WO2023109827A1 - 客户端筛选方法及装置、客户端及中心设备 - Google Patents

客户端筛选方法及装置、客户端及中心设备 Download PDF

Info

Publication number
WO2023109827A1
WO2023109827A1 PCT/CN2022/138755 CN2022138755W WO2023109827A1 WO 2023109827 A1 WO2023109827 A1 WO 2023109827A1 CN 2022138755 W CN2022138755 W CN 2022138755W WO 2023109827 A1 WO2023109827 A1 WO 2023109827A1
Authority
WO
WIPO (PCT)
Prior art keywords
client
model
training
candidate
clients
Prior art date
Application number
PCT/CN2022/138755
Other languages
English (en)
French (fr)
Inventor
孙布勒
孙鹏
杨昂
Original Assignee
维沃移动通信有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 维沃移动通信有限公司 filed Critical 维沃移动通信有限公司
Publication of WO2023109827A1 publication Critical patent/WO2023109827A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]

Definitions

  • the present application belongs to the technical field of communications, and in particular relates to a client screening method and device, a client and a central device.
  • AI Artificial Intelligence
  • neural network decision tree
  • support vector machine support vector machine
  • Bayesian classifier Bayesian classifier
  • the AI model of the terminal is generally trained offline by the network side, and then delivered to the terminal for execution. This is because the amount of data for a single terminal is limited, and it is difficult to train a better model. After receiving the model, the terminal generally performs fine-tuning based on a small amount of data to obtain better performance. Federated learning or federated meta-learning can be trained without exposing terminal data, which is a very promising direction.
  • the terminal uses its own data to update local model parameters or losses, and the terminal then aggregates the parameters or losses to the server for processing to obtain the global model, and the server then sends the global model to the terminal for a new round training.
  • the purpose of federated learning is to obtain a model that can converge to all terminals participating in training.
  • the purpose of federated meta-learning is to obtain a model initialization parameter that can quickly converge in new scenarios based on the terminal data participating in training.
  • Federated meta-learning can achieve better performance than federated learning when the amount of fine-tuning data at the terminal is small or the fine-tuning convergence time (or number of iterations) is high.
  • all candidate clients will participate in model training of specific federated learning or federated meta-learning, resulting in slow convergence speed of model training and high communication resource overhead between the central device and the client.
  • Embodiments of the present application provide a client screening method and device, a client and a central device, which can improve the convergence speed of model training.
  • a client screening method including:
  • the central device sends a first indication to the client, instructing the client to participate in model training of specific federated learning or federated meta-learning;
  • the central device receives the training result reported by the client, and the training result is a result or an intermediate result after the client performs a round of model training.
  • a client screening device including:
  • a sending module configured to send a first indication to the client, indicating that the client participates in model training of specific federated learning or federated meta-learning
  • the receiving module is configured to receive the training result reported by the client, where the training result is a result or an intermediate result after the client performs a round of model training.
  • a client screening method including:
  • the client receives a first instruction from the central device, the first instruction is used to instruct the client to participate in model training of specific federated learning or federated meta-learning;
  • the client performs specific federated learning or federated meta-learning model training, and reports the training result to the central device, and the training result is a result or an intermediate result after the client performs a round of model training.
  • a client screening device including:
  • a receiving module configured to receive a first instruction from the central device, where the first instruction is used to instruct the client to participate in model training of specific federated learning or federated meta-learning;
  • the reporting module is used to perform model training of specific federated learning or federated meta-learning, and report the training result to the central device, and the training result is the result or an intermediate result after the client performs a round of model training.
  • a central device includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and when the programs or instructions are executed by the processor, the The steps of the method as described in the first aspect.
  • a central device including a processor and a communication interface, wherein the communication interface is used to send a first indication to a client, instructing the client to participate in model training of specific federated learning or federated meta-learning ; receiving a training result reported by the client, the training result being a result or an intermediate result after the client performs a round of model training.
  • a client in a seventh aspect, includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and the programs or instructions are implemented when executed by the processor. The steps of the method as described in the third aspect.
  • a client including a processor and a communication interface, wherein the communication interface is used to receive a first instruction from a central device, and the first instruction is used to instruct the client to participate in a specific federated learning or federated meta-learning model training; the processor is used for specific federated learning or federated meta-learning model training, and reports the training result to the central device, and the training result is after the client performs a round of model training results or intermediate results.
  • a client screening system including: a central device and a client, the central device can be used to perform the steps of the client screening method as described in the first aspect, and the client can be used to perform the steps as described in the first aspect The steps of the client screening method described in the third aspect.
  • a readable storage medium is provided, and a program or an instruction is stored on the readable storage medium, and when the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented, or the steps of the method as described in the first aspect are implemented, or the The steps of the method described in the third aspect.
  • a chip in an eleventh aspect, includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run a program or an instruction to implement the method described in the first aspect. method, or implement the method as described in the third aspect.
  • a computer program/program product is provided, the computer program/program product is stored in a storage medium, and the computer program/program product is executed by at least one processor to implement the The client screening method, or the steps for realizing the client screening method as described in the third aspect.
  • the central device does not require all candidate clients to participate in the model training of specific federated learning or federated meta-learning, but the central device first screens the candidate clients and determines the clients that need model training , and then send a first indication to the client, instructing the client to participate in specific federated learning or federated meta-learning model training, and receive the training result reported by the client.
  • some candidate clients with poor conditions can be eliminated, the convergence speed of training can be improved, and the communication resource overhead between the central device and the client can be reduced.
  • FIG. 1 is a block diagram of a wireless communication system to which an embodiment of the present application is applicable;
  • FIG. 2 is a schematic diagram of channel state information feedback
  • Figure 3 is a schematic diagram of the performance of AI training with different iterations
  • FIG. 4 is a schematic flow diagram of a client screening method at the central device side according to an embodiment of the present application
  • FIG. 5 is a schematic flow diagram of a client side client screening method according to an embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of a communication device according to an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of a terminal according to an embodiment of the present application.
  • FIG. 8 is a schematic structural diagram of a network side device according to an embodiment of the present application.
  • first, second and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific sequence or sequence. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application are capable of operation in sequences other than those illustrated or described herein and that "first" and “second” distinguish objects. It is usually one category, and the number of objects is not limited. For example, there may be one or more first objects.
  • “and/or” in the description and claims means at least one of the connected objects, and the character “/” generally means that the related objects are an "or” relationship.
  • LTE Long Term Evolution
  • LTE-Advanced LTE-Advanced
  • LTE-A Long Term Evolution-Advanced
  • CDMA Code Division Multiple Access
  • TDMA Time Division Multiple Access
  • FDMA Frequency Division Multiple Access
  • OFDMA Orthogonal Frequency Division Multiple Access
  • SC-FDMA Single-carrier Frequency Division Multiple Access
  • system and “network” in the embodiments of the present application are often used interchangeably, and the described technology can be used for the above-mentioned system and radio technology, and can also be used for other systems and radio technologies.
  • the following description describes the New Radio (New Radio, NR) system for example purposes, and uses NR terminology in most of the following descriptions, but these techniques can also be applied to applications other than NR system applications, such as the 6th generation (6th Generation , 6G) communication system.
  • 6G 6th generation
  • Fig. 1 shows a block diagram of a wireless communication system to which the embodiment of the present application is applicable.
  • the wireless communication system includes a terminal 11 and a network side device 12 .
  • the terminal 11 can be a mobile phone, a tablet computer (Tablet Personal Computer), a laptop computer (Laptop Computer) or a notebook computer, a personal digital assistant (Personal Digital Assistant, PDA), a palmtop computer, a netbook, a super mobile personal computer (ultra-mobile personal computer, UMPC), mobile Internet device (Mobile Internet Device, MID), augmented reality (augmented reality, AR) / virtual reality (virtual reality, VR) equipment, robot, wearable device (Wearable Device) , Vehicle User Equipment (VUE), Pedestrian User Equipment (PUE), smart home (home equipment with wireless communication functions, such as refrigerators, TVs, washing machines or furniture, etc.), game consoles, personal computers (personal computer, PC), teller machine or self-service machine and other terminal side devices, wearable devices include: smart watches, smart bracelet
  • the network side device 12 may include an access network device or a core network device, where the access network device 12 may also be called a radio access network device, a radio access network (Radio Access Network, RAN), a radio access network function, or Wireless access network unit.
  • RAN Radio Access Network
  • RAN Radio Access Network
  • Wireless access network unit Wireless access network unit
  • the access network device 12 may include a base station, a WLAN access point, or a WiFi node, etc., and the base station may be called a Node B, an evolved Node B (eNB), an access point, a Base Transceiver Station (Base Transceiver Station, BTS), a radio Base station, radio transceiver, Basic Service Set (BSS), Extended Service Set (ESS), Home Node B, Home Evolved Node B, Transmitting Receiving Point (TRP) or all As long as the same technical effect is achieved, the base station is not limited to specific technical terms. It should be noted that in this embodiment of the application, only the base station in the NR system is used as an example for introduction, and The specific type of the base station is not limited.
  • the main way to improve the performance of the 5th Generation (5G) network with the help of AI is to enhance or replace existing algorithms or processing modules through algorithms and models based on neural networks.
  • algorithms and models based on neural networks can achieve better performance than deterministic algorithms.
  • the more commonly used neural networks include deep neural networks, convolutional neural networks, and recurrent neural networks.
  • the construction, training and verification of neural networks can be realized.
  • Selecting representative terminals for federated training can not only improve the training efficiency, but also improve the generalization performance of the model. Since the training purposes of federated learning and federated meta-learning are different, the terminal screening schemes corresponding to these two training methods should also be different.
  • the embodiment of this application provides a client screening method, as shown in Figure 4, including:
  • Step 101 the central device sends a first indication to the client, instructing the client to participate in model training of specific federated learning or federated meta-learning;
  • Step 102 The central device receives the training result reported by the client, and the training result is a result or an intermediate result after the client performs a round of model training.
  • the above training results are the results or intermediate results obtained after a round of training by clients participating in federated learning or federated meta-learning.
  • the training results or intermediate results can be gradient results, loss function results, performance results of specific tasks or the above results The encryption results, etc.
  • the central device does not require all candidate clients to participate in the model training of specific federated learning or federated meta-learning, but the central device first screens the candidate clients and determines the clients that need model training , and then send a first indication to the client, instructing the client to participate in specific federated learning or federated meta-learning model training, and receive the training result reported by the client.
  • some candidate clients with poor conditions can be eliminated, the convergence speed of training can be improved, and the communication resource overhead between the central device and the client can be reduced.
  • by selecting representative clients for federated training not only the training efficiency can be improved, but also the generalization performance of the model can be improved.
  • the sending of the first indication by the central device to the client includes:
  • the central device selects N clients from M candidate clients according to the preset first screening condition, and unicasts the first indication to the N clients, where M and N are positive integers, and N is less than or equal to M; or
  • the central device broadcasts the first indication to the M candidate clients, the first indication carries a second filter condition, and the second filter condition is used to filter the clients that report the training results, and the client The terminal meets the second screening condition.
  • the client within the communication range of the central device is the candidate client, and the client reporting the training result is selected from the candidate client.
  • All the candidate clients can be used as the client, and some candidate clients can also be selected as the client. end.
  • Broadcasting is to send the first instruction to all candidate clients, while unicast is only sending the first instruction to the selected clients.
  • Candidate clients who receive the unicast first instruction need to perform model training and report the training results.
  • Candidate clients that receive the broadcasted first indication need to judge whether they meet the second screening condition, and only those candidate clients that meet the second screening condition perform model training and report the training results.
  • the central device sends the first indication to the client through at least one of the following:
  • MAC Media access control
  • CE Control Element
  • Radio Resource Control Radio Resource Control, RRC
  • Non-access stratum Non-access stratum (Non-access stratum, NAS) message
  • SIB System Information Block
  • Physical Downlink shared channel Physical Downlink Shared Channel, PDSCH
  • Physical Random Access Channel Physical Random Access Channel (Physical Random Access Channel, PRACH) (Message, MSG) 2 information;
  • PC5 interface (a kind of interface) signaling
  • PSCCH Physical Sidelink Control Channel
  • PSSCH Physical Sidelink Shared Channel
  • PSBCH Physical Sidelink Broadcast Channel
  • PSDCH Physical Sidelink Discovery Channel
  • Physical Sidelink Feedback Channel Physical Sidelink Feedback Channel, PSFCH
  • the method before the central device sends the first indication to the client, the method further includes:
  • the central device receives the first training data and/or the first parameter reported by the candidate client, and the first parameter may be a judgment parameter of the first screening condition.
  • the candidate client may first report a small amount of training data (i.e. the first training data) and/or the first parameter, and the central device determines the client participating in the training according to the small amount of training data and/or the first parameter, Screen out the clients that participate in model training and report the training results to avoid all clients from participating in the training.
  • a small amount of training data i.e. the first training data
  • the central device determines the client participating in the training according to the small amount of training data and/or the first parameter, Screen out the clients that participate in model training and report the training results to avoid all clients from participating in the training.
  • the central device only receives the first training data reported by the candidate client, and determines the first parameter according to the first training data.
  • the central device can speculate, perceive, detect or infer the first parameter according to the first training data.
  • the central device may screen candidate clients according to the first parameter to determine the client.
  • the first parameter includes at least one of the following:
  • the service type of the candidate client such as Enhanced Mobile Broadband (Enhanced Mobile Broadband, eMBB), Ultra-Reliable Low-Latency Communications (Ultra-Reliable Low-Latency Communications, URLLC), Large-scale Machine Type Communication (Massive Machine Type Communication, mMTC) , other 6G new scenarios, etc.;
  • Enhanced Mobile Broadband Enhanced Mobile Broadband, eMBB
  • Ultra-Reliable Low-Latency Communications Ultra-Reliable Low-Latency Communications
  • URLLC Ultra-Reliable Low-Latency Communications
  • mMTC Large-scale Machine Type Communication
  • the working scenarios of the candidate clients include but are not limited to: high speed, low speed, line of sight (Line of Sight, LOS), non line of sight (Non Line of Sight, NLOS), high SNR, low SNR and other work scenes;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes the 2nd generation (2th Generation, 2G), the 3rd generation (3th Generation, 3G), the 4th generation (4th Generation , 4G), 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of graphics processing units (Graphics Processing Unit, GPU), the number of central processing units (Central Processing Unit, CPU), the number of cores, etc., or the computing power can be calculated by the number of calculations per second (Floating -Point Operations Per Second, FLOPS) or processor computing unit capability (one trillion calculations per second (Tera Operations Per Second, TOPS), one billion calculations per second (Giga Operations Per Second, GOPS) and / Or one million calculations per second (Million Operation Per Second, MOPS)) etc. can be expressed;
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the candidate client may first report a small amount of training data (i.e. the first training data) and/or the first parameter to the central device, wherein the first parameter may be a judgment parameter of the first screening condition, and the central device may The first screening condition and/or the second screening condition determine the client that needs to participate in model training and report the training result.
  • the client is selected from candidate clients. Specifically, there may be M candidate clients, from which N are determined The client needs to perform training client screening and reporting, and N can be less than M or equal to M.
  • the clients that need to participate in model training and report training results can be determined according to the data types of the candidate clients, and the candidate clients are grouped according to the data types of the candidate clients, and the data types of the candidate clients in each group are the same or similar .
  • screening clients select K1 candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and K1 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • K1 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the business types of the candidate clients, and the candidate clients are grouped according to the business types of the candidate clients, and the business types of the candidate clients in each group are the same or similar .
  • screening clients select K2 candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and K2 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • K2 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the data distribution parameters of the candidate clients, and the candidate clients are grouped according to the data distribution parameters of the candidate clients, and the data distribution parameters of the candidate clients in each group same or similar.
  • K3 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the working scenarios of the candidate clients, and the candidate clients are grouped according to the working scenarios of the candidate clients, and the working scenarios of the candidate clients in each group are the same or similar .
  • screening clients select A candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and A is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the difficulty of collecting data by the candidate clients, and the candidate clients are prioritized according to the difficulty of collecting data by the candidate clients, and the data collection The less difficult the candidate client, the higher the priority to be screened out.
  • D candidate clients are selected from the candidate clients as clients that need to participate in model training and report training results. D It is a positive integer, which can reduce the difficulty of data collection.
  • the clients that need to participate in the model training and report the training results can be determined according to the willingness of the candidate clients to participate in the model training of a specific federated learning or federated meta-learning, and the candidate clients are prioritized according to the degree of willingness, and the willingness The higher the degree of the candidate client, the higher the priority to be screened out.
  • G candidate clients are selected from the candidate clients as clients that need to participate in model training and report training results.
  • G is a positive integer, which ensures that candidate clients with high willingness to participate in model training.
  • the clients that need to participate in model training and report training results can be determined according to the number of times that candidate clients participate in model training of a specific federated learning or federated meta-learning, and the candidate clients are prioritized according to the number of times they participate in model training , the less the number of candidate clients that have participated in model training, the higher the priority to be screened out, and select K4 candidate clients from the candidate clients according to the priority from high to low as the ones that need to participate in model training and report training
  • K4 is a positive integer, which can balance the number of times that candidate clients participate in model training.
  • the clients that need to participate in model training and report training results can be determined according to the communication network access methods of the candidate clients, and the candidate clients are prioritized according to the communication network access methods of the candidate clients.
  • the methods include fixed network, WiFi and mobile network, and the mobile network includes 2G, 3G, 4G, 5G, 6G, etc.
  • the priority of the fixed network is greater than or equal to the priority of the WiFi
  • the priority of the WiFi is greater than or equal to the priority of the mobile network.
  • the higher the algebra in the mobile network the higher the priority of being screened out. For example, the priority of being screened out for 5G candidate clients is higher than that of 4G candidate clients.
  • the clients that need to participate in model training and report training results can be determined according to the channel quality of the candidate clients, and the candidate clients are prioritized according to the channel quality of the candidate clients, and the candidate clients with higher channel quality are selected
  • the higher the filtered priority select C candidate clients from the candidate clients according to the priority from high to low as the clients that need to participate in model training and report training results.
  • C is a positive integer, which can ensure the channel quality Good clients participate in model training and report training results to ensure the quality of model training.
  • the clients that need to participate in model training and report the training results can be determined according to the battery status of the candidate clients, and the candidate clients are prioritized according to the battery status of the candidate clients.
  • the higher the priority in addition, the candidate clients in the charging state are screened with the highest priority, and E candidate clients are selected from the candidate clients according to the priority from high to low as the ones that need to participate in model training and report training
  • E is a positive integer, which can ensure that the client participating in the model training and reporting the training result has enough power.
  • the clients that need to participate in model training and report training results can be determined according to the storage status of the candidate clients, and the candidate clients are prioritized according to the storage status of the candidate clients.
  • the higher the filtered priority select F candidate clients from the candidate clients according to the priority from high to low as the clients that need to participate in model training and report training results.
  • F is a positive integer, which can ensure that the participating model
  • the client for training and reporting training results has enough available storage space to store training data and training results.
  • the clients that need to participate in model training and report the training results can be determined according to the computing power of the candidate clients, and the candidate clients are prioritized according to the computing power of the candidate clients.
  • P is a positive integer, which can ensure participation in model training And the client that reports the training results has enough computing power for training.
  • the first indication of unicast includes at least one of the following:
  • the above models refer to models of specific federated learning or federated meta-learning.
  • the broadcasted first indication includes at least one of the following:
  • the above models refer to models of specific federated learning or federated meta-learning.
  • the identifiers of the candidate clients that perform client screening and the identifiers of candidate clients that do not perform client screening constitute the second screening condition, and the candidate clients can judge whether they meet the second screening condition according to their own identifiers.
  • the method further includes:
  • the central device sends the converged model and hyperparameters to L reasoning clients, where L is greater than M, equal to M or less than M.
  • the central device judges whether the model converges based on the received training results. If the model does not converge, then repeat the process of screening the client, sending the first instruction to the client, and receiving the training result reported by the client; if the model converges , then the converged model and hyperparameters are sent to L inference clients.
  • the inference clients can be selected from candidate clients, or other clients other than candidate clients.
  • client screening needs to be performed after at least one round of training.
  • the client is triggered to report the first training data and/or the first parameter at least in the round of client screening.
  • the central device sends the converged model and hyperparameters to the inference client, and the inference client performs performance verification and inference on the model.
  • the central device sends the converged model and hyperparameters to the inference client through at least one of the following:
  • Non-access stratum NAS message
  • SIB system information block
  • PSDCH information Physical direct link discovery channel
  • the physical direct link feedback channel PSFCH information The physical direct link feedback channel PSFCH information.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • federated meta-learning obtains an initialization parameter with the best generalization performance through multi-task training, and using this initialization parameter under a new task can quickly converge.
  • the training process is divided into inner iteration and outer iteration. Therefore, among the hyperparameters notified by the central device to the client, there will be hyperparameters not involved in federated learning, such as outer iteration learning rate, inner iteration learning rate, meta-learning rate, number of inner iterations, and number of outer iterations.
  • Federated meta-learning has a greater demand for client-side screening, because the advantage of federated meta-learning is better generalization performance. Therefore, try to be fair to all data when participating in training.
  • hyperparameters sent to different clients can be different.
  • the above-mentioned part of hyperparameters of each client can be determined according to the first parameter corresponding to each client (mainly according to the difficulty of data collection in the first parameter, the power status of the client, the storage status of the client, etc.).
  • Specific principles include at least one of the following:
  • Clients with high data collection difficulty are recommended to use fewer internal iterations and a larger internal iteration step, and clients with less difficult data collection are recommended to use more internal iterations and a smaller internal iteration step;
  • Clients with low battery power are recommended to use fewer number of inner iterations and larger inner iteration step size, and clients with more battery power are recommended to use more inner iteration times and smaller inner iteration step size;
  • Clients with less available storage space are recommended to use fewer number of inner iterations and larger inner iteration step size, and clients with more available storage space are recommended to use more inner iteration times and smaller inner iteration step size;
  • the central device is a network-side device or a terminal; the client is a network-side device or a terminal.
  • the scenario where multiple network-side devices jointly perform federated learning or federated meta-learning and the scenario where multiple terminals jointly perform federated learning or federated meta-learning.
  • the information exchange (including the first parameter, the first indication, etc.) between the central device and the client can be completed through one communication or multiple times of communication.
  • the embodiment of the present application also provides a client screening method, as shown in Figure 5, including:
  • Step 201 the client receives a first instruction from the central device, and the first instruction is used to instruct the client to participate in model training of specific federated learning or federated meta-learning;
  • Step 202 The client performs specific federated learning or federated meta-learning model training, and reports the training result to the central device, and the training result is a result or an intermediate result after the client performs a round of model training.
  • the above training results are the results or intermediate results obtained after a round of training by clients participating in federated learning or federated meta-learning.
  • the training results or intermediate results can be gradient results, loss function results, performance results of specific tasks or the above results The encryption results, etc.
  • the central device does not require all candidate clients to participate in the model training of specific federated learning or federated meta-learning, but the central device first screens the candidate clients and determines the clients that need model training , and then send a first indication to the client, instructing the client to participate in specific federated learning or federated meta-learning model training, and receive the training result reported by the client.
  • some candidate clients with poor conditions can be eliminated, the convergence speed of training can be improved, and the communication resource overhead between the central device and the client can be reduced.
  • by selecting representative clients for federated training not only the training efficiency can be improved, but also the generalization performance of the model can be improved.
  • the client reports the training result to the central device through at least one of the following:
  • Non-access stratum NAS message
  • PSDCH Physical direct link discovery channel
  • the client receiving the first indication from the central device includes:
  • the client receives the first indication unicast by the central device, and the client is a client selected by the central device from candidate clients according to a preset first filtering condition; or
  • the client receives the first instruction broadcast by the central device, the first instruction carries a second filter condition, and the second filter condition is used to filter the client that reports the training result, and the client satisfies The second filter condition.
  • the client performs specific federated learning or federated meta-learning model training, and reports the training results to the central device including:
  • the client receives the first instruction unicast by the central device, the client performs model training and reports a training result; or
  • the client If the client receives the first instruction broadcast by the central device, the client performs model training and reports a training result.
  • the client within the communication range of the central device is the candidate client, and the client reporting the training result is selected from the candidate client.
  • All the candidate clients can be used as the client, and some candidate clients can also be selected as the client. end.
  • Broadcasting is to send the first instruction to all candidate clients, while unicast is only sending the first instruction to the selected clients.
  • Candidate clients who receive the unicast first instruction need to perform model training and report the training results.
  • Candidate clients that receive the broadcasted first indication need to judge whether they meet the second screening condition, and only those candidate clients that meet the second screening condition perform model training and report the training results.
  • the method before the client receives the first instruction from the central device, the method further includes:
  • Candidate clients report first training data and/or first parameters to the central device, the first parameters may be judging parameters of the first screening conditions, and the first training data is used to determine the first parameter.
  • the candidate client may first report a small amount of training data (i.e. the first training data) and/or the first parameter, and the central device determines the client participating in the training according to the small amount of training data and/or the first parameter, Screen out the clients that participate in model training and report the training results to avoid all clients from participating in the training.
  • a small amount of training data i.e. the first training data
  • the central device determines the client participating in the training according to the small amount of training data and/or the first parameter, Screen out the clients that participate in model training and report the training results to avoid all clients from participating in the training.
  • the client reports the first training data and/or the first parameters to the central device through at least one of the following:
  • Non-access stratum NAS message
  • PSDCH Physical direct link discovery channel
  • the candidate client only reports the first training data to the central device, and the first training data is used to determine the first parameter.
  • the first parameter includes at least one of the following:
  • the business type of the candidate client such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), other 6G new scenarios, etc.;
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine type communication
  • the working scenarios of the candidate clients include but not limited to: high-speed, low-speed, line-of-sight propagation LOS, non-line-of-sight propagation NLOS, high signal-to-noise ratio, low signal-to-noise ratio and other working scenarios;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes 2G, 3G, 4G, 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of GPUs, the number of CPUs, the number of cores, etc., or the computing power can be represented by the number of operations per second FLOPS or the capability of the processor operation unit (TOPS, GOPS and/or MOPS);
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the candidate client may first report a small amount of training data (i.e. the first training data) and/or the first parameter to the central device, wherein the first parameter may be a judgment parameter of the first screening condition, and the central device may The first screening condition and/or the second screening condition determine the client that needs to participate in model training and report the training result.
  • the client is selected from candidate clients. Specifically, there may be M candidate clients, from which N are determined The client needs to perform training client screening and reporting, and N can be less than M or equal to M.
  • the clients that need to participate in model training and report training results can be determined according to the data types of the candidate clients, and the candidate clients are grouped according to the data types of the candidate clients, and the data types of the candidate clients in each group are the same or similar .
  • screening clients select K1 candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and K1 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • K1 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the business types of the candidate clients, and the candidate clients are grouped according to the business types of the candidate clients, and the business types of the candidate clients in each group are the same or similar .
  • screening clients select K2 candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and K2 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • K2 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the data distribution parameters of the candidate clients, and the candidate clients are grouped according to the data distribution parameters of the candidate clients, and the data distribution parameters of the candidate clients in each group same or similar.
  • K3 is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the working scenarios of the candidate clients, and the candidate clients are grouped according to the working scenarios of the candidate clients, and the working scenarios of the candidate clients in each group are the same or similar .
  • screening clients select A candidate clients from each group of candidate clients as clients that need to participate in model training and report training results, and A is a positive integer, which can ensure the diversity of clients participating in training and guarantee
  • Each group of candidate clients has clients to participate in model training, and the generalization performance is better, taking into account each group of clients.
  • the clients that need to participate in model training and report training results can be determined according to the difficulty of collecting data by the candidate clients, and the candidate clients are prioritized according to the difficulty of collecting data by the candidate clients, and the data collection The less difficult the candidate client, the higher the priority to be screened out.
  • D candidate clients are selected from the candidate clients as clients that need to participate in model training and report training results. D It is a positive integer, which can reduce the difficulty of data collection.
  • the clients that need to participate in the model training and report the training results can be determined according to the willingness of the candidate clients to participate in the model training of a specific federated learning or federated meta-learning, and the candidate clients are prioritized according to the degree of willingness, and the willingness The higher the degree of the candidate client, the higher the priority to be screened out.
  • G candidate clients are selected from the candidate clients as clients that need to participate in model training and report training results.
  • G is a positive integer, which ensures that candidate clients with high willingness to participate in model training.
  • the clients that need to participate in model training and report training results can be determined according to the number of times that candidate clients participate in model training of a specific federated learning or federated meta-learning, and the candidate clients are prioritized according to the number of times they participate in model training , the less the number of candidate clients that have participated in model training, the higher the priority to be screened out, and select K4 candidate clients from the candidate clients according to the priority from high to low as the ones that need to participate in model training and report training
  • K4 is a positive integer, which can balance the number of times that candidate clients participate in model training.
  • the clients that need to participate in model training and report training results can be determined according to the communication network access methods of the candidate clients, and the candidate clients are prioritized according to the communication network access methods of the candidate clients.
  • the methods include fixed network, WiFi and mobile network, and the mobile network includes 2G, 3G, 4G, 5G, 6G, etc.
  • the priority that the fixed network is selected is greater than or equal to the priority that the WiFi is selected
  • the priority that the WiFi is selected is greater than or equal to the priority that the mobile network is selected.
  • the higher the algebra in the mobile network the higher the priority of being screened out. For example, the priority of being screened out for 5G candidate clients is higher than that of 4G candidate clients.
  • the clients that need to participate in model training and report training results can be determined according to the channel quality of the candidate clients, and the candidate clients are prioritized according to the channel quality of the candidate clients, and the candidate clients with higher channel quality are selected
  • the higher the filtered priority select C candidate clients from the candidate clients according to the priority from high to low as the clients that need to participate in model training and report training results.
  • C is a positive integer, which can ensure the channel quality Good clients participate in model training and report training results to ensure the quality of model training.
  • the clients that need to participate in model training and report the training results can be determined according to the battery status of the candidate clients, and the candidate clients are prioritized according to the battery status of the candidate clients.
  • the higher the priority in addition, the candidate clients in the charging state are screened with the highest priority, and E candidate clients are selected from the candidate clients according to the priority from high to low as the ones that need to participate in model training and report training
  • E is a positive integer, which can ensure that the client participating in the model training and reporting the training result has enough power.
  • the clients that need to participate in model training and report training results can be determined according to the storage status of the candidate clients, and the candidate clients are prioritized according to the storage status of the candidate clients.
  • the higher the filtered priority select F candidate clients from the candidate clients according to the priority from high to low as the clients that need to participate in model training and report training results.
  • F is a positive integer, which can ensure that the participating model
  • the client for training and reporting training results has enough available storage space to store training data and training results.
  • the clients that need to participate in model training and report the training results can be determined according to the computing power of the candidate clients, and the candidate clients are prioritized according to the computing power of the candidate clients.
  • P is a positive integer, which can ensure participation in model training And the client that reports the training results has enough computing power for training.
  • the first indication of unicast includes at least one of the following:
  • the above models refer to models of specific federated learning or federated meta-learning.
  • the broadcasted first indication includes at least one of the following:
  • model initialization parameters The structure of the model; model initialization parameters;
  • the above models refer to models of specific federated learning or federated meta-learning.
  • the identifiers of the candidate clients that perform client screening and the identifiers of candidate clients that do not perform client screening constitute the second screening condition, and the candidate clients can judge whether they meet the second screening condition according to their own identifiers.
  • the method further includes:
  • the inference client receives the converged model and hyperparameters sent by the central device.
  • the central device judges whether the model converges based on the received training results. If the model does not converge, then repeat the process of screening the client, sending the first instruction to the client, and receiving the training result reported by the client; if the model converges , then the converged model and hyperparameters are sent to L inference clients.
  • the inference clients can be selected from candidate clients, or other clients other than candidate clients.
  • client screening needs to be performed after at least one round of training.
  • the client is triggered to report the first training data and/or the first parameter at least in the round of client screening.
  • the central device sends the converged model and hyperparameters to the inference client, and the inference client performs performance verification and inference on the model.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • Learning rate outer iteration learning rate, inner iteration learning rate, meta learning rate, number of iterations, number of inner iterations, number of outer iterations, amount of data required for training, batch size, mini-batch size, regularization parameter, neural network The number of layers, the number of neurons in each hidden layer, the number of learning rounds, the choice of cost function, and the activation function of neurons.
  • federated meta-learning obtains an initialization parameter with the best generalization performance through multi-task training, and using this initialization parameter under a new task can quickly converge.
  • the training process is divided into inner iteration and outer iteration. Therefore, among the hyperparameters notified by the central device to the client, there will be hyperparameters not involved in federated learning, such as outer iteration learning rate, inner iteration learning rate, meta-learning rate, number of inner iterations, and number of outer iterations.
  • Federated meta-learning has a greater demand for client-side screening, because the advantage of federated meta-learning is better generalization performance. Therefore, try to be fair to all data when participating in training.
  • hyperparameters sent to different clients can be different.
  • the above-mentioned part of hyperparameters of each client can be determined according to the first parameter corresponding to each client (mainly according to the difficulty of data collection in the first parameter, the power status of the client, the storage status of the client, etc.).
  • Specific principles include at least one of the following:
  • Clients with low battery power are recommended to use fewer number of inner iterations and larger inner iteration step size, and clients with more battery power are recommended to use more inner iteration times and smaller inner iteration step size;
  • Clients with less available storage space are recommended to use fewer number of inner iterations and larger inner iteration step size, and clients with more available storage space are recommended to use more inner iteration times and smaller inner iteration step size;
  • the method further includes:
  • the reasoning client performs performance verification on the model
  • the reasoning client uses the model for reasoning.
  • the first condition may be configured or pre-configured by the central device or stipulated in a protocol. After the inference client verifies the performance of the model, it may also report the result of whether to perform inference to the central device.
  • the model for performing performance verification is the model delivered by the central device, or the fine-tuned model of the model delivered by the central device.
  • the inference client may directly use the model delivered by the central device to perform performance verification, or perform performance verification after fine-tuning the model delivered by the central device.
  • the special hyperparameters related to meta-learning can be different for each inference client.
  • the special hyperparameters related to meta-learning of each inference client can be determined according to the first parameter corresponding to each inference client (mainly based on the client screening difficulty, battery status, storage status, etc. in the first parameter).
  • the central device is a network-side device or a terminal; the client is a network-side device or a terminal.
  • the scenario where multiple network-side devices jointly perform federated learning or federated meta-learning and the scenario where multiple terminals jointly perform federated learning or federated meta-learning.
  • the information exchange (including the first parameter, the first indication, etc.) between the central device and the client can be completed through one communication or multiple times of communication.
  • the model may be a channel estimation model, a mobility prediction model, and the like.
  • the technical solutions of the embodiments of the present application can be applied to 6G networks, and can also be applied to 5G and 5.5G networks.
  • the client screening method provided in the embodiment of the present application may be executed by a client screening device.
  • the client screening device provided by the embodiment of the present application is described by taking the client screening device executing the client screening method as an example.
  • An embodiment of the present application provides a client screening device, including:
  • a sending module configured to send a first indication to the client, indicating that the client participates in model training of specific federated learning or federated meta-learning
  • the receiving module is configured to receive the training result reported by the client, where the training result is a result or an intermediate result after the client performs a round of model training.
  • the central device does not require all candidate clients to participate in the model training of specific federated learning or federated meta-learning, but the central device first screens the candidate clients and determines the clients that need model training , and then send a first indication to the client, instructing the client to participate in specific federated learning or federated meta-learning model training, and receive the training result reported by the client.
  • some candidate clients with poor conditions can be eliminated, the convergence speed of training can be improved, and the communication resource overhead between the central device and the client can be reduced.
  • by selecting representative clients for federated training not only the training efficiency can be improved, but also the generalization performance of the model can be improved.
  • the sending module is specifically configured to select N clients from M candidate clients according to a preset first screening condition, and unicast the first indication to the N clients, M , N is a positive integer, N is less than or equal to M; or
  • the receiving module is further configured to receive first training data and/or a first parameter reported by the candidate client, and the first parameter may be a judgment parameter of the first screening condition.
  • the receiving module is configured to only receive first training data reported by the candidate client, and determine the first parameter according to the first training data.
  • the first parameter includes at least one of the following:
  • the business type of the candidate client such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), other 6G new scenarios, etc.;
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine type communication
  • the working scenarios of the candidate clients include but not limited to: high-speed, low-speed, line-of-sight propagation LOS, non-line-of-sight propagation NLOS, high signal-to-noise ratio, low signal-to-noise ratio and other working scenarios;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes 2G, 3G, 4G, 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of GPUs (graphics processing units), the number of CPUs (central processing units), the number of cores, etc., or the computing power can use the number of operations per second FLOPS or the computing unit capability of the processor (TOPS, GOPS and/or MOPS) etc.;
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the first indication of unicast includes at least one of the following:
  • the broadcasted first indication includes at least one of the following:
  • the sending module is further configured to judge that the model has reached convergence according to the training result, and send the converged model and hyperparameters to L reasoning clients, where L is greater than M, equal to M or less than M.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • Learning rate outer iteration learning rate, inner iteration learning rate, meta learning rate, number of iterations, number of inner iterations, number of outer iterations, amount of data required for training, batch size, mini-batch size, regularization parameter, neural network The number of layers, the number of neurons in each hidden layer, the number of learning rounds, the choice of cost function, and the activation function of neurons.
  • the central device is a network side device or a terminal
  • the client is a network side device or a terminal.
  • An embodiment of the present application provides a client screening device, including:
  • a receiving module configured to receive a first instruction from the central device, where the first instruction is used to instruct the client to participate in model training of specific federated learning or federated meta-learning;
  • the reporting module is used to perform model training of specific federated learning or federated meta-learning, and report the training result to the central device, and the training result is the result or an intermediate result after the client performs a round of model training.
  • the receiving module is configured to receive the first indication unicast by the central device, and the client is selected from candidate clients by the central device according to a preset first filtering condition client; or
  • the central device receiving the first instruction broadcast by the central device, the first instruction carrying a second filter condition, the second filter condition is used to filter the client that reports the training result, and the client satisfies the second filter condition. filter criteria.
  • the reporting module is configured to perform model training and report a training result if the client receives the first indication unicast by the central device; or
  • the client receives the first instruction broadcast by the central device, it performs model training and reports a training result.
  • the reporting module is further configured to report first training data and/or first parameters to the central device, the first parameters may be judgment parameters of the first screening conditions, and the first Training data is used to determine the first parameter.
  • the first parameter includes at least one of the following:
  • the business type of the candidate client such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), other 6G new scenarios, etc.;
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine type communication
  • the working scenarios of the candidate clients include but not limited to: high-speed, low-speed, line-of-sight propagation LOS, non-line-of-sight propagation NLOS, high signal-to-noise ratio, low signal-to-noise ratio and other working scenarios;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes 2G, 3G, 4G, 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of GPUs, the number of CPUs, the number of cores, etc., or the computing power can be represented by the number of operations per second FLOPS or the capability of the processor operation unit (TOPS, GOPS and/or MOPS);
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the first indication of unicast includes at least one of the following:
  • the broadcasted first indication includes at least one of the following:
  • model initialization parameters The structure of the model; model initialization parameters;
  • the receiving module is further configured to receive the converged model and hyperparameters sent by the central device.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • Learning rate outer iteration learning rate, inner iteration learning rate, meta learning rate, number of iterations, number of inner iterations, number of outer iterations, amount of data required for training, batch size, mini-batch size, regularization parameter, neural network The number of layers, the number of neurons in each hidden layer, the number of learning rounds, the choice of cost function, and the activation function of neurons.
  • the first part of the hyperparameter is determined by the first parameter corresponding to the inference client, and the first part includes at least one of the following:
  • Outer iteration learning rate inner iteration learning rate, meta learning rate, inner iteration count, outer iteration count.
  • the device also includes:
  • the processing module is used to verify the performance of the model; if the performance verification result meets the preset first condition, the model is used for reasoning.
  • the model for performance verification is the model delivered by the central device, or the model delivered by the central device is fine-tuned.
  • the central device is a network side device or a terminal
  • the client is a network side device or a terminal.
  • the client screening apparatus in this embodiment of the present application may be an electronic device, such as an electronic device with an operating system, or a component in the electronic device, such as an integrated circuit or a chip.
  • the electronic device may be a terminal, or other devices other than the terminal.
  • the terminal may include, but not limited to, the types of terminal 11 listed above, and other devices may be servers, Network Attached Storage (NAS), etc., which are not specifically limited in this embodiment of the present application.
  • NAS Network Attached Storage
  • the client screening device provided in the embodiment of the present application can realize the various processes realized by the method embodiments in FIG. 4 to FIG. 5 and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • this embodiment of the present application also provides a communication device 600, including a processor 601 and a memory 602, and the memory 602 stores programs or instructions that can run on the processor 601, such as
  • the communication device 600 is the central device, when the program or instruction is executed by the processor 601, each step in the above embodiment of the client screening method can be realized, and the same technical effect can be achieved.
  • the communication device 600 is a client, when the program or instruction is executed by the processor 601, each step of the above client screening method embodiment can be achieved, and the same technical effect can be achieved. To avoid repetition, details are not repeated here.
  • the embodiment of the present application also provides a central device, the central device includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and when the programs or instructions are executed by the processor Steps to implement the client screening method as described above.
  • the embodiment of the present application also provides a central device, including a processor and a communication interface, wherein the communication interface is used to send a first indication to the client, instructing the client to participate in a model of specific federated learning or federated meta-learning Training: receiving a training result reported by the client, where the training result is a result or an intermediate result after the client performs a round of model training.
  • the embodiment of the present application also provides a client, the client includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and when the programs or instructions are executed by the processor Steps to implement the client screening method as described above.
  • the embodiment of the present application also provides a client, including a processor and a communication interface, wherein the communication interface is used to receive a first indication from the central device, and the first indication is used to instruct the client to participate in a specific federation Model training of learning or federated meta-learning; the processor is used for model training of specific federated learning or federated meta-learning, and reports the training result to the central device, and the training result is a round of model training performed by the client subsequent or intermediate results.
  • a client including a processor and a communication interface, wherein the communication interface is used to receive a first indication from the central device, and the first indication is used to instruct the client to participate in a specific federation Model training of learning or federated meta-learning; the processor is used for model training of specific federated learning or federated meta-learning, and reports the training result to the central device, and the training result is a round of model training performed by the client subsequent or intermediate results.
  • the aforementioned central device may be a network-side device or a terminal
  • the client may be a network-side device or a terminal.
  • FIG. 7 is a schematic diagram of a hardware structure of a terminal implementing an embodiment of the present application.
  • the terminal 700 includes, but is not limited to: a radio frequency unit 701, a network module 702, an audio output unit 703, an input unit 704, a sensor 705, a display unit 706, a user input unit 707, an interface unit 708, a memory 709, and a processor 710. At least some parts.
  • the terminal 700 can also include a power supply (such as a battery) for supplying power to various components, and the power supply can be logically connected to the processor 710 through the power management system, so that the management of charging, discharging, and functions can be realized through the power management system. Consumption management and other functions.
  • a power supply such as a battery
  • the terminal structure shown in FIG. 7 does not constitute a limitation on the terminal, and the terminal may include more or fewer components than shown in the figure, or combine some components, or arrange different components, which will not be repeated here.
  • the input unit 704 may include a graphics processing unit (Graphics Processing Unit, GPU) 7041 and a microphone 7042, and the graphics processor 7041 is used by the image capture device (such as the image data of the still picture or video obtained by the camera) for processing.
  • the display unit 706 may include a display panel 7061, and the display panel 7061 may be configured in the form of a liquid crystal display, an organic light emitting diode, or the like.
  • the user input unit 707 includes at least one of a touch panel 7071 and other input devices 7072 .
  • the touch panel 7071 is also called a touch screen.
  • the touch panel 7071 may include two parts, a touch detection device and a touch controller.
  • Other input devices 7072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, switch buttons, etc.), trackballs, mice, and joysticks, which will not be described in detail here.
  • the radio frequency unit 701 may transmit the downlink data from the network side device to the processor 710 for processing after receiving the downlink data; in addition, the radio frequency unit 701 may send uplink data to the network side device.
  • the radio frequency unit 701 includes, but is not limited to, an antenna, an amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, and the like.
  • the memory 709 can be used to store software programs or instructions as well as various data.
  • the memory 709 may mainly include a first storage area for storing programs or instructions and a second storage area for storing data, wherein the first storage area may store an operating system, an application program or instructions required by at least one function (such as a sound playing function, image playback function, etc.), etc.
  • memory 709 may include volatile memory or nonvolatile memory, or, memory 709 may include both volatile and nonvolatile memory.
  • the non-volatile memory can be read-only memory (Read-Only Memory, ROM), programmable read-only memory (Programmable ROM, PROM), erasable programmable read-only memory (Erasable PROM, EPROM), electronically programmable Erase Programmable Read-Only Memory (Electrically EPROM, EEPROM) or Flash.
  • ROM Read-Only Memory
  • PROM programmable read-only memory
  • Erasable PROM Erasable PROM
  • EPROM erasable programmable read-only memory
  • Electrical EPROM Electrical EPROM
  • EEPROM electronically programmable Erase Programmable Read-Only Memory
  • Volatile memory can be random access memory (Random Access Memory, RAM), static random access memory (Static RAM, SRAM), dynamic random access memory (Dynamic RAM, DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDRSDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous connection dynamic random access memory (Synch link DRAM , SLDRAM) and Direct Memory Bus Random Access Memory (Direct Rambus RAM, DRRAM).
  • RAM Random Access Memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • DRAM synchronous dynamic random access memory
  • SDRAM double data rate synchronous dynamic random access memory
  • Double Data Rate SDRAM Double Data Rate SDRAM
  • DDRSDRAM double data rate synchronous dynamic random access memory
  • Enhanced SDRAM, ESDRAM enhanced synchronous dynamic random access memory
  • Synch link DRAM , SLDRAM
  • Direct Memory Bus Random Access Memory Direct Rambus
  • the processor 710 may include one or more processing units; optionally, the processor 710 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to the operating system, user interface, and application programs, etc., Modem processors mainly process wireless communication signals, such as baseband processors. It can be understood that the foregoing modem processor may not be integrated into the processor 710 .
  • the central device is a terminal
  • the processor 710 is configured to send a first indication to the client, instructing the client to participate in model training of specific federated learning or federated meta-learning; receiving the training result reported by the client,
  • the training result is a result or an intermediate result after the client performs a round of model training.
  • the processor 710 is configured to select N clients from M candidate clients according to a preset first screening condition, and unicast the first indication to the N clients, M, N is a positive integer and N is less than or equal to M; or
  • the processor 710 is configured to receive first training data and/or a first parameter reported by the candidate client, and the first parameter may be a judgment parameter of the first screening condition.
  • the processor 710 is configured to only receive first training data reported by the candidate client, and determine the first parameter according to the first training data.
  • the first parameter includes at least one of the following:
  • the business type of the candidate client such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), other 6G new scenarios, etc.;
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine type communication
  • the working scenarios of the candidate clients include but not limited to: high-speed, low-speed, line-of-sight propagation LOS, non-line-of-sight propagation NLOS, high signal-to-noise ratio, low signal-to-noise ratio and other working scenarios;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes 2G, 3G, 4G, 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of GPUs, the number of CPUs, the number of cores, etc., or the computing power can be represented by the number of operations per second FLOPS or the capability of the processor operation unit (TOPS, GOPS and/or MOPS);
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the first indication of unicast includes at least one of the following:
  • the broadcasted first indication includes at least one of the following:
  • the processor 710 is configured to judge that the model has reached convergence according to the training result, and send the converged model and hyperparameters to L reasoning clients, where L is greater than M, equal to M or less than M.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • Learning rate outer iteration learning rate, inner iteration learning rate, meta learning rate, number of iterations, number of inner iterations, number of outer iterations, amount of data required for training, batch size, mini-batch size, regularization parameter, neural network The number of layers, the number of neurons in each hidden layer, the number of learning rounds, the choice of cost function, and the activation function of neurons.
  • the central device is a network side device or a terminal
  • the client is a network side device or a terminal.
  • the client is a terminal
  • the processor 710 is configured to receive a first instruction from the central device, the first instruction is used to instruct the client to participate in model training of specific federated learning or federated meta-learning; perform specific federated learning Learning or federated meta-learning model training, and reporting the training result to the central device, the training result is the result or intermediate result after the client performs a round of model training.
  • the processor 710 is configured to receive the first indication unicast by the central device, and the client is a client selected from candidate clients by the central device according to a preset first screening condition. end; or
  • the central device receiving the first instruction broadcast by the central device, the first instruction carrying a second filter condition, the second filter condition is used to filter the client that reports the training result, and the client satisfies the second filter condition. filter criteria.
  • the processor 710 is configured to perform model training and report a training result if the client receives the first indication unicast by the central device; or
  • the client receives the first instruction broadcast by the central device, it performs model training and reports a training result.
  • the processor 710 is configured to report the first training data and/or the first parameter to the central device, the first parameter may be a judgment parameter of the first screening condition, and the first training data used to determine the first parameter.
  • the first parameter includes at least one of the following:
  • the business type of the candidate client such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), other 6G new scenarios, etc.;
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine type communication
  • the working scenarios of the candidate clients include but not limited to: high-speed, low-speed, line-of-sight propagation LOS, non-line-of-sight propagation NLOS, high signal-to-noise ratio, low signal-to-noise ratio and other working scenarios;
  • the communication network access mode of the candidate client includes mobile network, WiFi and fixed network, wherein the mobile network includes 2G, 3G, 4G, 5G and 6G;
  • the channel quality of the candidate client is the channel quality of the candidate client
  • the power state of the candidate client such as the specific value of the available remaining power, or the result of the classification description, charging or not charging, etc.
  • the storage status of the candidate client such as a specific value of available memory, or a hierarchical description result
  • the computing power of the candidate client such as the number of GPUs, the number of CPUs, the number of cores, etc., or the computing power can be represented by the number of operations per second FLOPS or the capability of the processor operation unit (TOPS, GOPS and/or MOPS);
  • the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning is the degree of willingness of the candidate client to participate in model training of a specific federated learning or federated meta-learning.
  • the first indication of unicast includes at least one of the following:
  • the broadcasted first indication includes at least one of the following:
  • model initialization parameters The structure of the model; model initialization parameters;
  • the processor 710 is configured to receive the converged model and hyperparameters sent by the central device.
  • the model is a federated meta-learning model, and the hyperparameters are determined by the first parameters.
  • the hyperparameters include at least one of the following:
  • Learning rate outer iteration learning rate, inner iteration learning rate, meta learning rate, number of iterations, number of inner iterations, number of outer iterations, amount of data required for training, batch size, mini-batch size, regularization parameter, neural network The number of layers, the number of neurons in each hidden layer, the number of learning rounds, the choice of cost function, and the activation function of neurons.
  • the first part of the hyperparameter is determined by the first parameter corresponding to the inference client, and the first part includes at least one of the following:
  • Outer iteration learning rate inner iteration learning rate, meta learning rate, inner iteration count, outer iteration count.
  • the processor 710 is configured to perform performance verification on the model; if the performance verification result satisfies a preset first condition, the model is used for reasoning.
  • the model for performance verification is the model delivered by the central device, or the model delivered by the central device is fine-tuned.
  • the central device is a network side device or a terminal
  • the client is a network side device or a terminal.
  • the embodiment of the present application further provides a network-side device, including a processor and a communication interface.
  • the network-side device embodiment corresponds to the above-mentioned network-side device method embodiment, and each implementation process and implementation mode of the above-mentioned method embodiment can be applied to this network-side device embodiment, and can achieve the same technical effect.
  • the embodiment of the present application also provides a network side device.
  • the network side device 800 includes: an antenna 81 , a radio frequency device 82 , a baseband device 83 , a processor 84 and a memory 85 .
  • the antenna 81 is connected to a radio frequency device 82 .
  • the radio frequency device 82 receives information through the antenna 81, and sends the received information to the baseband device 83 for processing.
  • the baseband device 83 processes the information to be sent and sends it to the radio frequency device 82
  • the radio frequency device 82 processes the received information and sends it out through the antenna 81 .
  • the method performed by the network side device in the above embodiments may be implemented in the baseband device 83, where the baseband device 83 includes a baseband processor.
  • the baseband device 83 can include at least one baseband board, for example, a plurality of chips are arranged on the baseband board, as shown in FIG.
  • the program executes the network device operations shown in the above method embodiments.
  • the network side device may also include a network interface 86, such as a common public radio interface (common public radio interface, CPRI).
  • a network interface 86 such as a common public radio interface (common public radio interface, CPRI).
  • the network side device 800 in the embodiment of the present invention also includes: instructions or programs stored in the memory 85 and operable on the processor 84, and the processor 84 invokes the instructions or programs in the memory 85 to execute the above-mentioned client terminal screening method, and achieve the same technical effect, in order to avoid repetition, it is not repeated here.
  • the embodiment of the present application also provides a readable storage medium, the readable storage medium stores a program or an instruction, and when the program or instruction is executed by the processor, each process of the above client screening method embodiment is realized, and can achieve The same technical effects are not repeated here to avoid repetition.
  • the processor is the processor in the terminal described in the foregoing embodiments.
  • the readable storage medium includes a computer-readable storage medium, such as a computer read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk, and the like.
  • the embodiment of the present application further provides a chip, the chip includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the above-mentioned embodiment of the client screening method Each process, and can achieve the same technical effect, in order to avoid repetition, will not repeat them here.
  • the chip mentioned in the embodiment of the present application may also be called a system-on-chip, a system-on-chip, a system-on-a-chip, or a system-on-a-chip.
  • the embodiment of the present application further provides a computer program/program product, the computer program/program product is stored in a storage medium, and the computer program/program product is executed by at least one processor to implement the above client screening method
  • the embodiment of the present application also provides a client screening system, including: a central device and a client, the central device can be used to perform the steps of the client screening method as described above, and the client can be used to perform the above steps The steps of the client screening method.
  • the term “comprising”, “comprising” or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or device. Without further limitations, an element defined by the phrase “comprising a " does not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising that element.
  • the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved. Functions are performed, for example, the described methods may be performed in an order different from that described, and various steps may also be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
  • the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation.
  • the technical solution of the present application can be embodied in the form of computer software products, which are stored in a storage medium (such as ROM/RAM, magnetic disk, etc.) , CD-ROM), including several instructions to make a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Algebra (AREA)
  • Computational Mathematics (AREA)
  • Medical Informatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

一种客户端筛选方法及装置、客户端及中心设备，属于通信技术领域，该客户端筛选方法包括：中心设备向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练（101）；所述中心设备接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果（102）。

Description

客户端筛选方法及装置、客户端及中心设备
相关申请的交叉引用
本申请主张在2021年12月15日在中国提交的中国专利申请No.202111537989.6的优先权，其全部内容通过引用包含于此。
技术领域
本申请属于通信技术领域，具体涉及一种客户端筛选方法及装置、客户端及中心设备。
背景技术
人工智能(Artificial Intelligence，AI)目前在各个领域获得了广泛的应用。AI模块有多种实现方式，例如神经网络、决策树、支持向量机、贝叶斯分类器等。
AI应用于无线通信系统中时，需要在终端上运行相应的神经网络。但是，终端的AI模型一般是由网络侧离线训练好，再下发到终端执行的。这是因为，单个终端的数据量有限，很难训练出一个较好的模型。终端接收到模型后一般先基于少量数据进行微调(fine-tuning)，才能获得比较好的性能。联邦学习或联邦元学习可以在不暴露终端数据的前提下进行训练，是非常有潜力的方向。在联邦学习和联邦元学习中都是终端用自己的数据更新本地模型参数或损失，终端再将参数或损失汇聚到服务器进行处理获得全局模型，服务器再将全局模型下发给终端进行新一轮的训练。联邦学习的目的是获得一个对所有参与训练的终端都可以收敛的模型，联邦元学习的目的是基于参与训练的终端数据，获得一个在新的场景中也能快速收敛的模型初始化参数。联邦元学习在终端处微调数据量少或对微调收敛时间(或迭代次数)要求较高时可以获得比联邦学习更好的性能。相关技术中，所有的候选客户端都会参与特定联邦学习或联邦元学习的模型训练，导致模型训练的收敛速度慢，中心设备与客户端之间的通信资源开销大。
发明内容
本申请实施例提供一种客户端筛选方法及装置、客户端及中心设备，能够提高模型训练的收敛速度。
第一方面，提供了一种客户端筛选方法，包括：
中心设备向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
所述中心设备接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第二方面，提供了一种客户端筛选装置，包括：
发送模块，用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
接收模块，用于接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第三方面，提供了一种客户端筛选方法，包括：
客户端接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
所述客户端进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第四方面，提供了一种客户端筛选装置，包括：
接收模块，用于接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
上报模块，用于进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第五方面，提供了一种中心设备，该中心设备包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如第一方面所述的方法的步骤。
第六方面，提供了一种中心设备，包括处理器及通信接口，其中，所述 通信接口用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第七方面，提供了一种客户端，该客户端包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如第三方面所述的方法的步骤。
第八方面，提供了一种客户端，包括处理器及通信接口，其中，所述通信接口用于接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；所述处理器用于进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
第九方面，提供了一种客户端筛选系统，包括：中心设备及客户端，所述中心设备可用于执行如第一方面所述的客户端筛选方法的步骤，所述客户端可用于执行如第三方面所述的客户端筛选方法的步骤。
第十方面，提供了一种可读存储介质，所述可读存储介质上存储程序或指令，所述程序或指令被处理器执行时实现如第一方面所述的方法的步骤，或者实现如第三方面所述的方法的步骤。
第十一方面，提供了一种芯片，所述芯片包括处理器和通信接口，所述通信接口和所述处理器耦合，所述处理器用于运行程序或指令，实现如第一方面所述的方法，或实现如第三方面所述的方法。
第十二方面，提供了一种计算机程序/程序产品，所述计算机程序/程序产品被存储在存储介质中，所述计算机程序/程序产品被至少一个处理器执行以实现如第一方面所述的客户端筛选方法，或实现如第三方面所述的客户端筛选方法的步骤。
在本申请实施例中，中心设备并不是要求所有的候选客户端参与特定联邦学习或联邦元学习的模型训练，而是由中心设备先对候选客户端进行筛选，确定需要进行模型训练的客户端，然后向客户端发送第一指示，指示客户端参与特定联邦学习或联邦元学习的模型训练，并接收客户端上报的训练结果。这样可以剔除一些条件不好的候选客户端，提高训练的收敛速度，降低中心 设备与客户端之间的通信资源开销。
附图说明
图1是本申请实施例可应用的一种无线通信系统的框图；
图2是信道状态信息反馈的示意图；
图3是AI训练不同迭代次数时的性能示意图；
图4是本申请实施例中心设备侧客户端筛选方法的流程示意图；
图5是本申请实施例客户端侧客户端筛选方法的流程示意图；
图6是本申请实施例通信设备的结构示意图；
图7是本申请实施例终端的结构示意图；
图8是本申请实施例网络侧设备的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚描述，显然，所描述的实施例是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员所获得的所有其他实施例，都属于本申请保护的范围。
本申请的说明书和权利要求书中的术语“第一”、“第二”等是用于区别类似的对象，而不用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换，以便本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施，且“第一”、“第二”所区别的对象通常为一类，并不限定对象的个数，例如第一对象可以是一个，也可以是多个。此外，说明书以及权利要求中“和/或”表示所连接对象的至少其中之一，字符“/”一般表示前后关联对象是一种“或”的关系。
值得指出的是，本申请实施例所描述的技术不限于长期演进型(Long Term Evolution，LTE)/LTE的演进(LTE-Advanced，LTE-A)系统，还可用于其他无线通信系统，诸如码分多址(Code Division Multiple Access，CDMA)、时分多址(Time Division Multiple Access，TDMA)、频分多址(Frequency Division Multiple Access，FDMA)、正交频分多址(Orthogonal Frequency  Division Multiple Access，OFDMA)、单载波频分多址(Single-carrier Frequency Division Multiple Access，SC-FDMA)和其他系统。本申请实施例中的术语“系统”和“网络”常被可互换地使用，所描述的技术既可用于以上提及的系统和无线电技术，也可用于其他系统和无线电技术。以下描述出于示例目的描述了新空口(New Radio，NR)系统，并且在以下大部分描述中使用NR术语，但是这些技术也可应用于NR系统应用以外的应用，如第6代(6th Generation，6G)通信系统。
图1示出本申请实施例可应用的一种无线通信系统的框图。无线通信系统包括终端11和网络侧设备12。其中，终端11可以是手机、平板电脑(Tablet Personal Computer)、膝上型电脑(Laptop Computer)或称为笔记本电脑、个人数字助理(Personal Digital Assistant，PDA)、掌上电脑、上网本、超级移动个人计算机(ultra-mobile personal computer，UMPC)、移动上网装置(Mobile Internet Device，MID)、增强现实(augmented reality，AR)/虚拟现实(virtual reality，VR)设备、机器人、可穿戴式设备(Wearable Device)、车载设备(Vehicle User Equipment，VUE)、行人终端(Pedestrian User Equipment，PUE)、智能家居(具有无线通信功能的家居设备，如冰箱、电视、洗衣机或者家具等)、游戏机、个人计算机(personal computer，PC)、柜员机或者自助机等终端侧设备，可穿戴式设备包括：智能手表、智能手环、智能耳机、智能眼镜、智能首饰(智能手镯、智能手链、智能戒指、智能项链、智能脚镯、智能脚链等)、智能腕带、智能服装等。需要说明的是，在本申请实施例并不限定终端11的具体类型。网络侧设备12可以包括接入网设备或核心网设备，其中，接入网设备12也可以称为无线接入网设备、无线接入网(Radio Access Network,RAN)、无线接入网功能或无线接入网单元。接入网设备12可以包括基站、WLAN接入点或WiFi节点等，基站可被称为节点B、演进节点B(eNB)、接入点、基收发机站(Base Transceiver Station，BTS)、无线电基站、无线电收发机、基本服务集(Basic Service Set，BSS)、扩展服务集(Extended Service Set，ESS)、家用B节点、家用演进型B节点、发送接收点(Transmitting Receiving Point，TRP)或所述领域中其他某个合适的术语，只要达到相同的技术效果，所述基站不限于特定技术词汇， 需要说明的是，在本申请实施例中仅以NR系统中的基站为例进行介绍，并不限定基站的具体类型。
借助AI提升第5代(5th Generation，5G)网络性能的主要方法是通过基于神经网络的算法和模型增强或者替代目前已有的算法或处理模块。在特定场景下，基于神经网络的算法和模型可以取得比基于确定性算法更好的性能。比较常用的神经网络包括深度神经网络、卷积神经网络和循环神经网络等。借助已有AI工具，可以实现神经网络的搭建、训练与验证工作。
通过AI方法替代现有系统中的模块能够有效提升系统性能。如图2所示的信道状态信息(Channel State Information，CSI)反馈，通过AI编码器(encoder)和AI解码器(decoder)替代常规的CSI计算，可以在相同开销的情况下大幅度提升相应的系统性能。通过基于AI的方案，系统的频谱效率可以提升30％左右。
AI训练不同迭代次数时的性能如图3所示，其中，横坐标为训练时期，纵坐标为相关性的平方。不同迭代需要不同的训练数据，可以看到需要大量的训练迭代才能达到性能收敛。
选取具有代表性的终端进行联邦式的训练不仅可以提高训练效率，还可以提高模型的泛化性能。而由于联邦学习和联邦元学习的训练目的不同，这两种训练方法所对应的终端筛选方案也应该不同。
下面结合附图，通过一些实施例及其应用场景对本申请实施例提供的客户端筛选方法进行详细地说明。
本申请实施例提供一种客户端筛选方法，如图4所示，包括：
步骤101：中心设备向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
步骤102：所述中心设备接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
其中，上述训练结果是参与联邦学习或联邦元学习的客户端进行一轮训练后得到的结果或中间结果，训练结果或中间结果可以是梯度结果、损失函数结果、特定任务的性能结果或上述结果的加密结果等。
在本申请实施例中，中心设备并不是要求所有的候选客户端参与特定联 邦学习或联邦元学习的模型训练，而是由中心设备先对候选客户端进行筛选，确定需要进行模型训练的客户端，然后向客户端发送第一指示，指示客户端参与特定联邦学习或联邦元学习的模型训练，并接收客户端上报的训练结果。这样可以剔除一些条件不好的候选客户端，提高训练的收敛速度，降低中心设备与客户端之间的通信资源开销。并且，通过选取具有代表性的客户端进行联邦式的训练不仅可以提高训练效率，还可以提高模型的泛化性能。
一些实施例中，所述中心设备向所述客户端发送第一指示包括：
所述中心设备按照预设的第一筛选条件从M个候选客户端中筛选出N个客户端，向所述N个客户端单播所述第一指示，M，N为正整数，N小于或等于M；或
所述中心设备向所述M个候选客户端广播所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
本实施例中，在中心设备通信范围内的为候选客户端，上报训练结果的客户端选自候选客户端，可以将所有的候选客户端作为客户端，也可以筛选出部分候选客户端作为客户端。广播是向所有的候选客户端发送第一指示，而单播只是向筛选出的客户端发送第一指示，收到单播的第一指示的候选客户端均需要执行模型训练并上报训练结果。收到广播的第一指示的候选客户端需要判断自身是否满足第二筛选条件，满足第二筛选条件的候选客户端才执行模型训练并上报训练结果。
一些实施例中，所述中心设备通过以下至少一项向所述客户端发送第一指示：
媒体介入控制(Medium Access Control，MAC)控制单元(Control Element，CE)；
无线资源控制(Radio Resource Control，RRC)消息；
非接入层(Non-access stratum，NAS)消息；
管理编排消息；
用户面数据；
下行控制信息；
系统信息块(System Information Block，SIB)；
物理下行控制信道(Physical Downlink Control Channel，PDCCH)的层1信令；
物理下行共享信道(Physical Downlink Shared Channel，PDSCH)的信息；
物理随机接入信道(Physical Random Access Channel，PRACH)的(Message，MSG)2信息；
物理随机接入信道PRACH的MSG 4信息；
物理随机接入信道PRACH的MSG B信息；
广播信道信息或信令；
Xn接口(一种接口)信令；
PC5接口(一种接口)信令；
物理侧边链路控制信道(Physical Sidelink Control Channel，PSCCH)的信息或信令；
物理侧边链路共享信道(Physical Sidelink Shared Channel，PSSCH)的信息；
物理侧边链路广播信道(Physical Sidelink Broadcast Channel，PSBCH)的信息；
物理直通链路发现信道(Physical Sidelink Discovery Channel，PSDCH)的信息；
物理直通链路反馈信道(Physical Sidelink Feedback Channel，PSFCH)的信息。
一些实施例中，所述中心设备向客户端发送第一指示之前，所述方法还包括：
所述中心设备接收所述候选客户端上报的第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数。
本实施例中，候选客户端可以先上报少量的训练数据(即第一训练数据)和/或第一参数，由中心设备根据少量的训练数据和/或第一参数确定参与训练的客户端，筛选出参与模型训练并上报训练结果的客户端，避免所有的客 户端都参与训练。
一些实施例中，所述中心设备仅接收所述候选客户端上报的第一训练数据，根据所述第一训练数据确定所述第一参数。中心设备可以根据第一训练数据推测、感知、检测或推理出第一参数。中心设备可以依据第一参数进行候选客户端的筛选，确定客户端。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(Enhanced Mobile Broadband，eMBB)，超可靠低延迟通信(Ultra-Reliable Low-Latency Communications，URLLC)，大规模机器类型通信(Massive Machine Type Communication，mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播(Line of Sight，LOS)、非视距传播(Non Line of Sight，NLOS)、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括第2代(2th Generation，2G)，第3代(3th Generation，3G)，第4代(4th Generation，4G)，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如图形处理器(Graphics Processing Unit，GPU)数、中央处理器(Central Processing Unit，CPU)数、核数等，或者，算力可以用每秒的运算次数(Floating-Point Operations Per Second，FLOPS)或处理器运算单元能力(每秒可进行一万亿次计算(Tera Operations Per Second，TOPS)、每秒可进行十亿次计算(Giga Operations Per Second，GOPS)和/或每秒可进行一百万次计算(Million Operation Per Second，MOPS)) 等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
本实施例中，候选客户端可以先向中心设备上报少量的训练数据(即第一训练数据)和/或第一参数，其中，第一参数可以是第一筛选条件的判断参数，中心设备根据第一筛选条件和/或第二筛选条件确定需要参与模型训练和上报训练结果的客户端，该客户端是选自候选客户端，具体地，可以有M个候选客户端，从其中确定N个客户端需要进行训练客户端筛选和上报，N可以小于M，也可以等于M。
一具体示例中，可以根据候选客户端的数据类型确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的数据类型对候选客户端进行分组，每组内的候选客户端的数据类型相同或相近。在筛选客户端时，从每一组候选客户端中选取K1个候选客户端作为需要参与模型训练和上报训练结果的客户端，K1为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端的业务类型确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的业务类型对候选客户端进行分组，每组内的候选客户端的业务类型相同或相近。在筛选客户端时，从每一组候选客户端中选取K2个候选客户端作为需要参与模型训练和上报训练结果的客户端，K2为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端的数据分布参数确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的数据分布参数对候选客户端进行分组，每组内的候选客户端的数据分布参数相同或相近。在筛选客户端时，从每一组候选客户端中选取K3个候选客户端作为需要参与模型训练和上报训练结果的客户端，K3为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾 及每一组的客户端。
一具体示例中，可以根据候选客户端的工作场景确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的工作场景对候选客户端进行分组，每组内的候选客户端的工作场景相同或相近。在筛选客户端时，从每一组候选客户端中选取A个候选客户端作为需要参与模型训练和上报训练结果的客户端，A为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端收集数据的难易程度确定需要参与模型训练和上报训练结果的客户端，按照候选客户端收集数据的难易程度对候选客户端进行优先级排序，收集数据难度越小的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取D个候选客户端作为需要参与模型训练和上报训练结果的客户端，D为正整数，这样可以降低数据采集的难度。
一具体示例中，可以根据候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度确定需要参与模型训练和上报训练结果的客户端，按照意愿程度对候选客户端进行优先级排序，意愿程度越高的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取G个候选客户端作为需要参与模型训练和上报训练结果的客户端，G为正整数，这样可以保证意愿程度高的候选客户端参与模型训练。
一具体示例中，可以根据候选客户端参与特定联邦学习或联邦元学习的模型训练的次数确定需要参与模型训练和上报训练结果的客户端，按照参与模型训练的次数对候选客户端进行优先级排序，已参与模型训练的次数越少的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取K4个候选客户端作为需要参与模型训练和上报训练结果的客户端，K4为正整数，这样可以均衡候选客户端参与模型训练的次数。
一具体示例中，可以根据候选客户端的通信网络接入方式确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的通信网络接入方式对候选客户端进行优先级排序，通信网络接入方式包括固网、WiFi和移动网络， 移动网络包括2G，3G，4G，5G，6G等。其中，固网被筛选到的优先级大于等于WiFi被筛选到的优先级，WiFi被筛选到的优先级大于等于移动网络被筛选到的优先级。移动网络中代数越高，被筛选到的优先级越高，比如5G候选客户端被筛选到的优先级高于4G候选客户端被筛选到的优先级，按照优先级的从高到低从候选客户端中选取B个候选客户端作为需要参与模型训练和上报训练结果的客户端，B为正整数。
一具体示例中，可以根据候选客户端的信道质量确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的信道质量对候选客户端进行优先级排序，信道质量越高的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取C个候选客户端作为需要参与模型训练和上报训练结果的客户端，C为正整数，这样可以保证信道质量好的客户端参与模型训练和上报训练结果，保证模型的训练质量。
一具体示例中，可以根据候选客户端的电量状态确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的电量状态对候选客户端进行优先级排序，候选客户端的电量越高，被筛选到的优先级越高，另外，处于充电状态的候选客户端被筛选到的优先级最高，按照优先级的从高到低从候选客户端中选取E个候选客户端作为需要参与模型训练和上报训练结果的客户端，E为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的电量。
一具体示例中，可以根据候选客户端的存储状态确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的存储状态对候选客户端进行优先级排序，候选客户端的可用存储空间越大，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取F个候选客户端作为需要参与模型训练和上报训练结果的客户端，F为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的可用存储空间来进行训练数据和训练结果的存储。
一具体示例中，可以根据候选客户端的算力确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的算力对候选客户端进行优先级排序，候选客户端的算力越大，被筛选到的优先级越高，按照优先级的从高到低从 候选客户端中选取P个候选客户端作为需要参与模型训练和上报训练结果的客户端，P为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的算力来进行训练。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
上述模型指特定联邦学习或联邦元学习的模型。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
上述模型指特定联邦学习或联邦元学习的模型。
其中，进行客户端筛选的候选客户端的标识和不进行客户端筛选的候选客户端的标识组成所述第二筛选条件，候选客户端可以根据自身的标识判断自身是否满足第二筛选条件。
一些实施例中，所述中心设备接收所述客户端上报的训练结果之后，所 述方法还包括：
根据所述训练结果判断模型达到收敛，所述中心设备将收敛的模型和超参数发送给L个推理客户端，所述L大于M，等于M或小于M。
本实施例中，中心设备基于接收到的训练结果判断模型是否收敛，如果模型不收敛，则重复筛选客户端，向客户端发送第一指示，接收客户端上报的训练结果的过程；如果模型收敛，则将收敛的模型和超参数发送给L个推理客户端，推理客户端可以选自候选客户端，还可以是候选客户端之外的其他客户端。
本实施例中，在经过至少一轮训练后需要做一次客户端筛选。在所有轮训练中，至少在进行客户端筛选的轮触发客户端上报第一训练数据和/或第一参数。经过多轮训练(包括多次的客户端筛选)后模型达到收敛时，中心设备将收敛的模型和超参数下发给推理客户端，推理客户端对模型进行性能验证和推理。
一些实施例中，所述中心设备通过以下至少一项向推理客户端发送收敛的模型和超参数：
媒体介入控制MAC控制单元CE；
无线资源控制RRC消息；
非接入层NAS消息；
管理编排消息；
用户面数据；
下行控制信息；
系统信息块SIB；
物理下行控制信道PDCCH的层1信令；
物理下行共享信道PDSCH的信息；
物理随机接入信道PRACH的MSG 2信息；
物理随机接入信道PRACH的MSG 4信息；
物理随机接入信道PRACH的MSG B信息；
广播信道信息或信令；
Xn接口信令；
PC5接口信令；
物理侧边链路控制信道PSCCH的信息或信令；
物理侧边链路共享信道PSSCH的信息；
物理侧边链路广播信道PSBCH的信息；
物理直通链路发现信道PSDCH的信息；
物理直通链路反馈信道PSFCH的信息。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批(batch)的大小，小批(mini batch)的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合(epoch)数，代价函数的选择，神经元激活函数。
相比于传统的联邦学习，联邦元学习通过多任务训练获得一个泛化性能最好的初始化参数，在新的任务下使用该初始化参数可以快速地收敛。联邦元学习的训练中有一些特殊的内容：训练过程分为内迭代和外迭代。因此由中心设备通知给客户端的超参数中会有联邦学习中不涉及的超参数，比如外迭代学习率，内迭代学习率，元学习率，内迭代次数，外迭代次数。
联邦元学习对客户端筛选的需求更大，因为联邦元学习的优势在于泛化性能更好。因此，参与训练的时候尽量做到对所有数据都公平。
对于联邦元学习，下发给不同客户端的一部分超参数可以不同。可以根据每个客户端对应的第一参数(主要是根据第一参数中的数据采集难易程度、客户端的电量状态、客户端的存储状态等)来决定每个客户端的上述一部分超参数。具体的原则包括以下至少一项：
数据采集难度大的客户端建议使用较少的内迭代次数和较大的内迭代步长，数据采集难度小的客户端建议使用较多的内迭代次数和较小的内迭代步长；
电量少的客户端建议使用较少的内迭代次数和较大的内迭代步长，电量多的客户端建议使用较多的内迭代次数和较小的内迭代步长；
可用存储空间少的客户端建议使用较少的内迭代次数和较大的内迭代步长，可用存储空间多的客户端建议使用较多的内迭代次数和较小的内迭代步长；
数据采集难度大的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；数据采集难度大的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率；
电量少的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；电量少的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率；
可用存储空间少的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；可用存储空间少的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率。
本实施例中，所述中心设备为网络侧设备或终端；所述客户端为网络侧设备或终端。如多个网络侧设备联合做联邦学习或联邦元学习的场景，以及多个终端联合做联邦学习或联邦元学习的场景。其中，中心设备与客户端之间的信息交互(包括第一参数、第一指示等)可以通过一次通信完成，也可以通过多次通信完成。
另外，候选客户端可以为网络侧设备，也可以为终端；推理客户端可以为网络侧设备，也可以为终端。
本申请实施例还提供了一种客户端筛选方法，如图5所示，包括：
步骤201：客户端接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
步骤202：所述客户端进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
其中，上述训练结果是参与联邦学习或联邦元学习的客户端进行一轮训练后得到的结果或中间结果，训练结果或中间结果可以是梯度结果、损失函数结果、特定任务的性能结果或上述结果的加密结果等。
在本申请实施例中，中心设备并不是要求所有的候选客户端参与特定联 邦学习或联邦元学习的模型训练，而是由中心设备先对候选客户端进行筛选，确定需要进行模型训练的客户端，然后向客户端发送第一指示，指示客户端参与特定联邦学习或联邦元学习的模型训练，并接收客户端上报的训练结果。这样可以剔除一些条件不好的候选客户端，提高训练的收敛速度，降低中心设备与客户端之间的通信资源开销。并且，通过选取具有代表性的客户端进行联邦式的训练不仅可以提高训练效率，还可以提高模型的泛化性能。
一些实施例中，所述客户端通过以下至少一项向所述中心设备上报训练结果：
媒体介入控制MAC控制单元CE；
无线资源控制RRC消息；
非接入层NAS消息；
物理上行控制信道PUCCH的层1信令；
物理随机接入信道PRACH的MSG 1信息；
物理随机接入信道PRACH的MSG 3信息；
物理随机接入信道PRACH的MSG A信息；
物理上行共享信道PUSCH的信息；
Xn接口信令；
PC5接口信令；
物理侧边链路控制信道PSCCH；
物理侧边链路共享信道PSSCH；
物理侧边链路广播信道PSBCH；
物理直通链路发现信道PSDCH；
物理直通链路反馈信道PSFCH。
一些实施例中，所述客户端接收中心设备的第一指示包括：
所述客户端接收所述中心设备单播的所述第一指示，所述客户端为所述中心设备按照预设的第一筛选条件从候选客户端中筛选出的客户端；或
所述客户端接收所述中心设备广播的所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
一些实施例中，所述客户端进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果包括：
若所述客户端接收所述中心设备单播的所述第一指示，所述客户端进行模型训练并上报训练结果；或
若所述客户端接收所述中心设备广播的所述第一指示，所述客户端进行模型训练并上报训练结果。
本实施例中，在中心设备通信范围内的为候选客户端，上报训练结果的客户端选自候选客户端，可以将所有的候选客户端作为客户端，也可以筛选出部分候选客户端作为客户端。广播是向所有的候选客户端发送第一指示，而单播只是向筛选出的客户端发送第一指示，收到单播的第一指示的候选客户端均需要执行模型训练并上报训练结果。收到广播的第一指示的候选客户端需要判断自身是否满足第二筛选条件，满足第二筛选条件的候选客户端才执行模型训练并上报训练结果。
一些实施例中，所述客户端接收中心设备的第一指示之前，所述方法还包括：
候选客户端向所述中心设备上报第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数，所述第一训练数据用于确定所述第一参数。
本实施例中，候选客户端可以先上报少量的训练数据(即第一训练数据)和/或第一参数，由中心设备根据少量的训练数据和/或第一参数确定参与训练的客户端，筛选出参与模型训练并上报训练结果的客户端，避免所有的客户端都参与训练。
一些实施例中，所述客户端通过以下至少一项向所述中心设备上报第一训练数据和/或第一参数：
媒体介入控制MAC控制单元CE；
无线资源控制RRC消息；
非接入层NAS消息；
物理上行控制信道PUCCH的层1信令；
物理随机接入信道PRACH的MSG 1信息；
物理随机接入信道PRACH的MSG 3信息；
物理随机接入信道PRACH的MSG A信息；
物理上行共享信道PUSCH的信息；
Xn接口信令；
PC5接口信令；
物理侧边链路控制信道PSCCH；
物理侧边链路共享信道PSSCH；
物理侧边链路广播信道PSBCH；
物理直通链路发现信道PSDCH；
物理直通链路反馈信道PSFCH。
一些实施例中，所述候选客户端向所述中心设备仅上报所述第一训练数据，所述第一训练数据用于确定所述第一参数。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(eMBB)，超可靠低延迟通信(URLLC)，大规模机器类型通信(mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播LOS、非视距传播NLOS、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括2G，3G，4G，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如GPU数、CPU数、核数等，或者，算力可以用每秒的运算次数FLOPS或处理器运算单元能力(TOPS、GOPS和/或MOPS)等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
本实施例中，候选客户端可以先向中心设备上报少量的训练数据(即第一训练数据)和/或第一参数，其中，第一参数可以是第一筛选条件的判断参数，中心设备根据第一筛选条件和/或第二筛选条件确定需要参与模型训练和上报训练结果的客户端，该客户端是选自候选客户端，具体地，可以有M个候选客户端，从其中确定N个客户端需要进行训练客户端筛选和上报，N可以小于M，也可以等于M。
一具体示例中，可以根据候选客户端的数据类型确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的数据类型对候选客户端进行分组，每组内的候选客户端的数据类型相同或相近。在筛选客户端时，从每一组候选客户端中选取K1个候选客户端作为需要参与模型训练和上报训练结果的客户端，K1为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端的业务类型确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的业务类型对候选客户端进行分组，每组内的候选客户端的业务类型相同或相近。在筛选客户端时，从每一组候选客户端中选取K2个候选客户端作为需要参与模型训练和上报训练结果的客户端，K2为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端的数据分布参数确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的数据分布参数对候选客户端进行分组，每组内的候选客户端的数据分布参数相同或相近。在筛选客户端时，从每一组候选客户端中选取K3个候选客户端作为需要参与模型训练和上报训练结果的客户端，K3为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端的工作场景确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的工作场景对候选客户端进行分组，每组内的候选客户端的工作场景相同或相近。在筛选客户端时，从每一组候选客户端中选取A个候选客户端作为需要参与模型训练和上报训练结果的客户端，A为正整数，这样可以保证参与训练的客户端的多样性，保证每一组候选客户端都有客户端参与模型训练，泛化性能更好，顾及每一组的客户端。
一具体示例中，可以根据候选客户端收集数据的难易程度确定需要参与模型训练和上报训练结果的客户端，按照候选客户端收集数据的难易程度对候选客户端进行优先级排序，收集数据难度越小的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取D个候选客户端作为需要参与模型训练和上报训练结果的客户端，D为正整数，这样可以降低数据采集的难度。
一具体示例中，可以根据候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度确定需要参与模型训练和上报训练结果的客户端，按照意愿程度对候选客户端进行优先级排序，意愿程度越高的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取G个候选客户端作为需要参与模型训练和上报训练结果的客户端，G为正整数，这样可以保证意愿程度高的候选客户端参与模型训练。
一具体示例中，可以根据候选客户端参与特定联邦学习或联邦元学习的模型训练的次数确定需要参与模型训练和上报训练结果的客户端，按照参与模型训练的次数对候选客户端进行优先级排序，已参与模型训练的次数越少的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取K4个候选客户端作为需要参与模型训练和上报训练结果的客户端，K4为正整数，这样可以均衡候选客户端参与模型训练的次数。
一具体示例中，可以根据候选客户端的通信网络接入方式确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的通信网络接入方式对候选客户端进行优先级排序，通信网络接入方式包括固网、WiFi和移动网络，移动网络包括2G，3G，4G，5G，6G等。其中，固网被筛选到的优先级大于等 于WiFi被筛选到的优先级，WiFi被筛选到的优先级大于等于移动网络被筛选到的优先级。移动网络中代数越高，被筛选到的优先级越高，比如5G候选客户端被筛选到的优先级高于4G候选客户端被筛选到的优先级，按照优先级的从高到低从候选客户端中选取B个候选客户端作为需要参与模型训练和上报训练结果的客户端，B为正整数。
一具体示例中，可以根据候选客户端的信道质量确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的信道质量对候选客户端进行优先级排序，信道质量越高的候选客户端，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取C个候选客户端作为需要参与模型训练和上报训练结果的客户端，C为正整数，这样可以保证信道质量好的客户端参与模型训练和上报训练结果，保证模型的训练质量。
一具体示例中，可以根据候选客户端的电量状态确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的电量状态对候选客户端进行优先级排序，候选客户端的电量越高，被筛选到的优先级越高，另外，处于充电状态的候选客户端被筛选到的优先级最高，按照优先级的从高到低从候选客户端中选取E个候选客户端作为需要参与模型训练和上报训练结果的客户端，E为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的电量。
一具体示例中，可以根据候选客户端的存储状态确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的存储状态对候选客户端进行优先级排序，候选客户端的可用存储空间越大，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取F个候选客户端作为需要参与模型训练和上报训练结果的客户端，F为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的可用存储空间来进行训练数据和训练结果的存储。
一具体示例中，可以根据候选客户端的算力确定需要参与模型训练和上报训练结果的客户端，按照候选客户端的算力对候选客户端进行优先级排序，候选客户端的算力越大，被筛选到的优先级越高，按照优先级的从高到低从候选客户端中选取P个候选客户端作为需要参与模型训练和上报训练结果的 客户端，P为正整数，这样可以保证参与模型训练和上报训练结果的客户端有足够的算力来进行训练。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
上述模型指特定联邦学习或联邦元学习的模型。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
上述模型指特定联邦学习或联邦元学习的模型。
其中，进行客户端筛选的候选客户端的标识和不进行客户端筛选的候选客户端的标识组成所述第二筛选条件，候选客户端可以根据自身的标识判断自身是否满足第二筛选条件。
一些实施例中，向所述中心设备上报训练结果之后，所述方法还包括：
推理客户端接收所述中心设备发送的收敛的模型和超参数。
本实施例中，中心设备基于接收到的训练结果判断模型是否收敛，如果模型不收敛，则重复筛选客户端，向客户端发送第一指示，接收客户端上报的训练结果的过程；如果模型收敛，则将收敛的模型和超参数发送给L个推理客户端，推理客户端可以选自候选客户端，还可以是候选客户端之外的其他客户端。
本实施例中，在经过至少一轮训练后需要做一次客户端筛选。在所有轮训练中，至少在进行客户端筛选的轮触发客户端上报第一训练数据和/或第一参数。经过多轮训练(包括多次的客户端筛选)后模型达到收敛时，中心设备将收敛的模型和超参数下发给推理客户端，推理客户端对模型进行性能验证和推理。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
相比于传统的联邦学习，联邦元学习通过多任务训练获得一个泛化性能最好的初始化参数，在新的任务下使用该初始化参数可以快速地收敛。联邦元学习的训练中有一些特殊的内容：训练过程分为内迭代和外迭代。因此由中心设备通知给客户端的超参数中会有联邦学习中不涉及的超参数，比如外迭代学习率，内迭代学习率，元学习率，内迭代次数，外迭代次数。
联邦元学习对客户端筛选的需求更大，因为联邦元学习的优势在于泛化性能更好。因此，参与训练的时候尽量做到对所有数据都公平。
对于联邦元学习，下发给不同客户端的一部分超参数可以不同。可以根据每个客户端对应的第一参数(主要是根据第一参数中的数据采集难易程度、客户端的电量状态、客户端的存储状态等)来决定每个客户端的上述一部分超参数。具体的原则包括以下至少一项：
数据采集难度大的客户端建议使用较少的内迭代次数和较大的内迭代步 长，数据采集难度小的客户端建议使用较多的内迭代次数和较小的内迭代步长；
电量少的客户端建议使用较少的内迭代次数和较大的内迭代步长，电量多的客户端建议使用较多的内迭代次数和较小的内迭代步长；
可用存储空间少的客户端建议使用较少的内迭代次数和较大的内迭代步长，可用存储空间多的客户端建议使用较多的内迭代次数和较小的内迭代步长；
数据采集难度大的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；数据采集难度大的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率；
电量少的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；电量少的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率；
可用存储空间少的客户端数较多时，建议使用较少的外迭代次数和较大的外迭代学习率；可用存储空间少的客户端数较少时，建议使用较多的外迭代次数和较小的外迭代学习率。
一些实施例中，所述推理客户端接收所述中心设备发送的收敛的模型和超参数之后，所述方法还包括：
所述推理客户端对所述模型进行性能验证；
若性能验证结果满足预设的第一条件，所述推理客户端将所述模型用于推理。
其中，第一条件可以是中心设备配置或预配置或协议约定的，推理客户端对模型进行性能验证后，还可以将是否进行推理的结果上报给中心设备。
一些实施例中，进行性能验证的模型为所述中心设备下发的模型，或，所述中心设备下发的模型经过微调后的模型。
本实施例中，推理客户端可以直接利用中心设备下发的模型进行性能验证，也可以是将中心设备下发的模型进行微调(fine-tuning)后再进行性能验证。对于元学习的微调，每个推理客户端对应的元学习相关的特殊超参数可以不同。可以根据每个推理客户端对应的第一参数(主要是根据第一参数 中的客户端筛选难易程度、电量状态、存储状态等)来决定每个推理客户端的元学习相关的特殊超参数。
本实施例中，所述中心设备为网络侧设备或终端；所述客户端为网络侧设备或终端。如多个网络侧设备联合做联邦学习或联邦元学习的场景，以及多个终端联合做联邦学习或联邦元学习的场景。其中，中心设备与客户端之间的信息交互(包括第一参数、第一指示等)可以通过一次通信完成，也可以通过多次通信完成。
另外，候选客户端可以为网络侧设备，也可以为终端；推理客户端可以为网络侧设备，也可以为终端。
上述实施例中，模型可以为信道估计模型、移动性预测模型等。本申请实施例的技术方案可以应用于6G网络中，还可以应用于5G和5.5G网络中。
本申请实施例提供的客户端筛选方法，执行主体可以为客户端筛选装置。本申请实施例中以客户端筛选装置执行客户端筛选方法为例，说明本申请实施例提供的客户端筛选装置。
本申请实施例提供一种客户端筛选装置，包括：
发送模块，用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
接收模块，用于接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
在本申请实施例中，中心设备并不是要求所有的候选客户端参与特定联邦学习或联邦元学习的模型训练，而是由中心设备先对候选客户端进行筛选，确定需要进行模型训练的客户端，然后向客户端发送第一指示，指示客户端参与特定联邦学习或联邦元学习的模型训练，并接收客户端上报的训练结果。这样可以剔除一些条件不好的候选客户端，提高训练的收敛速度，降低中心设备与客户端之间的通信资源开销。并且，通过选取具有代表性的客户端进行联邦式的训练不仅可以提高训练效率，还可以提高模型的泛化性能。
一些实施例中，所述发送模块具体用于按照预设的第一筛选条件从M个候选客户端中筛选出N个客户端，向所述N个客户端单播所述第一指示，M，N为正整数，N小于或等于M；或
向所述M个候选客户端广播所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
一些实施例中，所述接收模块还用于接收所述候选客户端上报的第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数。
一些实施例中，所述接收模块用于仅接收所述候选客户端上报的第一训练数据，根据所述第一训练数据确定所述第一参数。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(eMBB)，超可靠低延迟通信(URLLC)，大规模机器类型通信(mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播LOS、非视距传播NLOS、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括2G，3G，4G，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如GPU(图形处理器)数、CPU(中央处理器)数、核数等，或者，算力可以用每秒的运算次数FLOPS或处理器运算单元能力(TOPS、GOPS和/或MOPS)等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，所述发送模块还用于根据所述训练结果判断模型达到收敛，将收敛的模型和超参数发送给L个推理客户端，所述L大于M，等于M或小于M。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
一些实施例中，所述中心设备为网络侧设备或终端；
所述客户端为网络侧设备或终端。
另外，候选客户端可以为网络侧设备，也可以为终端；推理客户端可以为网络侧设备，也可以为终端。
本申请实施例提供一种客户端筛选装置，包括：
接收模块，用于接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
上报模块，用于进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
一些实施例中，所述接收模块用于接收所述中心设备单播的所述第一指示，所述客户端为所述中心设备按照预设的第一筛选条件从候选客户端中筛选出的客户端；或
接收所述中心设备广播的所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
一些实施例中，所述上报模块用于若所述客户端接收所述中心设备单播的所述第一指示，进行模型训练并上报训练结果；或
若所述客户端接收所述中心设备广播的所述第一指示，进行模型训练并上报训练结果。
一些实施例中，所述上报模块还用于向所述中心设备上报第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数，所述第一训练数据用于确定所述第一参数。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(eMBB)，超可靠低延迟通信(URLLC)，大规模机器类型通信(mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播LOS、非视距传播NLOS、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括2G，3G，4G，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如GPU数、CPU数、核数等，或者，算力可以用每秒的运算次数FLOPS或处理器运算单元能力(TOPS、GOPS和/或MOPS)等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，所述接收模块还用于接收所述中心设备发送的收敛的模型和超参数。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
一些实施例中，所述超参数的第一部分由所述推理客户端对应的第一参数决定，所述第一部分包括以下至少一项：
外迭代学习率，内迭代学习率，元学习率，内迭代次数，外迭代次数。
一些实施例中，所述装置还包括：
处理模块，用于对所述模型进行性能验证；若性能验证结果满足预设的第一条件，将所述模型用于推理。
一些实施例中，进行性能验证的模型为所述中心设备下发的模型，或，所述中心设备下发的模型经过微调后得到的模型。
一些实施例中，所述中心设备为网络侧设备或终端；
所述客户端为网络侧设备或终端。
另外，候选客户端可以为网络侧设备，也可以为终端；推理客户端可以为网络侧设备，也可以为终端。
本申请实施例中的客户端筛选装置可以是电子设备，例如具有操作系统的电子设备，也可以是电子设备中的部件，例如集成电路或芯片。该电子设备可以是终端，也可以为除终端之外的其他设备。示例性的，终端可以包括但不限于上述所列举的终端11的类型，其他设备可以为服务器、网络附属存储器(Network Attached Storage，NAS)等，本申请实施例不作具体限定。
本申请实施例提供的客户端筛选装置能够实现图4至图5的方法实施例实现的各个过程，并达到相同的技术效果，为避免重复，这里不再赘述。
可选的，如图6所示，本申请实施例还提供一种通信设备600，包括处理器601和存储器602，存储器602上存储有可在所述处理器601上运行的程序或指令，例如，该通信设备600为中心设备时，该程序或指令被处理器601执行时实现上述客户端筛选方法实施例的各个步骤，且能达到相同的技术效果。该通信设备600为客户端时，该程序或指令被处理器601执行时实现上述客户端筛选方法实施例的各个步骤，且能达到相同的技术效果，为避免重复，这里不再赘述。
本申请实施例还提供了一种中心设备，该中心设备包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如上所述的客户端筛选方法的步骤。
本申请实施例还提供了一种中心设备，包括处理器及通信接口，其中，所述通信接口用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
本申请实施例还提供了一种客户端，该客户端包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如上所述的客户端筛选方法的步骤。
本申请实施例还提供了一种客户端，包括处理器及通信接口，其中，所述通信接口用于接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；所述处理器用于进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
上述中心设备可以为网络侧设备或终端，客户端可以为网络侧设备或终端。
当中心设备和/或客户端为终端时，本申请实施例还提供一种终端，包括处理器和通信接口，该终端实施例与上述终端侧方法实施例对应，上述方法实施例的各个实施过程和实现方式均可适用于该终端实施例中，且能达到相 同的技术效果。具体地，图7为实现本申请实施例的一种终端的硬件结构示意图。
该终端700包括但不限于：射频单元701、网络模块702、音频输出单元703、输入单元704、传感器705、显示单元706、用户输入单元707、接口单元708、存储器709以及处理器710等中的至少部分部件。
本领域技术人员可以理解，终端700还可以包括给各个部件供电的电源(比如电池)，电源可以通过电源管理系统与处理器7 10逻辑相连，从而通过电源管理系统实现管理充电、放电、以及功耗管理等功能。图7中示出的终端结构并不构成对终端的限定，终端可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置，在此不再赘述。
应理解的是，本申请实施例中，输入单元704可以包括图形处理单元(Graphics Processing Unit，GPU)7041和麦克风7042，图形处理器7041对在视频捕获模式或图像捕获模式中由图像捕获装置(如摄像头)获得的静态图片或视频的图像数据进行处理。显示单元706可包括显示面板7061，可以采用液晶显示器、有机发光二极管等形式来配置显示面板7061。用户输入单元707包括触控面板7071以及其他输入设备7072中的至少一种。触控面板7071，也称为触摸屏。触控面板7071可包括触摸检测装置和触摸控制器两个部分。其他输入设备7072可以包括但不限于物理键盘、功能键(比如音量控制按键、开关按键等)、轨迹球、鼠标、操作杆，在此不再赘述。
本申请实施例中，射频单元701接收来自网络侧设备的下行数据后，可以传输给处理器710进行处理；另外，射频单元701可以向网络侧设备发送上行数据。通常，射频单元701包括但不限于天线、放大器、收发信机、耦合器、低噪声放大器、双工器等。
存储器709可用于存储软件程序或指令以及各种数据。存储器709可主要包括存储程序或指令的第一存储区和存储数据的第二存储区，其中，第一存储区可存储操作系统、至少一个功能所需的应用程序或指令(比如声音播放功能、图像播放功能等)等。此外，存储器709可以包括易失性存储器或非易失性存储器，或者，存储器709可以包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(Read-Only Memory，ROM)、可编程 只读存储器(Programmable ROM，PROM)、可擦除可编程只读存储器(Erasable PROM，EPROM)、电可擦除可编程只读存储器(Electrically EPROM，EEPROM)或闪存。易失性存储器可以是随机存取存储器(Random Access Memory，RAM)，静态随机存取存储器(Static RAM，SRAM)、动态随机存取存储器(Dynamic RAM，DRAM)、同步动态随机存取存储器(Synchronous DRAM，SDRAM)、双倍数据速率同步动态随机存取存储器(Double Data Rate SDRAM，DDRSDRAM)、增强型同步动态随机存取存储器(Enhanced SDRAM，ESDRAM)、同步连接动态随机存取存储器(Synch link DRAM，SLDRAM)和直接内存总线随机存取存储器(Direct Rambus RAM，DRRAM)。本申请实施例中的存储器709包括但不限于这些和任意其它适合类型的存储器。
处理器710可包括一个或多个处理单元；可选的，处理器710集成应用处理器和调制解调处理器，其中，应用处理器主要处理涉及操作系统、用户界面和应用程序等的操作，调制解调处理器主要处理无线通信信号，如基带处理器。可以理解的是，上述调制解调处理器也可以不集成到处理器710中。
一些实施例中，中心设备为终端，处理器710用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
一些实施例中，处理器710用于按照预设的第一筛选条件从M个候选客户端中筛选出N个客户端，向所述N个客户端单播所述第一指示，M，N为正整数，N小于或等于M；或
向所述M个候选客户端广播所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
一些实施例中，处理器710用于接收所述候选客户端上报的第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数。
一些实施例中，处理器710用于仅接收所述候选客户端上报的第一训练数据，根据所述第一训练数据确定所述第一参数。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(eMBB)，超可靠低延迟通信(URLLC)，大规模机器类型通信(mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播LOS、非视距传播NLOS、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括2G，3G，4G，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如GPU数、CPU数、核数等，或者，算力可以用每秒的运算次数FLOPS或处理器运算单元能力(TOPS、GOPS和/或MOPS)等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，处理器710用于根据所述训练结果判断模型达到收敛，将收敛的模型和超参数发送给L个推理客户端，所述L大于M，等于M或小于M。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
一些实施例中，所述中心设备为网络侧设备或终端；
所述客户端为网络侧设备或终端。
一些实施例中，客户端为终端，处理器710用于接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
一些实施例中，处理器710用于接收所述中心设备单播的所述第一指示，所述客户端为所述中心设备按照预设的第一筛选条件从候选客户端中筛选出 的客户端；或
接收所述中心设备广播的所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
一些实施例中，处理器710用于若所述客户端接收所述中心设备单播的所述第一指示，进行模型训练并上报训练结果；或
若所述客户端接收所述中心设备广播的所述第一指示，进行模型训练并上报训练结果。
一些实施例中，处理器710用于向所述中心设备上报第一训练数据和/或第一参数，所述第一参数可以是所述第一筛选条件的判断参数，所述第一训练数据用于确定所述第一参数。
一些实施例中，所述第一参数包括以下至少一项：
所述候选客户端的数据类型；
所述候选客户端的数据分布参数；
所述候选客户端的业务类型，比如增强移动宽带(eMBB)，超可靠低延迟通信(URLLC)，大规模机器类型通信(mMTC)，其他6G新场景等；
所述候选客户端的工作场景，包括但不限于：高速、低速、视距传播LOS、非视距传播NLOS、高信噪比、低信噪比等工作场景；
所述候选客户端的通信网络接入方式，包括移动网络、WiFi和固网，其中移动网络包括2G，3G，4G，5G和6G；
所述候选客户端的信道质量；
所述候选客户端收集数据的难易程度；
所述候选客户端的电量状态，比如可用剩余电量的具体值，或分级描述结果，充电或不充电等；
所述候选客户端的存储状态，比如可用内存的具体值，或分级描述结果；
所述候选客户端的算力，比如GPU数、CPU数、核数等，或者，算力可以用每秒的运算次数FLOPS或处理器运算单元能力(TOPS、GOPS和/或MOPS)等进行表示；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
一些实施例中，单播的所述第一指示包括以下至少一项：
模型文件；
模型的结构；
模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，广播的所述第一指示包括以下至少一项：
参与训练的候选客户端的标识；
不参与训练的候选客户端的标识；
所述第一筛选条件；
模型文件；
模型的结构；模型初始化参数；
模型的输入物理量；
模型的输出物理量；
模型对应的参考点；
模型的超参数；
通信信息。
一些实施例中，处理器710用于接收所述中心设备发送的收敛的模型和超参数。
一些实施例中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
一些实施例中，所述超参数包括以下至少一项：
学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代 价函数的选择，神经元激活函数。
一些实施例中，所述超参数的第一部分由所述推理客户端对应的第一参数决定，所述第一部分包括以下至少一项：
外迭代学习率，内迭代学习率，元学习率，内迭代次数，外迭代次数。
一些实施例中，处理器710用于对所述模型进行性能验证；若性能验证结果满足预设的第一条件，将所述模型用于推理。
一些实施例中，进行性能验证的模型为所述中心设备下发的模型，或，所述中心设备下发的模型经过微调后得到的模型。
一些实施例中，所述中心设备为网络侧设备或终端；
所述客户端为网络侧设备或终端。
另外，候选客户端可以为网络侧设备，也可以为终端；推理客户端可以为网络侧设备，也可以为终端。
当中心设备和/或客户端为网络侧设备时，本申请实施例还提供一种网络侧设备，包括处理器和通信接口。该网络侧设备实施例与上述网络侧设备方法实施例对应，上述方法实施例的各个实施过程和实现方式均可适用于该网络侧设备实施例中，且能达到相同的技术效果。
具体地，本申请实施例还提供了一种网络侧设备。如图8所示，该网络侧设备800包括：天线81、射频装置82、基带装置83、处理器84和存储器85。天线81与射频装置82连接。在上行方向上，射频装置82通过天线81接收信息，将接收的信息发送给基带装置83进行处理。在下行方向上，基带装置83对要发送的信息进行处理，并发送给射频装置82，射频装置82对收到的信息进行处理后经过天线81发送出去。
以上实施例中网络侧设备执行的方法可以在基带装置83中实现，该基带装置83包括基带处理器。
基带装置83例如可以包括至少一个基带板，该基带板上设置有多个芯片，如图8所示，其中一个芯片例如为基带处理器，通过总线接口与存储器85连接，以调用存储器85中的程序，执行以上方法实施例中所示的网络设备操作。
该网络侧设备还可以包括网络接口86，该接口例如为通用公共无线接口(common public radio interface，CPRI)。
具体地，本发明实施例的网络侧设备800还包括：存储在存储器85上并可在处理器84上运行的指令或程序，处理器84调用存储器85中的指令或程序执行如上所述的客户端筛选方法，并达到相同的技术效果，为避免重复，故不在此赘述。
本申请实施例还提供一种可读存储介质，所述可读存储介质上存储有程序或指令，该程序或指令被处理器执行时实现上述客户端筛选方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
其中，所述处理器为上述实施例中所述的终端中的处理器。所述可读存储介质，包括计算机可读存储介质，如计算机只读存储器ROM、随机存取存储器RAM、磁碟或者光盘等。
本申请实施例另提供了一种芯片，所述芯片包括处理器和通信接口，所述通信接口和所述处理器耦合，所述处理器用于运行程序或指令，实现上述客户端筛选方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
应理解，本申请实施例提到的芯片还可以称为系统级芯片，系统芯片，芯片系统或片上系统芯片等。
本申请实施例另提供了一种计算机程序/程序产品，所述计算机程序/程序产品被存储在存储介质中，所述计算机程序/程序产品被至少一个处理器执行以实现上述客户端筛选方法实施例的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
本申请实施例还提供了一种客户端筛选系统，包括：中心设备及客户端，所述中心设备可用于执行如上所述的客户端筛选方法的步骤，所述客户端可用于执行如上所述的客户端筛选方法的步骤。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。此外，需要指出的是，本申 请实施方式中的方法和装置的范围不限按示出或讨论的顺序来执行功能，还可包括根据所涉及的功能按基本同时的方式或按相反的顺序来执行功能，例如，可以按不同于所描述的次序来执行所描述的方法，并且还可以添加、省去、或组合各种步骤。另外，参照某些示例所描述的特征可在其他示例中被组合。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以计算机软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本申请各个实施例所述的方法。
上面结合附图对本申请的实施例进行了描述，但是本申请并不局限于上述的具体实施方式，上述的具体实施方式仅仅是示意性的，而不是限制性的，本领域的普通技术人员在本申请的启示下，在不脱离本申请宗旨和权利要求所保护的范围情况下，还可做出很多形式，均属于本申请的保护之内。

Claims (31)

  1. 一种客户端筛选方法，包括：
    中心设备向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
    所述中心设备接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
  2. 根据权利要求1所述的客户端筛选方法，其中，所述中心设备向所述客户端发送第一指示包括：
    所述中心设备按照预设的第一筛选条件从M个候选客户端中筛选出N个客户端，向所述N个客户端单播所述第一指示，M，N为正整数，N小于或等于M；或
    所述中心设备向所述M个候选客户端广播所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
  3. 根据权利要求2所述的客户端筛选方法，其中，所述中心设备向客户端发送第一指示之前，所述方法还包括：
    所述中心设备接收所述候选客户端上报的第一训练数据和/或第一参数，所述第一参数是所述第一筛选条件的判断参数。
  4. 根据权利要求3所述的客户端筛选方法，其中，
    所述中心设备仅接收所述候选客户端上报的第一训练数据，根据所述第一训练数据确定所述第一参数。
  5. 根据权利要求3或4所述的客户端筛选方法，其中，所述第一参数包括以下至少一项：
    所述候选客户端的数据类型；
    所述候选客户端的数据分布参数；
    所述候选客户端的业务类型；
    所述候选客户端的工作场景；
    所述候选客户端的通信网络接入方式；
    所述候选客户端的信道质量；
    所述候选客户端收集数据的难易程度；
    所述候选客户端的电量状态；
    所述候选客户端的存储状态；
    所述候选客户端的算力；
    所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
    所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
  6. 根据权利要求2所述的客户端筛选方法，其中，单播的所述第一指示包括以下至少一项：
    模型文件；
    模型的结构；
    模型初始化参数；
    模型的输入物理量；
    模型的输出物理量；
    模型对应的参考点；
    模型的超参数；
    通信信息。
  7. 根据权利要求2所述的客户端筛选方法，其中，广播的所述第一指示包括以下至少一项：
    参与训练的候选客户端的标识；
    不参与训练的候选客户端的标识；
    所述第一筛选条件；
    模型文件；
    模型的结构；
    模型初始化参数；
    模型的输入物理量；
    模型的输出物理量；
    模型对应的参考点；
    模型的超参数；
    通信信息。
  8. 根据权利要求3所述的客户端筛选方法，其中，所述中心设备接收所述客户端上报的训练结果之后，所述方法还包括：
    根据所述训练结果判断模型达到收敛，所述中心设备将收敛的模型和超参数发送给L个推理客户端，所述L大于M，等于M或小于M。
  9. 根据权利要求8所述的客户端筛选方法，其中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
  10. 根据权利要求8所述的客户端筛选方法，其中，所述超参数包括以下至少一项：
    学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
  11. 根据权利要求1所述的客户端筛选方法，其中，
    所述中心设备为网络侧设备或终端；
    所述客户端为网络侧设备或终端。
  12. 一种客户端筛选方法，包括：
    客户端接收中心设备的第一指示，所述第一指示用以指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
    所述客户端进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
  13. 根据权利要求12所述的客户端筛选方法，其中，所述客户端接收中心设备的第一指示包括：
    所述客户端接收所述中心设备单播的所述第一指示，所述客户端为所述中心设备按照预设的第一筛选条件从候选客户端中筛选出的客户端；或
    所述客户端接收所述中心设备广播的所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
  14. 根据权利要求13所述的客户端筛选方法，其中，所述客户端进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果包括：
    若所述客户端接收所述中心设备单播的所述第一指示，所述客户端进行模型训练并上报训练结果；或
    若所述客户端接收所述中心设备广播的所述第一指示，所述客户端进行模型训练并上报训练结果。
  15. 根据权利要求13所述的客户端筛选方法，其中，所述客户端接收中心设备的第一指示之前，所述方法还包括：
    候选客户端向所述中心设备上报第一训练数据和/或第一参数，所述第一参数是所述第一筛选条件的判断参数，所述第一训练数据用于确定所述第一参数。
  16. 根据权利要求15所述的客户端筛选方法，其中，所述第一参数包括以下至少一项：
    所述候选客户端的数据类型；
    所述候选客户端的数据分布参数；
    所述候选客户端的业务类型；
    所述候选客户端的工作场景；
    所述候选客户端的通信网络接入方式；
    所述候选客户端的信道质量；
    所述候选客户端收集数据的难易程度；
    所述候选客户端的电量状态；
    所述候选客户端的存储状态；
    所述候选客户端的算力；
    所述候选客户端参与特定联邦学习或联邦元学习的模型训练的次数；
    所述候选客户端参与特定联邦学习或联邦元学习的模型训练的意愿程度。
  17. 根据权利要求13所述的客户端筛选方法，其中，单播的所述第一指示包括以下至少一项：
    模型文件；
    模型的结构；
    模型初始化参数；
    模型的输入物理量；
    模型的输出物理量；
    模型对应的参考点；
    模型的超参数；
    通信信息。
  18. 根据权利要求13所述的客户端筛选方法，其中，广播的所述第一指示包括以下至少一项：
    参与训练的候选客户端的标识；
    不参与训练的候选客户端的标识；
    所述第一筛选条件；
    模型文件；
    模型的结构；模型初始化参数；
    模型的输入物理量；
    模型的输出物理量；
    模型对应的参考点；
    模型的超参数；
    通信信息。
  19. 根据权利要求15所述的客户端筛选方法，其中，向所述中心设备上报训练结果之后，所述方法还包括：
    推理客户端接收所述中心设备发送的收敛的模型和超参数。
  20. 根据权利要求19所述的客户端筛选方法，其中，所述模型为联邦元学习模型，所述超参数由所述第一参数决定。
  21. 根据权利要求19所述的客户端筛选方法，其中，所述超参数包括以下至少一项：
    学习率，外迭代学习率，内迭代学习率，元学习率，迭代次数，内迭代次数，外迭代次数，训练所需要的数据量，批的大小，小批的大小，正则化参数，神经网络的层数，每一个隐藏层中神经元的个数，学习的回合数，代价函数的选择，神经元激活函数。
  22. 根据权利要求21所述的客户端筛选方法，其中，所述超参数的第一部分由所述推理客户端对应的第一参数决定，所述第一部分包括以下至少一项：
    外迭代学习率，内迭代学习率，元学习率，内迭代次数，外迭代次数。
  23. 根据权利要求19所述的客户端筛选方法，其中，所述推理客户端接收所述中心设备发送的收敛的模型和超参数之后，所述方法还包括：
    所述推理客户端对所述模型进行性能验证；
    若性能验证结果满足预设的第一条件，所述推理客户端将所述模型用于推理。
  24. 根据权利要求23所述的客户端筛选方法，其中，进行性能验证的模型为所述中心设备下发的模型，或，所述中心设备下发的模型经过微调后得到的模型。
  25. 根据权利要求12所述的客户端筛选方法，其中，
    所述中心设备为网络侧设备或终端；
    所述客户端为网络侧设备或终端。
  26. 一种客户端筛选装置，包括：
    发送模块，用于向客户端发送第一指示，指示所述客户端参与特定联邦学习或联邦元学习的模型训练；
    接收模块，用于接收所述客户端上报的训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
  27. 根据权利要求26所述的客户端筛选装置，其中，
    所述发送模块具体用于按照预设的第一筛选条件从M个候选客户端中筛选出N个客户端，向所述N个客户端单播所述第一指示，M，N为正整数，N小于或等于M；或
    向所述M个候选客户端广播所述第一指示，所述第一指示携带有第二筛选条件，所述第二筛选条件用于筛选上报训练结果的客户端，所述客户端满足所述第二筛选条件。
  28. 一种客户端筛选装置，包括：
    接收模块，用于接收中心设备的第一指示，所述第一指示用以指示所述 客户端参与特定联邦学习或联邦元学习的模型训练；
    上报模块，用于进行特定联邦学习或联邦元学习的模型训练，并向所述中心设备上报训练结果，所述训练结果为所述客户端执行一轮模型训练后的结果或中间结果。
  29. 一种客户端，包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如权利要求12至25任一项所述的客户端筛选方法的步骤。
  30. 一种中心设备，包括处理器和存储器，所述存储器存储可在所述处理器上运行的程序或指令，所述程序或指令被所述处理器执行时实现如权利要求1至11任一项所述的客户端筛选方法的步骤。
  31. 一种可读存储介质，所述可读存储介质上存储程序或指令，所述程序或指令被处理器执行时实现如权利要求1-11任一项所述的客户端筛选方法，或者实现如权利要求12至25任一项所述的客户端筛选方法的步骤。
PCT/CN2022/138755 2021-12-15 2022-12-13 客户端筛选方法及装置、客户端及中心设备 WO2023109827A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111537989.6 2021-12-15
CN202111537989.6A CN116266216A (zh) 2021-12-15 2021-12-15 客户端筛选方法及装置、客户端及中心设备

Publications (1)

Publication Number Publication Date
WO2023109827A1 true WO2023109827A1 (zh) 2023-06-22

Family

ID=86742979

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/138755 WO2023109827A1 (zh) 2021-12-15 2022-12-13 客户端筛选方法及装置、客户端及中心设备

Country Status (2)

Country Link
CN (1) CN116266216A (zh)
WO (1) WO2023109827A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117476217A (zh) * 2023-12-26 2024-01-30 山东大学齐鲁医院 一种慢性心脏病病情发展趋势预测系统
CN117692939A (zh) * 2024-02-02 2024-03-12 南京邮电大学 一种动态通信环境下的客户端调度方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020180218A1 (en) * 2019-03-01 2020-09-10 Telefonaktiebolaget Lm Ericsson (Publ) Dynamic network configuration
CN113191484A (zh) * 2021-04-25 2021-07-30 清华大学 基于深度强化学习的联邦学习客户端智能选取方法及系统
CN113221470A (zh) * 2021-06-10 2021-08-06 南方电网科学研究院有限责任公司 一种用于电网边缘计算系统的联邦学习方法及其相关装置
CN113283185A (zh) * 2021-07-23 2021-08-20 平安科技（深圳）有限公司 联邦模型训练、客户画像方法、装置、设备及介质
CN113537518A (zh) * 2021-07-19 2021-10-22 哈尔滨工业大学 基于联邦学习的模型训练方法及装置、设备、存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020180218A1 (en) * 2019-03-01 2020-09-10 Telefonaktiebolaget Lm Ericsson (Publ) Dynamic network configuration
CN113191484A (zh) * 2021-04-25 2021-07-30 清华大学 基于深度强化学习的联邦学习客户端智能选取方法及系统
CN113221470A (zh) * 2021-06-10 2021-08-06 南方电网科学研究院有限责任公司 一种用于电网边缘计算系统的联邦学习方法及其相关装置
CN113537518A (zh) * 2021-07-19 2021-10-22 哈尔滨工业大学 基于联邦学习的模型训练方法及装置、设备、存储介质
CN113283185A (zh) * 2021-07-23 2021-08-20 平安科技（深圳）有限公司 联邦模型训练、客户画像方法、装置、设备及介质

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117476217A (zh) * 2023-12-26 2024-01-30 山东大学齐鲁医院 一种慢性心脏病病情发展趋势预测系统
CN117476217B (zh) * 2023-12-26 2024-03-26 山东大学齐鲁医院 一种慢性心脏病病情发展趋势预测系统
CN117692939A (zh) * 2024-02-02 2024-03-12 南京邮电大学 一种动态通信环境下的客户端调度方法
CN117692939B (zh) * 2024-02-02 2024-04-12 南京邮电大学 一种动态通信环境下的客户端调度方法

Also Published As

Publication number Publication date
CN116266216A (zh) 2023-06-20

Similar Documents

Publication Publication Date Title
WO2023109827A1 (zh) 客户端筛选方法及装置、客户端及中心设备
Xia et al. Federated-learning-based client scheduling for low-latency wireless communications
CN114363921A (zh) Ai网络参数的配置方法和设备
WO2023066288A1 (zh) 模型请求方法、模型请求处理方法及相关设备
US20230403206A1 (en) Model training method and apparatus
WO2023040887A1 (zh) 信息上报方法、装置、终端及可读存储介质
WO2022105907A1 (zh) Ai网络部分输入缺失的处理方法和设备
WO2023109828A1 (zh) 数据收集方法及装置、第一设备、第二设备
CN114501353B (zh) 通信信息的发送、接收方法及通信设备
Yan et al. Bayesian optimization for online management in dynamic mobile edge computing
CN115623509A (zh) Tci状态确定方法、装置、终端及网络侧设备
WO2023040885A1 (zh) 参数选择方法、参数配置方法、终端及网络侧设备
WO2023036309A1 (zh) 参考信号序列生成方法、装置、设备及介质
WO2023179540A1 (zh) 信道预测方法、装置及无线通信设备
WO2023179651A1 (zh) 波束处理方法、装置及设备
WO2023098535A1 (zh) 信息交互方法、装置及通信设备
WO2024032694A1 (zh) Csi预测处理方法、装置、通信设备及可读存储介质
WO2024067665A1 (zh) Csi预测处理方法、装置、通信设备及可读存储介质
WO2023198167A1 (zh) 模型微调的方法、装置及设备
US20240224082A1 (en) Parameter selection method, parameter configuration method, terminal, and network side device
WO2023186099A1 (zh) 信息反馈方法、装置及设备
WO2024093997A1 (zh) 确定模型适用性的方法、装置及通信设备
WO2022237822A1 (zh) 训练数据集获取方法、无线传输方法、装置及通信设备
WO2023179753A1 (zh) 波束信息指示方法、装置、终端及网络侧设备
WO2024067280A1 (zh) 更新ai模型参数的方法、装置及通信设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22906572

Country of ref document: EP

Kind code of ref document: A1