Communication method and related device

Abstract

The invention provides a communication method and a related device. In the method, a first access network device sends N pieces of first parameter information of a first model to N terminal devices, N being an integer greater than or equal to 1. And the first access network equipment receives N pieces of second parameter information from the N pieces of terminal equipment, and each piece of second parameter information is obtained after one piece of first parameter information is updated. And then, the first access network equipment updates the first model based on the N pieces of second parameter information, and sends third parameter information of the updated first model to the second access network equipment. And after receiving the third parameter information from the first access network equipment, the second access network equipment updates the first model based on the third parameter information. Therefore, the range and the number of terminal devices in one training iteration are reduced, the heterogeneity of the training data of the model is reduced, and the precision of model training is improved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a communication method and related apparatus. Background Technology

[0002] Federated learning is a distributed learning paradigm that enables the joint training of distributed user-local models through unified coordination and configuration by a central server, while ensuring that data does not leave the local machine.

[0003] In the federated learning process of wireless networks, a central server distributes model parameter information to a large number of terminal devices through multiple base stations. The terminal devices train the model based on their local data and then upload the trained parameter information to the central server. After collecting the parameter information uploaded by the terminal devices, the central server aggregates the parameter information from each terminal device to obtain an updated model.

[0004] In this case, the coverage area of ​​multiple base stations is very large, and the number of terminal devices participating in federated learning is also very large. However, the training data of the model comes from the data stored locally on each terminal device. Therefore, the training data of the model is non-independent and identically distributed (non-iid), which leads to low training accuracy of the model. Summary of the Invention

[0005] This application provides a communication method and related apparatus for improving the accuracy of model training.

[0006] Firstly, this application provides a communication method. This method can be applied to the access network side, such as an access network device or a communication / processing module within the access network device, or circuits or chips responsible for communication functions within the access network device (such as a modem chip, also known as a baseband chip, or a system-on-a-chip (SoC) chip containing a modem core, or a system-in-package (SIP) chip), or circuits or chips responsible for processing functions within the access network device (such as a graphics processing unit (GPU)). Taking the application of this method to a first access network device as an example, in this method, the first access network device sends N first parameter information of a first model to N terminal devices, where N is an integer greater than or equal to 1. Each of the N terminal devices receives one piece of first parameter information; in other words, the first access network device sends one piece of first parameter information to each of the N terminal devices.

[0007] After receiving the first parameter information from the first access network device, each of the N terminal devices combines it with its local data to train a model and obtain second parameter information. Each second parameter information is an updated version of the first parameter information. Next, each terminal device sends its trained second parameter information to the first access network device. Correspondingly, the first access network device receives N second parameter information messages from the N terminal devices, where each second parameter information message originates from a single terminal device.

[0008] Then, the first access network device updates the first model based on N second parameter information and sends the updated third parameter information of the first model to the second access network device. Upon receiving the third parameter information from the first access network device, the second access network device updates the first model based on the third parameter information. Next, the second access network device acts as a server, and the terminal devices in its serving cell act as computing nodes to continue training the first model in the second access network device. The specific process is similar to the aforementioned training process of the first model by the first access network device, and will not be repeated here.

[0009] In this application, only one access network device participates in each training iteration of the model. Furthermore, after one access network device completes the training of the model, the parameter information of the model can be transferred to another access network device for further training. Thus, while keeping the overall amount of training data constant, the scope and number of terminal devices in a single training iteration are reduced, the heterogeneity of the model's training data is decreased, and the accuracy of model training is improved.

[0010] On the other hand, the model's parameter information can be transferred between different access network devices, and the training results of the previous access network device can be accumulated in the next access network device, enabling terminal devices in different regions to participate in the joint training of the model, expanding the richness of the model's training data and improving the model's generalization.

[0011] Based on the first aspect, in one optional implementation, the first access network device sends the complete model parameter information of the first model to each of the N terminal devices. That is, each of the N first parameter information is the complete model parameter information of the first model; therefore, it can be considered that the N first parameter information are identical. Thus, each terminal device possesses the complete parameters of the model, allowing for more efficient use of the computing power of each terminal device and reducing the waste of computing resources caused by model segmentation. Furthermore, having the complete model parameters for each terminal device simplifies the integration and updating of the first model by the first access network device, reducing its computational overhead.

[0012] Based on the first aspect, in one optional implementation, the first model can be divided into M sub-models, where M is an integer greater than or equal to 1. Then, the N first parameter information sent by the first access network device to N terminal devices includes parameter information for the M sub-models. Each of the N first parameter information represents the parameter information of one of the M sub-models. Therefore, the aforementioned "N first parameter information sent by the first access network device to the N terminal devices" can be understood as the first access network device sending parameter information of one sub-model of the first model to each of the N terminal devices. Thus, by splitting the complete first model into multiple sub-models and sending the parameter information of these sub-models to each terminal device for training, the communication overhead required for transmitting parameter information can be reduced, the computing power overhead of the terminal devices can be reduced, and the computational efficiency can be improved.

[0013] Based on the first aspect, in one optional implementation, the first model includes M sub-models, wherein the parameter scales and / or input data modalities differ among the M sub-models. Correspondingly, in the N first parameter information sent by the first access network device to N terminal devices, each first parameter information corresponds to the parameter information of one of the M sub-models. Thus, each terminal device only needs to train one sub-model, instead of training the complete first model, reducing the computing power overhead and communication overhead for transmitting parameter information for the terminal devices.

[0014] Based on the first aspect, in an optional implementation, after a sub-model in the first access network device reaches the training convergence condition, the first access network device can send the parameter information of the already converged sub-model to the second access network device without waiting for other sub-models to complete training convergence. That is, the third parameter information in this application is the parameter information of one of the M sub-models. Optionally, this sub-model can be a sub-model in the first access network device that has completed training convergence. After receiving the third parameter information, the second access network device can perform auxiliary updates to the first model (including sub-models) in the second access network device based on the third parameter information. This allows the second access network device to participate in the training of the first model more quickly, improving the efficiency of model training.

[0015] On the other hand, after the other sub-models in the first access network device have successively completed training and converged, the first access network device continues to send the parameter information of these sub-models to the second access network device. Optionally, the second access network device performs auxiliary updates on the other sub-models received subsequently based on the updated sub-models, thereby reducing the process of model distribution and training of other sub-models on the second access network device, improving the efficiency of model training, and reducing the communication overhead required to transmit the parameter information of other sub-models.

[0016] Based on the first aspect, in one optional implementation, the first access network device can wait until all M sub-models have completed training and convergence before sending the parameter information of the M converged sub-models together to the second access network device. That is, the third parameter information includes the parameter information of each of the M sub-models. Therefore, the second access network device can obtain richer and more complete parameter information, enabling it to update the first model in the second access network device more accurately.

[0017] Based on the first aspect, in one optional implementation, the first access network device determines the next-hop node and transmits the parameter information of the first model to the next-hop node. This implementation is relatively simple, robust, and highly versatile for various federated frameworks and model structures. Furthermore, the algorithm has strong real-time performance and can effectively handle unexpected situations.

[0018] Based on the first aspect, in one optional implementation, the core network device pre-plans the transmission path of the parameter information (e.g., the third parameter information) of the first model. Then, the core network device sends first information to the first access network device, indicating that the second access network device is the next-hop node for the first access network device. Thus, when the training of the first model in the first access network device converges, the first access network device can more quickly transmit the third parameter information of the first model to the next-hop node (the second access network device) without needing to plan the route in real time, saving computational overhead and accelerating the training process.

[0019] Based on the first aspect, in one optional implementation, the first access network device is the next-hop node of the third access network device. The third access network device first trains the first model to obtain the fourth parameter information of the first model, which is the updated parameter information of the first model in the third access network device. Next, the third access network device sends the fourth parameter information of the first model to the first access network device. Correspondingly, the first access network device receives the fourth parameter information of the first model from the third access network device. The first access network device updates the first model in the first access network device based on the fourth parameter information. Then, the first access network device sends the N first parameter information of the first model to N terminal devices.

[0020] Based on the first aspect, in one optional implementation, the first access network device is the next-hop node of the third access network device, and the first model includes a first sub-model and a second sub-model. The first sub-model in the first model of the third access network device completes training convergence first. Then, the third access network device first sends the parameter information of the first sub-model to the first access network device, and the parameter information of the first sub-model is the fourth parameter information. Correspondingly, the first access network device first receives the fourth parameter information of the first sub-model.

[0021] Next, the second sub-model in the first model of the third access network device also completes training convergence. Then, the third access network device sends the parameter information of the second sub-model to the first access network device. This parameter information of the second sub-model is the fifth parameter information.

[0022] Then, the first access network device performs auxiliary updates (e.g., knowledge distillation) on the second sub-model based on the fifth parameter information and the updated first sub-model. This reduces the process of model distribution and training of the second sub-model on the first access network device, improves the efficiency of model training, and reduces the communication overhead required to transmit the parameter information of the second sub-model.

[0023] Secondly, this application provides a communication device, including a transceiver unit and a processing unit.

[0024] The transceiver unit is used to send N first parameter information of the first model to N terminal devices. Each terminal device is used to receive one first parameter information, where N is an integer greater than or equal to 1.

[0025] The transceiver unit is also used to receive N second parameter information from N terminal devices, where each second parameter information is obtained by updating a first parameter information;

[0026] The processing unit is used to update the first model based on N second parameter information;

[0027] The transceiver unit is also used to send the updated third parameter information of the first model to the second access network device, so that the second access network device can update the first model based on the third parameter information.

[0028] Based on the second aspect, in one optional implementation, each first parameter information is the parameter information of the complete model of the first model.

[0029] Based on the second aspect, in an optional implementation, the first model includes M sub-models, and each first parameter information is the parameter information of one of the M sub-models, where M is an integer greater than or equal to 1.

[0030] Based on the second aspect, in one optional implementation, the parameter scales and / or the modalities of the input data differ among the M sub-models.

[0031] Based on the second aspect, in one optional implementation, the third parameter information is the parameter information of one of the M sub-models.

[0032] Based on the second aspect, in one optional implementation, the third parameter information includes parameter information for each of the M sub-models.

[0033] Based on the second aspect, in one optional implementation, the processing unit is further configured to determine the second access network device as the next-hop node.

[0034] Based on the second aspect, in an optional implementation, the transceiver unit is further configured to receive first information from the core network device, the first information being used to indicate that the second access network device is the next-hop node of the first access network device.

[0035] Based on the second aspect, in an optional implementation, the transceiver unit is further configured to receive fourth parameter information from the first model of the third access network device.

[0036] The processing unit is also used to update the first model based on the fourth parameter information.

[0037] Based on the second aspect, in an optional implementation, the first model includes a first sub-model and a second sub-model, the fourth parameter information is the parameter information of the first sub-model, and the transceiver unit is further configured to receive fifth parameter information from the third access network device, the fifth parameter information being the parameter information of the second sub-model.

[0038] The processing unit is also used to update the second sub-model based on the first sub-model and the fifth parameter information.

[0039] Thirdly, this application provides a communication device including an interface circuit and one or more processors. The one or more processors are coupled to a memory. The memory stores part or all of the necessary computer program or instructions for implementing the functions described in the first aspect. The one or more processors can execute the computer program or instructions, causing the communication device to implement the methods in any possible design or implementation of the first aspect. The interface circuit is used to implement the communication functions within the communication device and / or the communication functions between the communication device and other devices or components.

[0040] In one possible design, the processor is used to communicate with other devices or components through the interface circuit.

[0041] In one possible design, the communication device may also include the memory.

[0042] The aforementioned communication device may be a first access network device, or a communication / processing module in the first access network device, or a chip in the first access network device responsible for communication functions such as a modem chip (also known as a baseband chip) or a SoC or SIP chip containing a modem module, or a circuit or chip (such as a GPU) in the first access network device responsible for processing functions.

[0043] Fourthly, this application provides a communication system, including the first access network device, the second access network device, and the terminal device mentioned in the first aspect.

[0044] Fifthly, this application provides a computer-readable storage medium storing computer-readable instructions that, when read and executed by a computer, cause the computer to perform any of the possible designs in the first aspect described above.

[0045] In a sixth aspect, this application provides a computer program product that, when read and executed by a computer, causes the computer to perform any of the possible design methods of the first aspect described above.

[0046] The technical effects of any of the design methods in aspects two through six can be found in the technical effects of the different design methods in aspect one above, and will not be repeated here. Attached Figure Description

[0047] Figures 1a to 1e This is a schematic diagram of the AI ​​processing involved in this application;

[0048] Figure 2 A schematic diagram of a possible system framework for federated learning;

[0049] Figure 3 This is a schematic diagram of one possible, non-limiting system used in the communication method and related apparatus of this application;

[0050] Figure 4 This is a schematic diagram of another possible, non-limiting system used in the communication method and related apparatus of this application;

[0051] Figure 5 This is a schematic diagram illustrating the implementation of model training in this application;

[0052] Figure 6 This is a schematic diagram illustrating the implementation of model inference in this application;

[0053] Figure 7 This is a schematic diagram illustrating one possible implementation of the communication method in this application;

[0054] Figure 8 This is a schematic diagram illustrating another possible implementation of the communication method in this application;

[0055] Figure 9 This is a schematic diagram illustrating another possible implementation of the communication method in this application;

[0056] Figure 10 This is a schematic diagram illustrating one implementation of the M sub-models in the first model of this application;

[0057] Figure 11 This is a schematic diagram illustrating another implementation of the M sub-models in the first model of this application;

[0058] Figure 12 This is a schematic diagram illustrating another implementation of the M sub-models in the first model of this application;

[0059] Figure 13 This is a schematic diagram illustrating one implementation of knowledge distillation in this application;

[0060] Figure 14 This is a schematic diagram illustrating another implementation of the communication method in this application;

[0061] Figure 15 This is a schematic diagram illustrating the implementation of M sub-models of different modalities in this application;

[0062] Figures 16 to 18 This is a schematic diagram of the structure of the communication device provided in the embodiments of this application. Detailed Implementation

[0063] The present application will now be described with reference to the accompanying drawings. The terminology used in the embodiments section is for illustrative purposes only and is not intended to limit the scope of the application. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in this application are equally applicable to similar technical problems.

[0064] First, some of the nouns or terms used in this application will be explained, and these nouns or terms are also part of the content of the invention.

[0065] (1) The terms “system” and “network” in this application are used interchangeably. “Multiple” refers to two or more. “And / or” describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, or B exists alone, where A and B can be singular or plural. The character “ / ” generally indicates that the related objects before and after are in an “or” relationship. “At least one of the following” or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, “at least one of A, B and C” includes A, B, C, AB, AC, BC or ABC. Unless otherwise specified, the ordinal numbers such as “first” and “second” mentioned in this application are used to distinguish multiple objects and are not used to limit the order, sequence, priority or importance of multiple objects. Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0066] (2) In this application, “sending information” can be understood as one device sending information to another device, or it can also be understood as one logical module within a device sending information to another logical module. For example, “terminal device sending information” can be understood as a terminal device sending information to another device (such as a network device), or it can be understood as logical module 1 in the terminal device sending information to logical module 2 in the network device.

[0067] In this application, "receiving information" can be understood as one device receiving information from another device, or it can also be understood as a logical module within a device receiving information from another logical module. For example, "terminal device receiving information" can be understood as a terminal device receiving information from another device (such as a network device), or it can be understood as logical module 1 in the terminal device receiving information from logical module 2 in the network device.

[0068] In this application, "sending information to... (e.g., a network device)" or the relevant illustrations in the accompanying drawings can be understood as the destination of the information being a network device. This can include sending information directly or indirectly to a network device. "Receiving information from... (e.g., a network device)" or "receiving information from... (e.g., a network device)" or "receiving information sent (e.g., by a network device)" or the relevant illustrations in the accompanying drawings can be understood as the source of the information being a network device. This can include receiving information directly or indirectly from a network device. Information may undergo necessary processing between the source and destination, such as format changes, encoding, modulation, etc., but the destination can understand the valid information from the source. Similar expressions in this application can be understood in a similar way, and will not be elaborated further here.

[0069] (3) Configuration and Pre-configuration: In this application, both configuration and pre-configuration are used. Configuration refers to the network device or server sending configuration information or parameter values ​​to the terminal device via messages or signaling, so that the terminal device can determine the communication parameters or resources for transmission based on these values ​​or information. Pre-configuration is similar to configuration; it can be parameter information or parameter values ​​pre-negotiated between the network device / server and the terminal device, parameter information or parameter values ​​specified by standard protocols for use by the base station / network device or terminal device, or parameter information or parameter values ​​pre-stored in the base station / server or terminal device. This application does not limit this.

[0070] It should be understood that these values ​​and parameters can change or be updated.

[0071] (4) In this application, “instruction” may include direct instruction and indirect instruction, and may also include explicit instruction and implicit instruction. When describing a certain instruction information for the purpose of instructing A, it can be understood that the instruction information carries A, directly instructs A, or indirectly instructs A.

[0072] In this application, the information indicated by the instruction information is called the information to be instructed. In specific implementations, there are many ways to indicate the information to be instructed, such as, but not limited to, directly indicating the information to be instructed, such as the information to be instructed itself or its index. It can also indirectly indicate the information to be instructed by indicating other information, where there is a relationship between the other information and the information to be instructed; or it can indicate only a part of the information to be instructed, while the other parts are known or pre-agreed upon, for example, by using a pre-agreed (e.g., protocol-predefined) arrangement of various information to indicate specific information, thereby reducing instruction overhead to some extent. This application does not limit the specific method of instruction. It is understood that for the sender of the instruction information, the instruction information can be used to indicate the information to be instructed, and for the receiver of the instruction information, the instruction information can be used to determine the information to be instructed.

[0073] (5) AI can enable machines to possess human-like intelligence, for example, allowing machines to use computer hardware and software to simulate certain intelligent human behaviors. To achieve artificial intelligence, machine learning methods can be used. In machine learning, machines learn (or train) a model using training data. This model represents the mapping between input and output. The learned model can be used for reasoning (or prediction), that is, it can be used to predict the output corresponding to a given input. This output can also be called the reasoning result (or prediction result).

[0074] Machine learning can include supervised learning, unsupervised learning, and reinforcement learning. Unsupervised learning can also be called learning without supervision.

[0075] Supervised learning, based on collected sample values ​​and labels, uses machine learning algorithms to learn the mapping relationship between sample values ​​and labels, and then expresses this learned mapping relationship using an AI model. The process of training the machine learning model is the process of learning this mapping relationship. During training, sample values ​​are input into the model to obtain the model's predicted values, and the model parameters are optimized by calculating the error between the model's predicted values ​​and the sample labels (ideal values). After the mapping relationship is learned, it can be used to predict new sample labels. The mapping relationship learned in supervised learning can include linear or non-linear mappings. Based on the type of label, the learning task can be divided into classification tasks and regression tasks.

[0076] Unsupervised learning relies on collected sample values ​​to discover inherent patterns within the samples themselves. One type of unsupervised learning algorithm uses the samples themselves as supervisory signals, meaning the model learns the mapping relationship from sample to sample; this is called self-supervised learning. During training, model parameters are optimized by calculating the error between the model's predictions and the samples themselves. Self-supervised learning can be used for signal compression and decompression recovery applications; common algorithms include autoencoders and generative adversarial networks.

[0077] Reinforcement learning, unlike supervised learning, is a type of algorithm that learns problem-solving strategies through interaction with the environment. Unlike supervised and unsupervised learning, reinforcement learning problems do not have explicit "correct" action labels. The algorithm needs to interact with the environment to obtain reward signals from the environment, and then adjust its decision actions to obtain a larger reward signal value. For example, in downlink power control, the reinforcement learning model adjusts the downlink transmission power of each user based on the total system throughput feedback from the wireless network, aiming to achieve a higher system throughput. The goal of reinforcement learning is also to learn the mapping relationship between the environment state and a better (e.g., optimal) decision action. However, because the label of the "correct action" cannot be obtained in advance, the network cannot be optimized by calculating the error between the action and the "correct action." Reinforcement learning training is achieved through iterative interaction with the environment.

[0078] Neural networks (NNs) are a specific model in machine learning techniques. According to the general approximation theorem, neural networks can theoretically approximate any continuous function, thus enabling them to learn arbitrary mappings. Traditional communication systems rely on extensive expert knowledge to design communication modules, while deep learning communication systems based on neural networks can automatically discover hidden pattern structures from large datasets, establish mapping relationships between data, and achieve performance superior to traditional modeling methods.

[0079] The idea behind neural networks comes from the neuronal structure of the brain. For example, each neuron performs a weighted summation of its input values ​​and outputs the result through an activation function.

[0080] like Figure 1a The diagram shown is a schematic representation of a neuron structure. Assume the neuron's input is x = [x0, x1, ..., x...]. n The weights corresponding to each input are w = [w0, w1, ..., w] n ], where n is a positive integer, w i and x i It can be any possible type, such as a decimal, an integer (e.g., 0, a positive integer, or a negative integer), or a complex number. i As x i The weights are used to assign weights to x.i Weighting is applied. The bias for the weighted sum of the input values ​​is, for example, b. Activation functions can take many forms. Assuming the activation function of a neuron is y = f(z) = max(0, z), then the output of that neuron is: For example, if the activation function of a neuron is y = f(z) = z, then the output of that neuron is: Here, b can be any possible type, such as a decimal, an integer (e.g., 0, a positive integer, or a negative integer), or a complex number. The activation functions of different neurons in a neural network can be the same or different.

[0081] Furthermore, neural networks generally consist of multiple layers, each of which may include one or more neurons. Increasing the depth and / or width of a neural network can improve its expressive power, providing more powerful information extraction and abstract modeling capabilities for complex systems. The depth of a neural network can refer to the number of layers it includes, and the number of neurons in each layer can be called the width of that layer. In one implementation, a neural network includes an input layer and an output layer. The input layer processes the received input information through neurons and passes the processing result to the output layer, which then obtains the output of the neural network. In another implementation, a neural network includes an input layer, hidden layers, and an output layer. The input layer processes the received input information through neurons and passes the processing result to the hidden layer. The hidden layer calculates the received processing result and passes the calculation result to the output layer or the next adjacent hidden layer, ultimately obtaining the output of the neural network. A neural network may include one hidden layer or multiple sequentially connected hidden layers, without limitation.

[0082] Neural networks, for example, are deep neural networks (DNNs). Depending on how the network is constructed, DNNs can include feedforward neural networks (FNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

[0083] Figure 1b This is a schematic diagram of a Free-Nearest Neural Network (FNN). A key characteristic of FNNs is that neurons in adjacent layers are completely connected pairwise. This characteristic makes FNNs typically require a large amount of storage space, leading to high computational complexity.

[0084] CNNs are neural networks specifically designed to process data with a grid-like structure. For example, time-series data (e.g., discrete sampling along a time axis) and image data (e.g., two-dimensional discrete sampling) can both be considered grid-like data. CNNs do not use all the input information at once for computation; instead, they use a fixed-size window to extract a portion of the information for convolution operations, which significantly reduces the computational cost of model parameters. Furthermore, depending on the type of information extracted by the window (e.g., people and objects in an image represent different types of information), each window can use different convolution kernels, allowing CNNs to better extract features from the input data.

[0085] Recurrent Neural Networks (RNNs) are a type of distributed neural network (DNN) that utilizes feedback time-series information. The input to an RNN includes the current input value and its own output value from the previous time step. RNNs are well-suited for acquiring temporally correlated sequence features, and are particularly applicable to applications such as speech recognition and channel coding / decoding.

[0086] In the model training process described above, a loss function can be defined. The loss function describes the difference between the model's output value and the ideal target value. The loss function can be expressed in various forms, and there are no restrictions on its specific form. The model training process can be viewed as follows: by adjusting some or all of the model's parameters, the value of the loss function is made to be less than a threshold or to meet the target requirement.

[0087] A model can also be called an AI model, a rule, or other names. An AI model can be considered a specific method for implementing AI functions. An AI model represents the mapping relationship or function between the model's input and output. AI functions can include one or more of the following: data collection, model training (or model learning), model information dissemination, model inference (or model reasoning, inference, or prediction, etc.), model monitoring or model validation, or inference result publication, etc. AI functions can also be called AI (related) operations or AI-related functions.

[0088] The implementation process of a fully connected neural network will be described below with reference to the accompanying drawings. A fully connected neural network is also called a multilayer perceptron (MLP).

[0089] like Figure 1c As shown, an MLP consists of an input layer (left side), an output layer (right side), and multiple hidden layers (middle). Each layer of an MLP contains several nodes, called neurons. Neurons in adjacent layers are connected pairwise.

[0090] Optionally, considering neurons in two adjacent layers, the output h of a neuron in the next layer is the weighted sum of all neurons x in the previous layer connected to it, after passing through an activation function, and can be expressed as:

[0091] h = f(wx + b).

[0092] Where w is the weight matrix, b is the bias vector, and f is the activation function.

[0093] Alternatively, the output of the neural network can be recursively expressed as:

[0094] y = f n (w n f n-1 (…)+b n ).

[0095] Where n is the index of the neural network layer, n is greater than or equal to 1 and less than or equal to N, where N is the total number of layers in the neural network.

[0096] In other words, a neural network can be understood as a mapping from an input data set to an output data set. Neural networks are typically initialized randomly; the process of obtaining this mapping from random values ​​w and b using existing data is called training the neural network.

[0097] Optionally, the training method involves using a loss function to evaluate the output of the neural network.

[0098] like Figure 1d As shown, the error can be backpropagated, and the neural network parameters (including w and b) can be iteratively optimized using gradient descent until the loss function reaches its minimum. Figure 1d The term "relative advantage (e.g., optimal advantage)" is used. This is understandable. Figure 1d The neural network parameters corresponding to the "better points (e.g., the best points)" in the data can be used as neural network parameters in the trained AI model information.

[0099] Alternatively, the gradient descent process can be represented as:

[0100]

[0101] Where θ represents the parameters to be optimized (including w and b), L is the loss function, and η is the learning rate, controlling the step size of gradient descent. This represents the differentiation operation. This indicates taking the derivative of θ with respect to L.

[0102] Alternatively, the backpropagation process can utilize the chain rule for partial derivatives.

[0103] like Figure 1eAs shown, the gradient of the parameters in the previous layer can be recursively calculated from the gradient of the parameters in the next layer, and can be expressed as:

[0104]

[0105] Among them, w ij Let s be the weight of the connection between node j and node i. i The weighted sum of the inputs at node i.

[0106] Next, we will introduce the possible, non-limiting scenarios involved in this application.

[0107] With the advent of the big data era, traditional cloud computing models are no longer sufficient to meet requirements such as real-time performance, privacy, and communication bottlenecks. In a system context where communication and computing are intertwined, network architecture is gradually evolving from centralized cloud computing to distributed edge computing, a process driven by several key factors. First, breakthroughs in artificial intelligence, particularly the continuous exploration of deep learning applications across various fields, have demonstrated superior performance in classification, generation, representation, and prediction. At the architecture and algorithm levels, a large number of high-performance neural network structures and lightweight models for mobile devices have emerged to match more real-world scenarios. Second, the continuous development of the communications industry has led to a surge in communication volume. The gradual deployment of 5G networks in recent years has brought faster and more reliable data transmission services, providing fundamental data transmission support for distributed intelligent networks. Third, users' increasing emphasis on data security highlights the necessity of intelligent edge computing. In the big data era, the risk of personal information leakage is exacerbated. Edge computing, by offloading computing tasks to the edge, ensures that data does not leave the local machine, thus mitigating the risk of long-distance data transmission over the network and reducing privacy breaches.

[0108] Federated learning, as a typical method of distributed edge computing, is gaining increasing attention from academia and industry due to its simple training model and protection of data privacy. With the continuous development of communication systems and smart devices, federated learning-based solutions have broad application potential in many fields such as mobile edge computing, network-computing integration, and the Internet of Things. Federated learning is a distributed learning paradigm that enables the joint training of distributed user-local models through unified coordination and configuration by a central server, ensuring that data does not leave the local environment.

[0109] Please see Figure 2 , Figure 2 A schematic diagram of a possible system framework for federated learning. For example... Figure 2 As shown: The central server transmits data to a large number of terminal devices (e.g., [missing information]) through multiple base stations. Figure 2Terminal devices 1, 2, and K (as shown) send parameter information for the model. Each terminal device trains its model using local data and uploads the trained parameter information to the central server. The central server collects the uploaded parameter information from each terminal device, aggregates the parameter information, and obtains an updated model.

[0110] In this case, the coverage area of ​​multiple base stations is very large, and the number of terminal devices participating in federated learning is also very large. However, the training data of the model comes from the data stored locally on each terminal device. Therefore, the training data of the model is non-independent and identically distributed (non-iid), which leads to low training accuracy of the model.

[0111] To address the aforementioned problems, this application provides a communication method and related apparatus for improving the accuracy of model training. The communication method and related apparatus provided in this application can be applied to various communication systems. For example, 5th generation (5G) mobile communication systems, new radio (NR) systems, long term evolution (LTE) systems, LTE frequency division duplex (FDD) systems, LTE time division duplex (TDD) systems, future communication systems, vehicle-to-everything (V2X) communication systems, device-to-device (D2D) communication systems, Internet of Things (IoT) communication systems, industrial internet communication systems, or satellite communication systems, etc. The wireless communication systems involved in this application also include, but are not limited to, narrowband Internet of Things (NB-IoT) systems.

[0112] For example, please refer to Figure 3 , Figure 3 This is a schematic diagram of one possible, non-limiting system used in the communication method and related apparatus of this application. Figure 3 As shown, the communication system 10 includes a radio access network (RAN) 100 and a core network (CN) 200. Optionally, the communication system 10 may also include an Internet 300. The RAN 100 includes at least one RAN node (e.g., Figure 3 110a and 110b (collectively referred to as 110) and at least one terminal device (such as Figure 3RAN 100, denoted as RAN 120a-120j, is collectively referred to as RAN 120. RAN 100 may also include other RAN nodes, such as wireless relay equipment and / or wireless backhaul equipment. Figure 3 (Not shown in the image). Terminal device 120 is connected to RAN node 110 wirelessly. RAN node 110 is connected to core network 200 wirelessly or via wired connection. The core network equipment in core network 200 and RAN node 110 in RAN 100 can be different physical devices, or they can be the same physical device integrating core network logical functions and radio access network logical functions. Terminal devices and RAN nodes can be interconnected via wired or wireless connections.

[0113] RAN 100 can be a cellular system related to the 3rd Generation Partnership Project (3GPP), such as a fourth-generation (4G) mobile communication system, a fifth-generation (5G) mobile communication system, or a future communication system. RAN 100 can also be an open access network (openRAN, O-RAN, or ORAN), a cloud radio access network (CRAN), an evolved universal terrestrial radio access (E-UTRA) system, or a wireless fidelity (WiFi) system. RAN 100 can also be a communication system that integrates two or more of the above systems.

[0114] RAN node 110, sometimes also referred to as network device, access network device, RAN equipment, RAN entity, or access node, constitutes part of the communication system and is used to help terminal devices achieve wireless access. Multiple RAN nodes 110 in the communication system 10 can be of the same type or different types. In some scenarios, the roles of RAN node 110 and terminal device 120 are relative, for example... Figure 3 Network element 120i can be a helicopter or a drone, and it can be configured as a mobile base station. For terminal devices 120j that access RAN 100 through network element 120i, network element 120i is a base station; however, for base station 110a, network element 120i is a terminal device. RAN node 110 and terminal device 120 are sometimes referred to as communication devices, for example... Figure 3Network elements 110a and 110b can be understood as communication devices with base station functions, while network elements 120a-120j can be understood as communication devices with terminal equipment functions.

[0115] In one possible scenario, RAN node 110 can be a base station, an evolved NodeB (eNodeB), an access point (AP), a transmission reception point (TRP), a next-generation NodeB (gNB), a base station in a future mobile communication system, or an access node in a WiFi system. Optionally, RAN node 110 can also be a macro base station (such as...). Figure 3 110a), micro base stations or indoor stations (such as Figure 3 The RAN node 110 can be a relay node or donor node, or a wireless controller in a CRAN scenario. Optionally, the RAN node 110 can also be a server, wearable device, vehicle, or in-vehicle equipment. For example, the access network device in vehicle-to-everything (V2X) technology can be a roadside unit (RSU). All or part of the functions of the RAN node 110 in this application can also be implemented through software functions running on hardware, or through virtualization functions instantiated on a platform (e.g., a cloud platform). The RAN node 110 can also be equipped with communication modules, circuits, or chips that perform corresponding communication functions. The RAN node 110 can also be configured with program instructions for performing corresponding communication functions and corresponding program instructions. The RAN node 110 in this application can also be a logical node, logical module, or software that can implement all or part of the functions of the RAN node 110.

[0116] In another possible scenario, multiple RAN nodes collaborate to assist terminal devices in achieving wireless access, with different RAN nodes each implementing a portion of the base station's functions. For example, RAN nodes can be central units (CUs), distributed units (DUs), CU-control plane (CPs), CU-user plane (UPs), or radio units (RUs), etc. CUs and DUs can be set up separately or included in the same network element, such as a baseband unit (BBU). RUs can be included in radio frequency equipment or radio frequency units, such as remote radio units (RRUs), active antenna units (AAUs), or remote radio heads (RRHs).

[0117] In different systems, CU (or CU-CP and CU-UP), DU, or RU may have different names, but those skilled in the art will understand their meaning. For example, in an ORAN system, CU can also be called O-CU (open CU), DU can also be called O-DU, CU-CP can also be called O-CU-CP, CU-UP can also be called O-CU-UP, and RU can also be called O-RU. For ease of description, this application uses CU, CU-CP, CU-UP, DU, and RU as examples. Any of the units among CU (or CU-CP, CU-UP), DU, and RU in this application can be implemented through software modules, hardware modules, or a combination of software and hardware modules.

[0118] Terminal equipment can be any device or module that connects to the communication system shown above and has corresponding communication functions. Terminal equipment can also be referred to as a terminal, user equipment (UE), mobile station (MS), mobile terminal (MT), fixed wireless access (FWA), or customer premises equipment (CPE), etc. Terminal equipment includes wireless communication functions (providing voice / data connectivity to users). Examples include handheld devices with wireless connectivity, in-vehicle devices, and machine-type communication (MTC) terminals. Currently, terminal devices can include: mobile phones, tablets, laptops, PDAs, mobile internet devices (MIDs), wearable devices, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminals in industrial control, wireless terminals in self-driving (e.g., drones, vehicles), wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, and wireless terminals in smart homes. For example, wireless terminals in self-driving can be drones, helicopters, or airplanes. For example, wireless terminals in vehicle-to-everything (V2X) can be in-vehicle equipment, vehicle-mounted equipment, in-vehicle modules, vehicles, or ships. Wireless terminals in industrial control can be cameras, robots, or robotic arms. Wireless terminals in smart homes can be televisions, air conditioners, robot vacuums, speakers, or set-top boxes. Terminal devices typically contain communication modules, circuits, or chips that perform corresponding communication functions, and they also contain program instructions for performing those functions.

[0119] Please see Figure 4 , Figure 4 This is a schematic diagram of another possible, non-limiting system used in the communication method and related apparatus of this application. The communication system adopts a network architecture of server-network-terminal equipment. Figure 4 As shown, the communication system includes one or more of the following:

[0120] 1. UE (also known as terminal device): For a detailed description of the terminal device, please refer to the preceding text. Figure 3 The descriptions in the corresponding embodiments will not be repeated here;

[0121] 2. Data Network (DN), used to provide services such as carrier services, internet access, or third-party services. Servers (such as...) Figure 4 The server in this application can be considered part of the DN. Optionally, in this application, an AI model is deployed on the server to implement AI processing capabilities, AI-related functions, and / or AI-related services.

[0122] 3. Access Network (AN): including, for example Figure 3 The RAN node 100 is shown. For a detailed description of RAN node 100, please refer to the preceding text. Figure 3 The descriptions in the corresponding embodiments will not be repeated here;

[0123] 4. Core Network: Used to perform three main functions: registration, connection, and session management. The core network mainly includes Network Exposure Function (NEF) network elements, Policy Control Function (PCF) network elements, Application Function (AF) network elements, Access and Mobility Management Function (AMF) network elements, Session Management Function (SMF) network elements, and User Plane Function (UPF) network elements, etc.

[0124] The following section introduces the various network elements in the core network.

[0125] NEF network element: Used to expose the services and capabilities of 3GPP network functions to AF network element, and at the same time, it can also allow AF network element to provide information to 3GPP network functions. The corresponding interface is N33 interface.

[0126] PCF network element: Used for policy management of charging and QoS policies;

[0127] AF (AF Element): Used to transmit the application's requirements to the network side;

[0128] AMF network elements are primarily used for mobility management, access authentication / authorization, and other functions. They are also responsible for transmitting user policies between the UE and PCF network elements. The N1 interface is the signaling plane interface between the UE and AMF network elements; since the UE cannot directly interact with the core network, it needs to pass NAS (non-access stratum) information through the AN. The N2 interface is the signaling plane interface through which the AMF network element requests resources from the AN for Protocol Data Unit (PDU) sessions.

[0129] SMF network element: Used to complete session management functions such as Internet Protocol (IP) address allocation for UE, UPF selection, and billing and QoS policy control;

[0130] UPF network elements: As the interface with the data network, they are used to perform functions such as user plane data forwarding, session / flow-based billing statistics, and bandwidth limiting. The N3 interface is the interface between the radio access network (RAN) and UPF network elements, mainly used to transmit uplink and downlink user plane data between the 5G RAN and UPF network elements.

[0131] Optionally, AI nodes may be introduced into the network to support artificial intelligence (AI) technology.

[0132] AI nodes can be deployed in one or more of the following locations within the communication system: access network nodes (RAN nodes), terminal devices, or core network devices. Alternatively, AI nodes can be deployed independently, for example, in a location other than any of the aforementioned devices, such as in the host or cloud server of an over-the-top (OTT) system. AI nodes can communicate with other devices in the communication system, which can be one or more of the following: network devices, terminal devices, or core network elements.

[0133] It is understood that this application does not limit the number of AI nodes. For example, when there are multiple AI nodes, these nodes can be divided based on function, such as different AI nodes being responsible for different functions.

[0134] It can also be understood that AI nodes can be independent devices, or they can be integrated into the same device to achieve different functions. Alternatively, they can be network elements in hardware devices, software functions running on dedicated hardware, or virtualization functions instantiated on a platform (e.g., a cloud platform). This application does not limit the specific form of the aforementioned AI nodes.

[0135] AI nodes can be AI network elements or AI modules.

[0136] The communication method and related apparatus of this application will be further described below with reference to the accompanying drawings.

[0137] In this application, Figure 2 and Figure 3 The RAN node shown in this application can be replaced with other terms, such as "access network equipment". For ease of description, unless otherwise specified, "access network equipment" will be used throughout this application. It should be understood that the technical solutions provided in this application are also applicable to other different expressions or types of "access network equipment".

[0138] In the communication method of this application, the access network device acts as a server, while the terminal device acts as a computing node, thereby realizing model training and inference. Please refer to... Figure 5 , Figure 5 This is a schematic diagram illustrating the model training implementation in this application. For example... Figure 5 As shown, a model is deployed in the access network device. When the model in the access network device needs to be updated, the access network device sends a training request to the terminal device. If the terminal device sends an acknowledgment message for the training request to the access network device, the access network device sends the model's parameter information to the terminal device, which then trains the model using locally stored data. After training is complete, the terminal device uploads the trained model's parameter information back to the access network device, which then aggregates and updates the parameter information from all the terminal devices.

[0139] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating the implementation of model inference in this application. For example... Figure 6 As shown, when a terminal device needs to use a model for inference, it sends a model access request to the access network device. After receiving the model access request, the access network device sends the model's parameter information to the terminal device, allowing the terminal device to use the model for inference locally.

[0140] The communication method of this application is applied to Figure 5 This illustrates the model training phases to improve model training accuracy. Please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic diagram illustrating one possible implementation of the communication method in this application. It should be understood that this application uses access network devices (including a first access network device, a second access network device, and a third access network device) and terminal devices as examples to illustrate the method, but this application does not limit the entities executing the interaction. For example, Figure 7The method executed by the access network device can also be implemented by chips, baseband chips, modem chips, system-on-chip (SoC) chips containing modem cores, system-in-package (SIP) chips, communication modules, chip systems, processors, logic modules, or software within the access network device. In this application, the term "access network device" can refer to the access network device itself, or to chips, communication modules, integrated circuits, processors, logic modules, or software within the access network device used to implement the communication method provided in this application; no specific limitation is made in this application. Similarly, Figure 7 The method executed by the terminal device can also be implemented by chips, baseband chips, modem chips, SoC chips containing modem cores, SIP chips, communication modules, chip systems, processors, logic modules, or software within the terminal device. In this application, the term "terminal device" can refer to the terminal device itself, or to chips, communication modules, integrated circuits, processors, logic modules, or software within the terminal device used to implement the communication method provided in this application; no specific limitation is made in this application.

[0141] like Figure 7 As shown, the communication method of this application includes, but is not limited to, steps 401 to 404.

[0142] 401. The first access network device sends N first parameter information of the first model to N terminal devices.

[0143] In this application, a first model is deployed in the first access network device. During the training process of the first model, the first access network device acts as a server, distributing model training tasks to various terminal devices, i.e., the terminal devices act as computing nodes participating in model training. Specifically, the first access network device sends N first parameter information of the first model to N terminal devices, where N is an integer greater than or equal to 1. Each of the N terminal devices receives one piece of first parameter information; in other words, the first access network device sends one piece of first parameter information to each of the N terminal devices.

[0144] Optionally, the N terminal devices can be some or all of the terminal devices in the serving cell of the first access network device.

[0145] In this application, the term "first model" can be replaced with other descriptions, such as first AI model, first neural network, first neural network model, first AI neural network model, first machine learning model, first mathematical model, or first AI processing model, etc.

[0146] Optionally, the first parameter information includes the weights and / or biases of the first model.

[0147] Optionally, some terminal devices may not have saved the relevant information of the first model before step 401. Therefore, the first parameter information sent by the first access network device to the terminal device, in addition to the aforementioned weight parameters and / or bias parameters, also includes the architecture definition and activation function of the first model. For example, the type of layers, the number of layers, and the specific configuration of each layer in the first model.

[0148] 402. The first access network device receives N second parameter information from N terminal devices.

[0149] After receiving the first parameter information from the first access network device, each of the N terminal devices combines it with local data to train a model and obtain the second parameter information. Each second parameter information is an updated version of the first parameter information. For example, if the first parameter information includes the weight parameters and / or bias parameters of the first model, then the second parameter information is the updated weight parameters and / or bias parameters.

[0150] Next, each terminal device sends the second parameter information obtained during training to the first access network device. Correspondingly, the first access network device receives N pieces of second parameter information from N terminal devices, where each piece of second parameter information comes from a single terminal device.

[0151] 403. The first access network device updates the first model based on N second parameter information.

[0152] Specifically, the first access network device can aggregate N second parameter information from N terminal devices to update the first model.

[0153] For example, the first access network device may perform aggregation updates in the following ways, including but not limited to: FedAvg, FedProx, or Secure Aggregation.

[0154] Optionally, the training of the first model requires multiple iterations. The first access network device can then repeatedly execute steps 401 to 403 until the convergence condition for training the first model is met. For example, this convergence condition could be that the number of iterations in steps 401 to 403 reaches an upper limit threshold, or that the accuracy of the first model meets the requirements.

[0155] 404. The first access network device sends the updated third parameter information of the first model to the second access network device.

[0156] After the first access network device achieves convergence in training the first model, it then sends the updated third parameter information of the first model to the second access network device.

[0157] Optionally, the third parameter information includes the updated weight parameters and / or bias parameters in the first model of the first access network device.

[0158] Optionally, the second access network device may not have saved the relevant information of the first model before step 404. Therefore, the third parameter information sent by the first access network device to the terminal device, in addition to the aforementioned weight parameters and / or bias parameters, also includes the architecture definition and activation function of the first model. For example, the type of layers, the number of layers, and the specific configuration of each layer in the first model.

[0159] 405. The second access network device updates the first model based on the third parameter information.

[0160] After receiving the third parameter information from the first access network device, the second access network device updates the first model based on the third parameter information. Next, the second access network device acts as a server, and the terminal devices in its serving cell act as computing nodes to continue training the first model in the second access network device. The specific process is similar to steps 401 to 405, and will not be described in detail here.

[0161] It should be understood that, in this application, the AI ​​functions enabled by the first model in the second access network are the same as the AI ​​functions enabled by the first model in the first access network device. Optionally, the parameter information of the first model in the second access network may be the same as or different from the parameter information of the first model in the first access network.

[0162] In this application, only one access network device participates in each training iteration of the model. Furthermore, after one access network device completes the training of the model, the parameter information of the model can be transferred to another access network device for further training. Thus, while keeping the overall amount of training data constant, the scope and number of terminal devices in a single training iteration are reduced, the heterogeneity of the model's training data is decreased, and the accuracy of model training is improved.

[0163] On the other hand, the model's parameter information can be transferred between different access network devices, and the training results of the previous access network device can be accumulated in the next access network device, enabling terminal devices in different regions to participate in the joint training of the model, expanding the richness of the model's training data and improving the model's generalization.

[0164] In one possible implementation, the first access network device is the next-hop node of the third access network device. The third access network device first trains the first model to obtain the fourth parameter information of the first model, which is the updated parameter information of the first model in the third access network device. Next, the third access network device sends the fourth parameter information of the first model to the first access network device. Correspondingly, the first access network device receives the fourth parameter information of the first model from the third access network device. The first access network device updates the first model in the first access network device based on the fourth parameter information. Then, the first access network device sends the N first parameter information of the first model to N terminal devices (i.e., executes step 401).

[0165] It should be understood that, in this application, the AI ​​functions enabled by the first model in the third access network are the same as the AI ​​functions enabled by the first model in the first access network device.

[0166] Optional, in Figure 7 In step 404, as shown, this application provides several optional implementation methods. These will be described below.

[0167] Implementation Method 1: The access network device (e.g., the first access network device) determines the next-hop node and transmits the parameter information of the first model to the next-hop node.

[0168] Please see Figure 8 , Figure 8 This is a schematic diagram illustrating another possible implementation of the communication method in this application. For example... Figure 8 As shown, the communication method in this application includes, but is not limited to, the following steps 501 to 508.

[0169] 501. Initialize the first model to obtain the initial architecture and parameter information of the first model.

[0170] 502. The first access network device determines N terminal devices within its serving cell to participate in the training of the first model. Please refer to the description related to step 401 above for details; it will not be repeated here.

[0171] 503. The first access network device sends N first parameter information to N terminal devices, and the N terminal devices perform model training respectively. Please refer to the description related to step 402 above for details, which will not be repeated here.

[0172] 504. N terminal devices send the second parameter information to the first access network device. Please refer to the description related to step 403 above for details, which will not be repeated here.

[0173] 505. The first access network device updates the first model based on N second parameter information. Please refer to the description related to step 404 above for details, which will not be repeated here.

[0174] 506. The first access network device determines whether the training of the first model has reached the convergence condition. If it has, it means that the first access network device no longer needs to train the first model, and the first access network device executes step 507; if it has not, it means that the first access network device still needs to train the first model, and the first access network device executes the next round of training for the first model, that is, jumps to step 503.

[0175] 507. The first access network device determines whether the training task of the first model has been completed. If yes, it means that the first model no longer needs to be trained, and the first access network device ends the process; if no, it proceeds to step 508.

[0176] 508. The first access network device selects another access network device as the next-hop node based on indicators such as physical distance and / or channel conditions. In this application, the second access network device is the next-hop node of the first access network device. Then, the first access network device sends the third parameter information of the first model to the second access network device. The second access network device then continues to execute step 502 above.

[0177] For example, the first access network device can select the nearest access network device as the next-hop node based on physical distance; or, the first access network device can also select the access network device with the lowest transmission latency as the next-hop node based on channel conditions. Thus, the deployment of method one is relatively simple, robust, and highly versatile for various federated frameworks and model structures. Furthermore, the algorithm has strong real-time performance and can effectively cope with unexpected situations.

[0178] Implementation Method 2: The core network equipment plans the transmission path of the parameter information of the first model.

[0179] Please see Figure 9 , Figure 9 This is a schematic diagram illustrating another possible implementation of the communication method in this application. For example... Figure 9 As shown, the communication method in this application includes, but is not limited to, the following steps 601 to 609.

[0180] 601. Initialize the first model to obtain the initial architecture and parameter information of the first model.

[0181] 602. The core network device pre-plans the transmission path for the parameter information (e.g., the third parameter information) of the first model. After receiving the parameter information of the first model, the access network device on the transmission path performs training on the first model. Optionally, the core network device sends first information to the first access network device, which is used to indicate that the second access network device is the next-hop node of the first access network device.

[0182] 603. The first access network device determines N terminal devices within its serving cell to participate in the training of the first model. Please refer to the description related to step 401 above for details; it will not be repeated here.

[0183] 604. The first access network device sends N first parameter information to N terminal devices, and the N terminal devices perform model training respectively. Please refer to the description related to step 402 above for details, which will not be repeated here.

[0184] 605. N terminal devices send the second parameter information to the first access network device. Please refer to the description related to step 403 above for details, which will not be repeated here.

[0185] 606. The first access network device updates the first model based on N second parameter information. Please refer to the description related to step 404 above for details, which will not be repeated here.

[0186] 607. The first access network device determines whether the training of the first model has reached the convergence condition. If it has, it means that the first access network device no longer needs to train the first model, and the first access network device executes step 608; if it has not, it means that the first access network device still needs to train the first model, and the first access network device executes the next round of training for the first model, that is, jumps to step 604.

[0187] 608. The first access network device determines whether the training task of the first model has been completed. If yes, it means that the first model no longer needs to be trained, and the first access network device ends the process; if no, it proceeds to step 609.

[0188] 609. Based on the first information from the core network device, the first access network device determines the second access network device as the next-hop node. Then, the first access network device sends the third parameter information of the first model to the second access network device. The second access network device then continues to execute step 603 above.

[0189] In the second implementation method, since the core network device plans the transmission path of the parameter information (e.g., the third parameter information) of the first model in advance, when the training of the first model in the first access network device is completed and converged (i.e., the convergence condition is met in step 606), the first access network device can transmit the third parameter information of the first model to the next hop node (the second access network device) more quickly, without the first access network device having to plan the route in real time, thus saving the computational overhead of the first access network device and speeding up the training process.

[0190] Optional, in Figure 7 In step 401 as shown, this application provides several optional implementation methods. These will be described below.

[0191] Implementation Method A: The first access network device sends the complete parameter information of the first model to each of the N terminal devices. That is, each of the N first parameter information sets represents the complete parameter information of the first model; therefore, the N first parameter information sets can be considered identical. Thus, each terminal device possesses the complete parameters of the model, allowing for more efficient use of the computing power of each terminal device and reducing the waste of computing resources caused by model segmentation. Furthermore, having complete parameters for the model at each terminal device simplifies the integration and updating of the first model for the first access network device, reducing its computational overhead.

[0192] Implementation Method B: The first model can be divided into M sub-models, where M is an integer greater than or equal to 1. The N first parameter messages sent by the first access network device to N terminal devices include parameter information for the M sub-models. Each of the N first parameter messages represents the parameter information of one of the M sub-models. Therefore, the aforementioned "N first parameter messages sent by the first access network device to the N terminal devices" can be understood as the first access network device sending parameter information for one of the sub-models of the first model to each of the N terminal devices. Thus, by splitting the complete first model into multiple sub-models and sending the parameter information of these sub-models to each terminal device for training, the communication overhead required for transmitting parameter information can be reduced, the computing power overhead of the terminal devices can be reduced, and computational efficiency can be improved.

[0193] Optionally, in implementation B, the parameter scales and / or modalities of the input data differ among the M sub-models. The following sections describe the possible implementation scenarios for the M sub-models in this application.

[0194] Scenario 1: The M sub-models have different parameter scales, but the modalities of their input data are the same, and the AI ​​functions enabled by the M sub-models are also the same. In other words, the M sub-models are sub-models of the first model with different sizes and parameter scales.

[0195] For example, please refer to Figure 10 , Figure 10 This is a schematic diagram illustrating one implementation of the M sub-models in the first model of this application. For example... Figure 10 As shown, the first access network device can divide the M sub-models into basic sub-models according to their different sizes and parameter scales. Figure 10 The B model in the model), the enhanced sub-model ( Figure 10 The E model and the complete sub-model (in the model) Figure 10 In the C model, the complete sub-model can be the sub-model with the largest parameter size in the first sub-model, or the complete sub-model is the first model, where the parameter size of the complete sub-model is larger than that of the enhanced sub-model, and the parameter size of the enhanced sub-model is larger than that of the basic sub-model. Therefore, the first access network device can send parameter information of a sub-model (i.e., the first parameter information) to each terminal device based on the training conditions (e.g., computing power resources, storage resources, and / or network resources) and the model usage requirements, so that the terminal device can train the sub-model. This improves the flexibility of model training; the first access network device can send parameter information of sub-models with corresponding parameter sizes to terminal devices with different computing capabilities and usage requirements, further reducing communication costs and improving training efficiency.

[0196] Optionally, the first access network device can store a complete first model, or it can store M sub-models from the first model. See also... Figure 11 , Figure 11 This is a schematic diagram illustrating another implementation of the M sub-models in the first model of this application. For example... Figure 11 As shown, if the first access network device stores a complete first model, before step 401, the first access network device first compresses (e.g., prunes) the first model into M sub-models of different sizes and parameter scales, and then sends the parameter information of a sub-model (i.e., the first parameter information) to each terminal device; if the first access network device stores M sub-models of the first model, then the first access network device can directly send the parameter information of a sub-model (i.e., the first parameter information) to each terminal device.

[0197] After training the sub-model, the terminal device sends second parameter information to the first access network device. This second parameter information contains the parameter information of the trained sub-model. If a terminal device has trained the complete first model, the first access network device can extract the parameter information of each sub-model from the second parameter information reported by that terminal device to update the sub-models in the first access network device. (See also: [link to relevant documentation]) Figure 12 , Figure 12 This is a schematic diagram illustrating another implementation of the M sub-models in the first model of this application. For example... Figure 12 As shown, if a terminal device is training a certain sub-model, the first access network device, after receiving the second parameter information of the terminal device, updates the sub-model corresponding to the second parameter information. Furthermore, after the sub-model corresponding to the second parameter information is updated, it can also be used to assist in updating other sub-models (e.g., using knowledge distillation). The model training method of knowledge distillation in this application will be described below as an example.

[0198] Please see Figure 13 , Figure 13 This is a schematic diagram illustrating one implementation of knowledge distillation in this application. For example... Figure 13 As shown, the terminal device trains a certain sub-model. After receiving the second parameter information from the terminal device, the first access network device updates the sub-model corresponding to the second parameter information. In the subsequent knowledge distillation process, the first access network device collects training data from a public database, using the sub-model corresponding to the second parameter information as the "teacher model" and other sub-models in the first access network device as "student models" to update the other sub-models. Specifically, input data (such as...) is selected from the public database. Figure 13 The input x shown is fed into the teacher model and the student model respectively, yielding the output values ​​of the teacher model and the student model. Then, the gradient between the output values ​​of the teacher model and the student model is calculated, and the relationship between the student model's output value and the label value of the input data is also calculated. Figure 13 The gradient between the label y) shown above. Based on the above two gradients, the student model is updated.

[0199] The public database can be stored in the first access network device, or in the core network device or a third-party public storage device. When the public database is stored in the first access network device, it can be transmitted to the second access network device along with the third parameter information in step 404. When the public database is stored in the core network device or a third-party public storage device, it can also be used by other access network devices (the second access network device and the third access network device).

[0200] In one possible implementation, after a sub-model in the first access network device reaches the training convergence condition, the first access network device can send the parameter information of the converged sub-model to the second access network device without waiting for other sub-models to complete training convergence. That is, the third parameter information in this application is the parameter information of one of the M sub-models. Optionally, this sub-model can be a sub-model in the first access network device that has completed training convergence. After receiving the third parameter information, the second access network device can perform auxiliary updates to the first model (including sub-models) in the second access network device based on this third parameter information. This allows the second access network device to participate in the training of the first model more quickly, improving the efficiency of model training.

[0201] On the other hand, after the other sub-models in the first access network device have successively completed training and converged, the first access network device continues to send the parameter information of these sub-models to the second access network device. Optionally, the second access network device performs auxiliary updates on the other sub-models received subsequently based on the updated sub-models, thereby reducing the process of model distribution and training of other sub-models on the second access network device, improving the efficiency of model training, and reducing the communication overhead required to transmit the parameter information of other sub-models.

[0202] Please see Figure 14 , Figure 14 This is a schematic diagram illustrating another implementation of the communication method in this application. For example... Figure 14 As shown, the communication method includes, but is not limited to, steps 701 to 705.

[0203] 701. The first access network device sends the parameter information of M sub-models to N terminal devices. Each of the N terminal devices trains the received sub-model and sends the trained sub-model's parameter information back to the first access network device. The sub-model with smaller parameter size (hereinafter referred to as the small model) will complete training faster than the model / sub-model with larger parameter size (hereinafter referred to as the large model), meaning the small model will converge faster in the first access network device.

[0204] 702. The first access network device sends the parameter information of the pre-trained small model to the second access network device.

[0205] 703. While the first access network device is still training the large model, the second access network device can simultaneously train the small model because it has already received the parameter information of the small model.

[0206] 704. After the first access network device has trained the large model, it sends the parameter information of the large model to the second access network device.

[0207] 705. After receiving the parameter information of the large model, the second access network device can train the large model based on the small model, so that the second access network device does not need to send the parameter information of the large model to the terminal devices covered by the second access network device.

[0208] In one possible implementation, the first access network device can wait until all M sub-models have completed training and converged before sending the parameter information of the M converged sub-models together to the second access network device. That is, the third parameter information includes the parameter information of each of the M sub-models. Therefore, the second access network device can obtain richer and more complete parameter information, enabling it to update the first model in the second access network device more accurately.

[0209] Scenario 2: If the input data of the M sub-models have different modalities, then each of the M sub-models is used to perform tasks of a different modality. Therefore, the parameter scales of the M sub-models are also different. Thus, in the N first parameter information sent by the first access network device to the N terminal devices, each first parameter information represents the parameter information of a single-modality sub-model. For example, in... Figure 15 In this model, the different modalities of the M sub-models include, but are not limited to, image modality, text modality, and / or audio modality. After training a sub-model for a single modality, the terminal device reports the parameter information (i.e., second parameter information) of the sub-model for that single modality to the first access network. The first access network device updates the corresponding sub-model based on the second parameter information and, based on this sub-model, assists in updating the first model. Thus, the basic capabilities of the first model are trained, as well as its ability to process specific modalities of people.

[0210] Specifically, the training process of each sub-model and the first model in Scenario 2 is similar to that in Scenario 1, and will not be repeated here.

[0211] Accordingly, this application also provides related apparatus for implementing the above-described scheme. Please refer to [link / reference]. Figure 16 , Figure 16 This is a schematic diagram of a communication device 800 provided in an embodiment of this application. The communication device 800 can realize the functions of the first access network device in the above method embodiments, and therefore can also achieve the beneficial effects of the above method embodiments. In this embodiment, the communication device 800 can be the first access network device, or it can be an integrated circuit or component inside the first access network device, such as a chip, baseband chip, modem chip, SoC chip containing a modem core, system-in-package (SIP) chip, communication module, chip system, processor, etc.

[0212] like Figure 16As shown, the communication device 800 includes a transceiver unit 801 and a processing unit 802. Optionally, the transceiver unit 801 may include a transmitting unit and a receiving unit, respectively used to perform transmitting and receiving.

[0213] The transceiver unit 801 is used to send N first parameter information of the first model to N terminal devices, and each terminal device is used to receive one first parameter information, where N is an integer greater than or equal to 1;

[0214] The transceiver unit 801 is also used to receive N second parameter information from N terminal devices, where each second parameter information is obtained by updating a first parameter information.

[0215] Processing unit 802 is used to update the first model based on N second parameter information;

[0216] The transceiver unit 801 is also used to send the updated third parameter information of the first model to the second access network device, so that the second access network device can update the first model based on the third parameter information.

[0217] In one possible design, each first parameter information is the parameter information of the complete model of the first model.

[0218] In one possible design, the first model includes M sub-models, and each first parameter information is the parameter information of one of the M sub-models, where M is an integer greater than or equal to 1.

[0219] In one possible design, the M sub-models have different parameter scales and / or different modalities of input data.

[0220] In one possible design, the third parameter information is the parameter information of one of the M sub-models.

[0221] In one possible design, the third parameter information includes the parameter information of each of the M sub-models.

[0222] In one possible design, the processing unit 802 is also used for the first access network device to determine the second access network device as the next-hop node.

[0223] In one possible design, the transceiver unit 801 is also used to receive first information from the core network device, the first information being used to indicate that the second access network device is the next-hop node of the first access network device.

[0224] In one possible design, the transceiver unit 801 is also used to receive fourth parameter information from the first model of the third access network device;

[0225] The processing unit 802 is also used to update the first model based on the fourth parameter information.

[0226] In one possible design, the first model includes a first sub-model and a second sub-model, the fourth parameter information is the parameter information of the first sub-model, and the transceiver unit 801 is also used to receive the fifth parameter information from the third access network device, the fifth parameter information is the parameter information of the second sub-model;

[0227] The processing unit 802 is also used to update the second sub-model based on the first sub-model and the fifth parameter information.

[0228] It should be noted that the information interaction and execution process between the modules / units in the communication device 800 are different from those in this application. Figures 7 to 9 The corresponding method embodiments are based on the same concept, and the details can be found in the descriptions of the method embodiments shown above in this application, which will not be repeated here.

[0229] Please see Figure 17 This is another schematic structural diagram of the communication device 900 provided in this application. The communication device 900 includes a logic circuit 901 and an input / output interface 902. The communication device 900 can be a chip or an integrated circuit.

[0230] in, Figure 16 The transceiver unit 801 shown can be a communication interface, which can be... Figure 17 The input / output interface 902 in the communication interface may include an input interface and an output interface. Alternatively, the communication interface may also be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0231] The logic circuit 901 and the input / output interface 902 can also perform other steps executed by the first access network device in any embodiment and achieve corresponding beneficial effects, which will not be elaborated here.

[0232] In one possible implementation, Figure 16 The processing unit 802 shown can be Figure 17 The logic circuit 901 in the middle.

[0233] Optionally, the logic circuit 901 can be a processing device, the functions of which can be partially or entirely implemented in software.

[0234] Optionally, the processing apparatus may include a memory and a processor, wherein the memory is used to store a computer program, and the processor reads and executes the computer program stored in the memory to perform the corresponding processing and / or steps in any of the method embodiments.

[0235] Optionally, the processing device may consist of only a processor. A memory for storing computer programs is located outside the processing device, and the processor is connected to the memory via circuitry / wires to read and execute the computer programs stored in the memory. The memory and processor may be integrated together or physically independent of each other.

[0236] Optionally, the processing device may be one or more chips, or one or more integrated circuits. For example, the processing device may be one or more field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), system on-chips (SoCs), central processing units (CPUs), network processors (NPs), digital signal processors (DSPs), microcontroller units (MCUs), programmable logic controllers (PLDs), or other integrated chips, or any combination of the above chips or processors.

[0237] Please see Figure 18 The above-described embodiment of the present application provides a structural schematic diagram of the communication device 1000 involved in the embodiment. Specifically, the communication device 1000 can be the communication device that serves as the first access network device in the above embodiment.

[0238] The communication device 1000 includes at least one processor 1011 and at least one network interface 1014. Optionally, the communication device further includes at least one memory 1012, at least one transceiver 1013, and one or more antennas 1014. The processor 1011, memory 1012, transceiver 1013, and network interface 1014 are connected, for example, via a bus. In this embodiment, the connection may include various interfaces, transmission lines, or buses, etc., and this embodiment is not limited thereto. The antenna 1015 is connected to the transceiver 1013. The network interface 1014 enables the communication device to communicate with other communication devices through a communication link. For example, the network interface 1014 may include a network interface between the communication device and core network equipment, such as an S1 interface; the network interface may also include a network interface between the communication device and other communication devices (e.g., other network devices or core network equipment), such as an X2 or Xn interface.

[0239] in, Figure 16The transceiver unit 801 shown can be a communication interface, which can be... Figure 18 The network interface 1014 may include an input interface and an output interface. Alternatively, the network interface 1014 may also be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0240] The processor 1011 is primarily used to process communication protocols and communication data, control the entire communication device, execute software programs, and process data from the software programs, for example, to support the communication device in performing the actions described in the embodiments. The communication device may include a baseband processor and a central processing unit (CPU). The baseband processor is primarily used to process communication protocols and communication data, while the CPU is primarily used to control the entire terminal device, execute software programs, and process data from the software programs. Figure 18 The processor 1011 can integrate the functions of a baseband processor and a central processing unit. Those skilled in the art will understand that the baseband processor and the central processing unit can also be independent processors interconnected via technologies such as buses. Those skilled in the art will understand that a terminal device can include multiple baseband processors to adapt to different network standards, and a terminal device can include multiple central processing units to enhance its processing capabilities. The various components of the terminal device can be connected via various buses. The baseband processor can also be described as a baseband processing circuit or a baseband processing chip. The central processing unit can also be described as a central processing circuit or a central processing chip. The function of processing communication protocols and communication data can be built into the processor or stored in memory as a software program, with the processor executing the software program to implement the baseband processing function.

[0241] The memory is primarily used to store software programs and data. The memory 1012 can exist independently or be connected to the processor 1011. Optionally, the memory 1012 can be integrated with the processor 1011, for example, integrated within a single chip. The memory 1012 can store program code that executes the technical solutions of the embodiments of this application, and its execution is controlled by the processor 1011. The various types of computer program code being executed can also be considered as drivers for the processor 1011.

[0242] Figure 18 Only one memory and one processor are shown. In actual terminal devices, there may be multiple processors and multiple memories. Memory can also be called storage medium or storage device, etc. Memory can be a storage element on the same chip as the processor, i.e., an on-chip storage element, or it can be a separate storage element; this application does not limit this.

[0243] Transceiver 1013 can be used to support the reception or transmission of radio frequency (RF) signals between a communication device and a terminal. Transceiver 1013 can be connected to antenna 1015. Transceiver 1013 includes a transmitter Tx and a receiver Rx. Specifically, one or more antennas 1015 can receive RF signals. The receiver Rx of transceiver 1013 is used to receive the RF signals from the antennas, convert the RF signals into digital baseband signals or digital intermediate frequency (IF) signals, and provide the digital baseband signals or IF signals to processor 1011 so that processor 1011 can perform further processing on the digital baseband signals or IF signals, such as demodulation and decoding. In addition, the transmitter Tx in transceiver 1013 is also used to receive modulated digital baseband signals or IF signals from processor 1011, convert the modulated digital baseband signals or IF signals into RF signals, and transmit the RF signals through one or more antennas 1015. Specifically, the receiver Rx can selectively perform one or more stages of downmixing and analog-to-digital conversion on the radio frequency signal to obtain a digital baseband signal or a digital intermediate frequency (IF) signal. The order of these downmixing and IF conversion processes is adjustable. The transmitter Tx can selectively perform one or more stages of upmixing and digital-to-analog conversion on the modulated digital baseband signal or digital IF signal to obtain a radio frequency signal. The order of these upmixing and IF conversion processes is also adjustable. The digital baseband signal and the digital IF signal can be collectively referred to as digital signals.

[0244] The transceiver 1013 can also be called a transceiver unit, transceiver, transceiver device, etc. Optionally, the device in the transceiver unit that performs the receiving function can be regarded as the receiving unit, and the device in the transceiver unit that performs the transmitting function can be regarded as the transmitting unit. That is, the transceiver unit includes a receiving unit and a transmitting unit. The receiving unit can also be called a receiver, input port, receiving circuit, etc., and the transmitting unit can be called a transmitter, transmitter, or transmitting circuit, etc.

[0245] It should be noted that, Figure 18 The communication device 1000 shown can be used to implement the steps implemented by the network device in the aforementioned method embodiments, and to achieve the corresponding technical effects of the network device. Figure 18 The specific implementation of the communication device 1000 shown can be referred to the description in the foregoing method embodiments, and will not be repeated here.

[0246] This application also provides a chip device, including a processor, configured to call computer programs or computer instructions stored in the memory, so that the processor executes the above-described... Figures 7 to 9 The method provided in the illustrated embodiment.

[0247] In one possible implementation, the input of the chip device corresponds to the above. Figures 7 to 9In any of the embodiments shown, the receiving operation of the chip device corresponds to the above-described... Figures 7 to 9 The sending operation in any of the embodiments shown.

[0248] Optionally, the processor is coupled to the memory via an interface.

[0249] Optionally, the chip device may also include a memory that stores computer programs or computer instructions.

[0250] The processor mentioned above can be a general-purpose central processing unit, a microprocessor, an application-specific integrated circuit (ASIC), or one or more devices used to control the above. Figures 7 to 9 The illustrated embodiments provide an integrated circuit for program execution of the method provided in any of the embodiments. The memory mentioned above may be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, such as random access memory (RAM).

[0251] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0252] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0253] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0254] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0255] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms. Whether a function is implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0256] It should be understood that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines.

[0257] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0258] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to it, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0259] In the various embodiments of this application, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions between different embodiments are consistent and can be referenced by each other. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0260] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A communication method, characterized in that, include: The first access network device sends N first parameter information of the first model to N terminal devices, and each terminal device is used to receive one first parameter information, where N is an integer greater than or equal to 1; The first access network device receives N second parameter information from the N terminal devices, where each second parameter information is obtained by updating the first parameter information. The first access network device updates the first model based on the N second parameter information; The first access network device sends the updated third parameter information of the first model to the second access network device, so that the second access network device can update the first model based on the third parameter information.

2. The method according to claim 1, characterized in that, Each of the first parameter information is the parameter information of the complete model of the first model.

3. The method according to claim 1, characterized in that, The first model includes M sub-models, and each of the first parameter information is the parameter information of one of the M sub-models, where M is an integer greater than or equal to 1.

4. The method according to claim 3, characterized in that, The M sub-models have different parameter scales and / or different input data modalities.

5. The method according to claim 3 or 4, characterized in that, The third parameter information is the parameter information of one of the M sub-models.

6. The method according to claim 3 or 4, characterized in that, The third parameter information includes the parameter information of each of the M sub-models.

7. The method according to any one of claims 1 to 6, characterized in that, Before the first access network device sends the updated third parameter information of the first model to the second access network device, the method further includes: The first access network device determines the second access network device as the next-hop node.

8. The method according to any one of claims 1 to 6, characterized in that, The method further includes: The first access network device receives first information from the core network device, the first information being used to indicate that the second access network device is the next-hop node of the first access network device.

9. The method according to any one of claims 1 to 8, characterized in that, Before the first access network device sends the N first parameter information of the first model to the N terminal devices, the method further includes: The first access network device receives the fourth parameter information of the first model from the third access network device; The first access network device updates the first model based on the fourth parameter information.

10. The method according to claim 9, characterized in that, The first model includes a first sub-model and a second sub-model, and the fourth parameter information is the parameter information of the first sub-model. The method further includes: The first access network device receives fifth parameter information from the third access network device, wherein the fifth parameter information is the parameter information of the second sub-model; The first access network device updates the second sub-model based on the first sub-model and the fifth parameter information.

11. A communication device, characterized in that, It includes at least one processor coupled to a memory; the at least one processor is used to perform the method as described in any one of claims 1 to 18.

12. The communication device according to claim 11, characterized in that, The communication device is a chip or chip system.

13. A readable storage medium, characterized in that, The storage medium stores a computer program or instructions, which, when executed by a communication device, implement the method as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, When the computer program product is run on a computer, it causes the computer to perform the method as described in any one of claims 1 to 10.

Images