actual application, the above functions may be allocated to different functional modules for implementation as needed. In other words, the internal structure of the device is divided into different functional modules to implement all or part of the above functions.

[0245] It should be understood that in some embodiments provided in the present application, the disclosed devices and methods may be implemented in other ways. For example, the described device embodiments are merely examples. For example, the division into modules or units is merely a logical division of function, and in actual implementation, other divisions may be used. For example, multiple units or components may be combined or integrated into another device, or some features may be omitted or not implemented. In addition, the shown or discussed mutual couplings or direct couplings or communication connections may be implemented via some interfaces. Indirect couplings or communication connections between devices or units may be implemented in electrical, mechanical, or other forms.

[0246] The units described as separate parts may or may not be physically separate, and the parts shown as units may be one or more physical units, in other words, located in one place or distributed in different places. Some or all of the units may be selected based on actual requirements to achieve the objectives of the solutions of the embodiments.

[0247] In addition, the functional units of the embodiments of the present application may be integrated into one processing unit, each unit may exist physically alone, or two or more units may be integrated into one unit. The integrated unit may be implemented in the form of hardware or in the form of a software functional unit. When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, the integrated unit may be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application, or portions that contribute to the prior art, or all or part of the technical solutions, may be implemented in the form of a software product. The software product is stored in a storage medium and includes several instructions that instruct a device, such as a single-chip microcomputer or chip, or a processor, to execute all or part of the steps of the method of the embodiments of the present application. The aforementioned storage medium includes any medium capable of storing program code, such as a USB flash drive, a removable hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

[0248] The above description is merely a specific implementation of the present application and is not intended to limit the scope of protection of the present application. Any variation or replacement within the technical scope disclosed in the present application shall fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the scope of protection of the claims. [Explanation of symbols]

[0249] 301 Central Node 302a, 302b, 302c Distributed nodes 1100 Communication equipment 1101 processor 1102 Communication lines 1103 Transceiver 1107 processor 1104 Memory 1105 Output Device 1106 Input Devices 120 Communication equipment 121 Receiver Module 122 Transmitting Module 123 Processing Module 130 Communication equipment 131 Transmitting Module 132 Receiver Module 133 Processing Module 140 Central Node Devices 141 child node devices

Claims

1. 1. A model training method, the method being applied on a central node device, receiving capability information from a plurality of child node devices, the capability information of the child node devices representing capabilities of training a model by the child node devices, and the plurality of child node devices being node devices participating in federated learning of the model; transmitting first instruction information and the model to each of the plurality of child node devices, the first instruction information indicating first model parameters to be updated when the child node device trains the model, the first model parameters to be updated of the child node devices being determined based on the capability information of the child node devices; receiving updated first model parameters from the plurality of child node devices; fusing global model parameters of the plurality of child node devices using a predefined fusion algorithm to obtain fused parameters, wherein the global model parameters include the updated first model parameters and second model parameters of the model other than the first model parameters; transmitting the fused parameters to the plurality of child node devices; A method comprising:

2. separately predicting model training periods for the plurality of child node devices based on the capability information; determining the first model parameters of the child node device based on the model training period; The method of claim 1 further comprising:

3. determining the first model parameters of the child node device based on the model training period; selecting a portion of the model parameters as the first model parameters if the model training period does not satisfy a predetermined period condition; or selecting all parameters of the model as the first model parameters if the model training period satisfies a predetermined period condition; 3. The method of claim 2, comprising:

4. The step of selecting a portion of the model parameters as the first model parameters comprises: randomly selecting the portion of the parameters of the model as the first model parameters; 4. The method of claim 3, comprising:

5. The model includes a first network layer and a second network layer, a model application scenario indicates the first network layer, and the first network layer and the second network layer are different; The step of selecting a portion of the model parameters as the first model parameters comprises: selecting the parameters of the first network layer and a portion of the parameters of the second network layer as the first model parameters; 4. The method of claim 3, comprising:

6. Before a predetermined first network layer parameter in the first network layer is selected as the first model parameter, the method includes: determining the first network layer required to build the model based on the model application scenario; 6. The method of claim 5, further comprising:

7. The capability information further indicates a type of model that can be trained by the child node device, and the method further comprises: determining a child node device whose capability information includes a model type of the model as a target device to which the first instruction information and the model are to be transmitted; 7. The method of claim 1, further comprising:

8. The step of fusing global model parameters of the plurality of child node devices using a predefined fusion algorithm to obtain fused parameters includes: For model parameters having the same function of different child node devices, setting the weight of the second model parameter to a first predetermined weight value and setting the weight of the first model parameter to a second predetermined weight value, and performing weighted averaging to obtain the fused parameters; or For model parameters having the same function of different child node devices, setting both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and performing weighted averaging to obtain the fused parameter.

8. The method of any one of claims 1 to 7, comprising:

9. A beam information selection method, the method being applied to a child node device, receiving beam information; inputting said beam information into a model and outputting target beam information, said model being obtained through training according to the method of any one of claims 1 to 8; A method comprising:

10. 1. A method for training a model, the method being applied to a child node device, sending capability information to a central node device, the capability information of the child node device representing a capability of training a model by the child node device; receiving first instruction information and the model from the central node device, the first instruction information indicating first model parameters to be updated when the child node device trains the model, and the first model parameters to be updated of the child node device are determined based on the capability information of the child node device; training the model based on a training set and the model parameters to be updated to obtain updated first model parameters, wherein the training set includes beam information collected by the child node device; sending the updated first model parameters to the central node device; receiving fused parameters from the central node device, the fused parameters being calculated based on the updated first model parameters; training the model based on the fused parameters; A method comprising:

11. A model training apparatus, the apparatus being used in a central node device; a receiving module configured to receive capability information from a plurality of child node devices, the capability information of the child node devices representing capabilities of training a model by the child node devices, and the plurality of child node devices being node devices participating in federated learning of the model; a transmitting module configured to transmit first instruction information and the model to each of the plurality of child node devices, the first instruction information instructing first model parameters to be updated when the child node device trains the model, the first model parameters to be updated of the child node device being determined based on the capability information of the child node device, and the receiving module further configured to receive the updated first model parameters from the plurality of child node devices; a processing module configured to fuse global model parameters of the plurality of child node devices using a predefined fusion algorithm to obtain fused parameters, the global model parameters including the updated first model parameters and second model parameters of the model other than the first model parameters, the processing module further configured to transmit the fused parameters to the plurality of child node devices; An apparatus comprising:

12. The processing module includes: separately predicting model training periods for the plurality of child node devices based on the capability information; determining the first model parameters of the child node device based on the model training period; The apparatus of claim 11 further configured to:

13. The processing module includes: selecting a portion of the model parameters as the first model parameters if the model training period does not satisfy a predetermined period condition; or If the model training period satisfies a predetermined period condition, all parameters of the model are selected as the first model parameters. The apparatus of claim 12 further configured to:

14. The processing module includes: Randomly selecting the portion of the model parameters as the first model parameters. The apparatus of claim 13 further configured to:

15. The model includes a first network layer and a second network layer, a model application scenario indicates the first network layer, and the first network layer and the second network layer are different; The processing module includes: selecting the parameters of the first network layer and a portion of the parameters of the second network layer as the first model parameters; The apparatus of claim 13 further configured to:

16. The processing module includes: Before selecting a predetermined first network layer parameter in the first network layer as the first model parameter, the first network layer necessary for constructing the model is determined based on the model application scenario.

16. The apparatus of claim 15, further configured to:

17. The capability information further indicates a type of model that can be trained by the child node device, and the processing module: A child node device whose capability information includes a model type of the model is determined as a target device to which the first instruction information and the model are to be transmitted.

17. The apparatus of claim 11, further configured to:

18. The processing module includes: For model parameters having the same function of different child node devices, setting the weight of the second model parameter to a first predetermined weight value, setting the weight of the first model parameter to a second predetermined weight value, and performing weighted averaging to obtain the fused parameters; or For model parameters having the same function of different child node devices, set both the weight of the second model parameter and the weight of the first model parameter to a second default weight value, and perform weighted averaging to obtain the fused parameters.

18. The apparatus of claim 11, further configured to:

19. A communication apparatus, the apparatus being used in a child node device; a receiving module configured to receive beam information; a processing module configured to input said beam information into a model and output target beam information, said model being obtained through training by an apparatus according to any one of claims 11 to 18; An apparatus comprising:

20. A communication apparatus, the apparatus being used in a child node device; a sending module configured to send capability information to a central node device, the capability information of the child node device representing a capability of training a model by the child node device; a receiving module configured to receive first instruction information and the model from the central node device, the first instruction information indicating first model parameters to be updated when the child node device trains the model, the first model parameters to be updated of the child node device being determined based on the capability information of the child node device; a processing module configured to train the model based on a training set and the model parameters to be updated to obtain updated first model parameters, the training set including beam information collected by the child node device; and Equipped with the transmitting module is configured to transmit the updated first model parameters to the central node device; the receiving module is configured to receive fused parameters from the central node device, the fused parameters being calculated based on the updated first model parameters; the processing module is configured to train the model based on the fused parameters. Device.

21. 11. A communications device, the communications device comprising a processor and a transceiver, the processor and the transceiver configured to support the communications device in performing the method of any one of claims 1 to 10.

22. 11. A computer-readable storage medium having stored thereon computer instructions that, when executed, perform the method of any one of claims 1 to 10.

Images