WO2023104169A1 - 一种无线网络中的人工智能ai模型训练方法及装置 - Google Patents

一种无线网络中的人工智能ai模型训练方法及装置 Download PDF

Info

Publication number
WO2023104169A1
WO2023104169A1 PCT/CN2022/137671 CN2022137671W WO2023104169A1 WO 2023104169 A1 WO2023104169 A1 WO 2023104169A1 CN 2022137671 W CN2022137671 W CN 2022137671W WO 2023104169 A1 WO2023104169 A1 WO 2023104169A1
Authority
WO
WIPO (PCT)
Prior art keywords
terminal
training
model
model training
node
Prior art date
Application number
PCT/CN2022/137671
Other languages
English (en)
French (fr)
Inventor
朱勇合
曾清海
曾宇
耿婷婷
Original Assignee
华为技术有限公司
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 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2023104169A1 publication Critical patent/WO2023104169A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W8/00—Network data management
    • H04W8/22—Processing or transfer of terminal data, e.g. status or physical capabilities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W8/00—Network data management
    • H04W8/22—Processing or transfer of terminal data, e.g. status or physical capabilities
    • H04W8/24—Transfer of terminal data

Definitions

  • the embodiments of the present application relate to the field of artificial intelligence (AI), and in particular to an AI model training method and device in a wireless network.
  • AI artificial intelligence
  • a wireless communication network such as a mobile communication network
  • services supported by the network are becoming more and more diverse, and therefore requirements to be met are becoming more and more diverse.
  • the network needs to be able to support ultra-high speed, ultra-low latency, and/or very large connections.
  • This feature makes network planning, network configuration, and/or resource scheduling increasingly complex.
  • These new requirements, new scenarios, and new features have brought unprecedented challenges to network planning, O&M, and efficient operations.
  • artificial intelligence technology can be introduced into the wireless communication network to realize network intelligence.
  • federated learning is a popular model training architecture, which can mobilize distributed data located on edge devices to participate in model training without compromising privacy.
  • central nodes and edge devices need to interact with parameters or gradients of artificial intelligence (AI) models. How to deploy federated learning in wireless networks is a problem worthy of research.
  • This application provides an AI model training method and device in a wireless network to solve the problem of time-frequency resources caused by the need to use mutually orthogonal uplink time-frequency resources for edge nodes participating in federated learning when deploying federated learning in a wireless network.
  • the occupancy rate is too high, and the time delay is large.
  • a method for training an artificial intelligence AI model in a wireless network where the execution body of the method is a second node, and may also be a component (processor, chip or others) configured in the second node, or It may be a software module, including: sending first configuration information to terminals participating in federated learning, and the first configuration information is at least used to configure: training duration, time-frequency resources, and reporting time; wherein, for different terminals participating in federated learning The configured training duration, time-frequency resources, and reporting time are all the same; receive the signal after the gradient reported by the terminal participating in the federated learning is superimposed in the air, and the gradient is reported by the terminal using the time-frequency resource at the reporting time The gradient of the AI model trained within the training duration.
  • the second node may also be called an access network device, and the first node may also be called a central node, and so on.
  • the access network device (which may be called the second node) or the central node (which may be called the first node) configures the same training duration, time-frequency resources and reporting time for the terminals participating in the federated learning.
  • the number of terminals participating in federated learning is n
  • the access network device or the central node allocates the same time-frequency resources to the n terminals.
  • n orthogonal time-frequency resources are allocated to n terminals participating in federated learning, which can reduce the overhead of time-frequency resources.
  • each terminal uses its own time-frequency resource to report the gradient of the AI model.
  • the access network device can receive n radio frequency signals, and process the above n radio frequency signals respectively to restore the gradient reported by each terminal, and the time delay is relatively large.
  • the access network device allocates one time-frequency resource to n terminals participating in federated learning, and the n terminals all report gradients on the time-frequency resource. Then, according to the superposition characteristics of the wireless channel, these n gradients will be superimposed together during air transmission.
  • the access network equipment can process the received signal on the above-mentioned time-frequency resources after receiving the above-mentioned signal, and restore the superimposed signal Y, and the terminal uses the superimposed signal Y Gradient aggregation can be performed.
  • the gradient aggregation process can be an arithmetic average process.
  • the above-mentioned superimposed signal Y can be divided by 3, and the result can be used as an aggregation gradient.
  • a superimposed signal Y can be obtained by processing one radio frequency signal, while in the existing scheme, it is necessary to sequentially process three radio frequency signals in different time slots to restore the corresponding gradients respectively, and then
  • adopting the solution of the present application can reduce the time-frequency resource occupation to a certain extent, and can also reduce the delay of gradient aggregation.
  • the training duration configured for the terminal by the second node may not be the actual training duration of the terminal, but refers to the upper limit of time required by the terminal for each round of training. That is, the terminal can complete the current round of model training within the training duration and report the training completion indication to the base station; otherwise, the terminal can terminate the current round of model training and wait for the arrival of the next training duration.
  • the purpose of setting the training duration is to ensure that different terminals can report the gradient of model training to the access network device at the same time.
  • the gradient of the model training reported by the terminal to the access network device is the gradient of the current round of model training, not the gradient of other rounds of model training, such as the gradient of the last round of model training.
  • the above-mentioned training duration may be a global parameter determined comprehensively by taking into account the computing power of each terminal participating in the federated learning and the complexity of the model.
  • the method further includes: receiving a training completion indication from the terminal, where the training completion indication is that the terminal completes the training of the AI model within the training duration and sends a message to the second node Sent; according to the training completion instruction sent by the terminal, count the number of terminals that have completed AI model training within the training duration.
  • it also includes: if the number of terminals that have completed the AI model training is greater than or equal to the terminal number threshold, then determine the average gradient of the current round of model training according to the gradients reported by different terminals participating in federated learning; otherwise , taking the average gradient of the previous round of model training as the average gradient of the current round of model training; updating the parameters of the AI model according to the average gradient of the current round of model training, and sending the current round of model training to the terminal average gradient.
  • the first node can determine the threshold of the number of terminals.
  • the first node may set a threshold for the number of terminals.
  • the average gradient of the current round of model training is calculated and sent to the terminal; otherwise, the average gradient of the previous round of model training is sent to the terminal, or it can also be described as the previous round
  • the average gradient of model training is sent to the terminal as the average gradient of the current round of model training; thus ensuring that the accuracy of the calculated average gradient of the current round of model training meets the requirements.
  • the method further includes: sending to the first node the number of terminals that have completed the training of the AI model within the training period, and a gradient superimposition signal reported by the terminals in the air.
  • the first configuration information is further used to configure at least one of the following: dedicated bearer RB resource, modulation mode, initial AI model, or transmit power.
  • the process of determining the transmission power includes: determining the uplink channel quality of the terminal according to the measurement of the sounding reference signal SRS from the terminal; transmit power of the terminal.
  • the second node determines the optimal transmission power by measuring the uplink channel, and configures the transmission power to the terminal to send the gradient of the current round of model training, thereby improving the accuracy of air calculation and further improving the accuracy of gradient aggregation .
  • the first configuration information is further used to configure at least one of the following: dedicated bearer RB resources, modulation mode, initial AI model, channel state information CSI interval, or channel inversion parameters.
  • it also includes: receiving second configuration information from the first node, where the second configuration information is used to configure at least one of the following: a list of terminals participating in federated learning, an initial AI model, and a group temporary identifier , training duration, terminal number threshold, transmission block size or uplink requirements.
  • the method further includes: receiving first terminal information from the terminal, and sending second terminal information to the first node; wherein, the first terminal information includes at least one of the following: terminal communication capabilities, computing capabilities of the terminal, or data set characteristics of the terminal; the second terminal information includes at least one of the following: communication capabilities of the terminal, computing capabilities of the terminal, data set characteristics of the terminal, or a temporary identifier of the terminal, the The terminal temporary identifier is allocated to the terminal by the second node.
  • the communication capability of the terminal includes, for example, the maximum transmit power that the terminal can support, the antenna configuration of the terminal, and the like.
  • the computing capability of the terminal includes, for example, performance of a central processing unit (central processing unit, CPU), performance of a graphics processing unit (graphics processing unit, GPU), storage space, power, and the like.
  • Data set characteristics of the terminal such as the size of the data set, the distribution of the data set, whether the data set is complete, and whether the labels of the data set are complete.
  • the dataset can be further divided by percentage into training, validation and test sets. For example, 60% of the dataset is training set, 20% is validation set, 20% is test set, etc.
  • the training set is used to train the AI model
  • the verification set is used to evaluate the trained AI model
  • the test set is used to test the trained AI model.
  • the terminal temporary identifier which may be a cell-radio network temporary identifier (C-RNTI) or other temporary identifiers.
  • it further includes: when the model training termination condition is met, sending a termination model training instruction to the terminal; or receiving a termination model training instruction from the first node, and forwarding the termination instruction to the terminal Model training instructions.
  • the second aspect provides an artificial intelligence AI model training method in a wireless network.
  • This method corresponds to the above first aspect.
  • the execution subject of this method is a terminal, and it can also be A component (processor, chip or other) configured in the terminal, or may be a software module, etc., including: receiving first configuration information from the second node, the first configuration information is at least used to configure: training duration, time Frequency resources, and reporting time; wherein, the training duration, time-frequency resources and reporting time configured for different terminals participating in federated learning are the same; within the training duration, the AI model is trained to obtain the current round of model training.
  • the gradient of the model at the reporting moment, report the gradient of the AI model of the current round of model training to the second node by using the time-frequency resource.
  • the method further includes: when the training duration ends, if the AI model finishes training, sending a training completion indication to the second node.
  • the method further includes: if the training of the AI model is not completed within the training period, terminating the training of the AI model.
  • it also includes: receiving the average gradient of the last round of model training from the second node; updating the gradient of the AI model in the current round of model training according to the average gradient of the last round of model training ; Or, according to the average gradient of the current round of model training, update the parameters and gradients of the AI model in the current round of model training.
  • the first configuration information is further used to configure at least one of the following: dedicated bearer RB resource, modulation mode, initial AI model or transmit power.
  • the first configuration information is further used to configure at least one of the following: dedicated bearer RB resources, modulation mode, initial AI model, channel state information CSI interval, or channel inversion parameters.
  • the method further includes: if the frequency resources configured for the downlink channel and the uplink channel are the same , then determine the CSI of the uplink channel of the terminal according to the measured CSI of the downlink channel; if the CSI of the uplink channel meets the requirements of the CSI interval, then determine the transmit power according to the channel inversion parameter;
  • the second node reporting the gradient of the AI model of the current round of model training includes: reporting the gradient of the AI model of the current round of model training to the second node based on the determined transmit power.
  • the first configuration information further includes a group temporary identifier, where the group temporary identifier is a group temporary identifier assigned to the terminal by the first node.
  • the method further includes: receiving a scheduling instruction from the second node, where the scheduling instruction includes a group temporary identifier; when the group temporary identifier included in the scheduling instruction is identical to the first node When the group temporary identifiers assigned to the terminals are the same, perform AI model training in the current round of model training; otherwise, do not perform AI model training in the current round of model training.
  • the method further includes: receiving a model training termination instruction from the second node; and terminating the training of the AI model according to the model training termination instruction.
  • the method further includes: sending first terminal information to the second node, where the first terminal information includes at least one of the following: communication capability of the terminal, computing capability of the terminal, or data of the terminal set features.
  • the third aspect provides an artificial intelligence AI model training method in a wireless network.
  • This method corresponds to the above first aspect.
  • the execution subject of this method is the first node, and it can also be It is a component (processor, chip or other) configured in the first node, or may be a software module, etc., including: determining second configuration information, and the second configuration information is used to configure at least one of the following: participate in federated learning terminal list, initial AI model, group temporary identifier, training duration, terminal number threshold, transmission block size, or uplink requirement; and send the second configuration information to the second node.
  • the method further includes: receiving second terminal information from the second node, where the second terminal information includes at least one of the following: communication capability of the terminal, computing capability of the terminal, data of the terminal Set features, or terminal temporary identifiers, where the terminal temporary identifiers are allocated to the terminals by the second node; and determine a list of terminals participating in federated learning according to the terminal information.
  • it also includes: receiving from the second node the signal of the gradient superimposed in the air reported by the terminals participating in the federated learning and the number of terminals that have completed AI model training within the training period; if the If the number of terminals that have completed model training within the training period is greater than or equal to the threshold number of terminals, the average gradient of the current round of model training is determined according to the gradient of the AI model reported by the terminals; otherwise, the average gradient of the previous round of model training is The gradient is used as the average gradient of the current round of model training; the parameters of the AI model are updated according to the average gradient of the current round of model training, and the average gradient of the current round of model training is sent to the second node, so that the The second node sends the average gradient of the current round of model training to the terminal.
  • the method further includes: sending a scheduling instruction to the second node, the scheduling instruction includes a group temporary identifier, and the scheduling instruction is used to schedule a terminal corresponding to the group temporary identifier.
  • the scheduling instruction includes a group temporary identifier
  • the scheduling instruction is used to schedule a terminal corresponding to the group temporary identifier.
  • it also includes: when the model training termination condition is met, sending a termination model training instruction to the second node, used to instruct the terminal to stop training the AI model in the current round of model training .
  • the device includes a one-to-one corresponding unit or module for performing the method/operation/step/action described in the first aspect, the second aspect or the third aspect, and the unit or module may be hardware
  • the circuit may also be software, or may be implemented by combining hardware circuits with software.
  • a communication device includes a processor and a memory.
  • the memory is used to store computer programs or instructions
  • the processor is coupled to the memory; when the processor executes the computer programs or instructions, the device is made to execute the method of the first aspect, the second aspect or the third aspect.
  • a device including a processor and an interface circuit, the processor is used to communicate with other devices through the interface circuit, and execute the method described in any one of the first aspect, the second aspect, or the third aspect. method.
  • the processor includes one or more.
  • an apparatus including a processor coupled to a memory, the processor is used to execute a program stored in the memory, so as to execute the method described in any one of the first aspect, the second aspect, or the third aspect. method.
  • the memory may be located within the device or external to the device. And there may be one or more processors.
  • a chip system including: a processor or a circuit, configured to execute the method described in any one of the first aspect, the second aspect, or the third aspect.
  • a computer-readable storage medium stores computer programs or instructions.
  • the device executes the above-mentioned first aspect, second aspect, or third-party methods.
  • a computer program product includes a computer program or an instruction, and when the computer program or instruction is executed by a device, the device executes the method of the first aspect, the second aspect, or the third aspect.
  • a system including a device for performing the method of the aforementioned first aspect and a device for performing the method of the aforementioned second aspect.
  • a device for performing the method in the third aspect above may also be included.
  • Figure 1 is a schematic diagram of the network architecture provided by the present application.
  • Fig. 2a and Fig. 2b are the schematic diagrams of the neural network provided by the present application.
  • Figure 2c is a schematic diagram of the AI model provided by the present application.
  • Figure 3 is a schematic diagram of the federated learning training provided by this application.
  • FIG. 4 and Figure 5 are a schematic flow chart provided by the present application.
  • FIG. 6 is a schematic diagram of channel measurement provided by the present application.
  • FIG. 7 and Figure 8 are another schematic flow diagram provided by the present application.
  • FIG. 9 and Figure 10 are schematic diagrams of the device provided in this application.
  • FIG. 1 is a schematic structural diagram of a communication system 1000 to which the present application can be applied.
  • the communication system includes a radio access network 100 and a core network 200 , and optionally, the communication system 1000 may also include the Internet 300 .
  • the radio access network 100 may include at least one access network device (such as 110a and 110b in Figure 1), and may also include at least one terminal (such as 120a-120j in Figure 1). The terminal is connected to the access network device in a wireless manner, and the access network device is connected to the core network in a wireless or wired manner.
  • the core network device and the access network device can be independent and different physical devices, or the functions of the core network device and the logical functions of the access network device can be integrated on the same physical device, or they can be integrated on one physical device Part of the functions of the core network device and part of the functions of the access network device are specified. Terminals and access network devices may be connected to each other in a wired or wireless manner.
  • FIG. 1 is only a schematic diagram.
  • the communication system may also include other network devices, such as wireless relay devices and wireless backhaul devices, which are not shown in FIG. 1 .
  • the access network equipment can be a base station (base station), an evolved base station (evolved NodeB, eNodeB), a transmission reception point (transmission reception point, TRP), and a next-generation base station in the fifth generation (5th generation, 5G) mobile communication system (next generation NodeB, gNB), access network equipment in the open radio access network (open radio access network, O-RAN), next-generation base stations in the sixth generation (6th generation, 6G) mobile communication system, future mobile
  • DU distributed unit
  • CU control plane centralized unit control plane
  • CU-CP centralized unit control plane
  • CU user plane centralized unit user plane
  • the access network device may be a macro base station (such as 110a in Figure 1), a micro base station or an indoor station (such as 110b in Figure 1), or a relay node or a donor node.
  • a macro base station such as 110a in Figure 1
  • a micro base station such as 110b in Figure 1
  • a relay node or a donor node.
  • the device used to realize the function of the access network device may be the access network device; it may also be a device capable of supporting the access network device to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware A circuit plus a software module, the device can be installed in the access network equipment or can be matched with the access network equipment for use.
  • a system-on-a-chip may consist of a chip, or may include a chip and other discrete devices.
  • the technical solutions provided by the present application are described below by taking the apparatus for realizing the functions of the access network equipment as the access network equipment and the access network equipment as the base station as an example.
  • the protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure.
  • the control plane protocol layer structure may include a radio resource control (radio resource control, RRC) layer, a packet data convergence protocol (packet data convergence protocol, PDCP) layer, a radio link control (radio link control, RLC) layer, a media The access control (media access control, MAC) layer and the function of the protocol layer such as the physical layer.
  • the user plane protocol layer structure may include the functions of the PDCP layer, the RLC layer, the MAC layer, and the physical layer.
  • the PDCP layer may also include a service data adaptation protocol (service data adaptation protocol). protocol, SDAP) layer.
  • the protocol layer structure between the access network device and the terminal may also include an artificial intelligence (AI) layer, which is used to transmit data related to the AI function.
  • AI artificial intelligence
  • Access devices may include CUs and DUs. Multiple DUs can be centrally controlled by one CU.
  • the interface between the CU and the DU may be referred to as an F1 interface.
  • the control plane (control panel, CP) interface may be F1-C
  • the user plane (user panel, UP) interface may be F1-U.
  • the present application does not limit the specific names of the interfaces.
  • CU and DU can be divided according to the protocol layer of the wireless network: for example, the functions of the PDCP layer and above protocol layers are set in the CU, and the functions of the protocol layers below the PDCP layer (such as RLC layer and MAC layer, etc.) are set in the DU; another example, PDCP The functions of the protocol layer above the layer are set in the CU, and the functions of the PDCP layer and the protocol layer below are set in the DU, without restriction.
  • the CU or DU may be divided into functions having more protocol layers, and for example, the CU or DU may also be divided into part processing functions having protocol layers.
  • part of the functions of the RLC layer and the functions of the protocol layers above the RLC layer are set in the CU, and the rest of the functions of the RLC layer and the functions of the protocol layers below the RLC layer are set in the DU.
  • the functions of the CU or DU can also be divided according to the business type or other system requirements, for example, according to the delay, and the functions whose processing time needs to meet the delay requirement are set in the DU, which does not need to meet the delay
  • the required feature set is in the CU.
  • the CU may also have one or more functions of the core network.
  • the CU can be set on the network side to facilitate centralized management.
  • the wireless unit (radio unit, RU) of the DU is set remotely.
  • the RU may have a radio frequency function.
  • DUs and RUs can be divided in a physical layer (physical layer, PHY).
  • the DU can implement high-level functions in the PHY layer
  • the RU can implement low-level functions in the PHY layer.
  • the functions of the PHY layer may include at least one of the following: adding a cyclic redundancy check (cyclic redundancy check, CRC) code, channel coding, rate matching, scrambling, modulation, layer mapping, precoding, Resource mapping, physical antenna mapping, or radio frequency transmission functions.
  • CRC cyclic redundancy check
  • the functions of the PHY layer may include at least one of the following: CRC check, channel decoding, de-rate matching, descrambling, demodulation, de-layer mapping, channel detection, resource de-mapping, physical antenna de-mapping, or RF receiving function.
  • the high-level functions in the PHY layer may include a part of the functions of the PHY layer, for example, this part of the functions is closer to the MAC layer, and the lower-level functions in the PHY layer may include another part of the functions of the PHY layer, for example, this part of the functions is closer to the radio frequency function.
  • high-level functions in the PHY layer may include adding CRC codes, channel coding, rate matching, scrambling, modulation, and layer mapping
  • low-level functions in the PHY layer may include precoding, resource mapping, physical antenna mapping, and radio transmission functions
  • high-level functions in the PHY layer may include adding CRC codes, channel coding, rate matching, scrambling, modulation, layer mapping, and precoding
  • low-level functions in the PHY layer may include resource mapping, physical antenna mapping, and radio frequency send function.
  • the high-level functions in the PHY layer may include CRC check, channel decoding, de-rate matching, decoding, demodulation, and de-layer mapping
  • the low-level functions in the PHY layer may include channel detection, resource de-mapping, physical antenna de-mapping, and RF receiving functions
  • the high-level functions in the PHY layer may include CRC check, channel decoding, de-rate matching, decoding, demodulation, de-layer mapping, and channel detection
  • the low-level functions in the PHY layer may include resource de-mapping , physical antenna demapping, and RF receiving functions.
  • the function of the CU may be implemented by one entity, or may also be implemented by different entities.
  • the functions of the CU can be further divided, that is, the control plane and the user plane are separated and realized by different entities, namely, the control plane CU entity (ie, the CU-CP entity) and the user plane CU entity (ie, the CU-UP entity) .
  • the CU-CP entity and CU-UP entity can be coupled with the DU to jointly complete the functions of the access network equipment.
  • any one of the foregoing DU, CU, CU-CP, CU-UP, and RU may be a software module, a hardware structure, or a software module+hardware structure, without limitation.
  • the existence forms of different entities may be different, which is not limited.
  • DU, CU, CU-CP, and CU-UP are software modules
  • RU is a hardware structure.
  • the access network device includes CU-CP, CU-UP, DU and RU.
  • the execution subject of this application includes DU, or includes DU and RU, or includes CU-CP, DU and RU, or includes CU-UP, DU and RU, without limitation.
  • the methods executed by each module are also within the protection scope of the present application.
  • a terminal may also be called terminal equipment, user equipment (user equipment, UE), mobile station, mobile terminal, and so on.
  • the terminal can be widely used in communication in various scenarios, including but not limited to at least one of the following scenarios: device-to-device (device-to-device, D2D), vehicle-to-everything (V2X), machine-type communication (machine -type communication, MTC), Internet of Things (IOT), virtual reality, augmented reality, industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, or smart city wait.
  • device-to-device device-to-device, D2D
  • V2X vehicle-to-everything
  • machine-type communication machine -type communication
  • MTC machine -type communication
  • IOT Internet of Things
  • virtual reality augmented reality
  • industrial control automatic driving
  • telemedicine smart grid
  • smart furniture smart office
  • smart wear smart transportation
  • smart city wait smart city wait.
  • the terminal can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a wearable device, a vehicle, a drone, a helicopter, an airplane, a ship, a robot, a mechanical arm, or a smart home device, etc.
  • This application does not limit the specific technology and specific equipment form adopted by the terminal.
  • the device used to realize the function of the terminal may be a terminal; it may also be a device capable of supporting the terminal to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module. Installs in the terminal or can be used with the terminal.
  • a terminal as an example of an apparatus for realizing functions of a terminal.
  • Base stations and terminals can be fixed or mobile.
  • Base stations and/or terminals can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on water; they can also be deployed on aircraft, balloons and artificial satellites in the sky.
  • This application does not limit the application scenarios of the base station and the terminal.
  • the base station and the terminal can be deployed in the same scene or in different scenes. For example, the base station and the terminal are deployed on land at the same time; or, the base station is deployed on land, and the terminal is deployed on water, etc., and no more examples are given here.
  • the roles of the base station and the terminal can be relative.
  • the helicopter or UAV 120i in FIG. base station for base station 110a, 120i is a terminal, that is, communication between 110a and 120i is performed through a wireless air interface protocol. Communication between 110a and 120i may also be performed through an interface protocol between base stations.
  • relative to 110a, 120i is also a base station. Therefore, both the base station and the terminal can be collectively referred to as a communication device, 110a and 110b in FIG. 1 can be referred to as a communication device with a base station function, and 120a-120j in FIG. 1 can be referred to as a communication device with a terminal function.
  • the communication between the base station and the terminal, between the base station and the base station, and between the terminal and the terminal can be carried out through the licensed spectrum, the communication can also be carried out through the unlicensed spectrum, and the communication can also be carried out through the licensed spectrum and the unlicensed spectrum at the same time; Communications may be performed on frequency spectrums below megahertz (gigahertz, GHz), or communications may be performed on frequency spectrums above 6 GHz, or communications may be performed using both frequency spectrums below 6 GHz and frequency spectrums above 6 GHz. This application does not limit the frequency spectrum resources used by wireless communication.
  • the base station sends a downlink signal or downlink information to the terminal, and the downlink information is carried on the downlink channel;
  • the terminal sends an uplink signal or uplink information to the base station, and the uplink information is carried on the uplink channel.
  • the terminal can establish a wireless connection with the cell controlled by the base station.
  • a cell with which a terminal has established a wireless connection is called a serving cell of the terminal.
  • the terminal communicates with the serving cell, it may be interfered by signals from neighboring cells.
  • an independent network element such as a central node, AI network element, or AI node, etc.
  • the central node can communicate with the communication system
  • the access network devices in the network are directly connected, or can be indirectly connected through a third-party network element and the access network device, wherein the third-party network element can be an authentication management function (authentication management function, AMF) network element, or a user core network elements such as user plane function (UPF) network elements; or, AI functions, AI modules or AI entities can be configured in other network elements in the communication system to implement AI-related operations, such as other network elements
  • AMF authentication management function
  • UPF user plane function
  • AI functions, AI modules or AI entities can be configured in other network elements in the communication system to implement AI-related operations, such as other network elements
  • It can be an access network device (such as gNB), a core network device, or a network management (operation, administration and maintenance, OAM), etc.
  • the network element that performs AI-related operations is a network element with a built-in AI function.
  • OAM is used to operate, manage and/or maintain core network equipment, and/or is used to operate, manage and/or maintain access network equipment.
  • the AI model is a specific method for realizing the AI function, and the AI model represents the mapping relationship between the input and output of the model.
  • AI models can be neural networks or other machine learning models.
  • the AI model may be referred to as a model for short.
  • AI-related operations may include at least one of the following: data collection, model training, model information release, model inference (model reasoning), or release of reasoning results.
  • neural network is a specific implementation form of machine learning technology.
  • the neural network can theoretically approximate any continuous function, so that the neural network has the ability to learn any mapping.
  • Traditional communication systems need to rely on rich expert knowledge to design communication modules, while deep learning communication systems based on neural networks can automatically discover hidden pattern structures from a large number of data sets, establish mapping relationships between data, and achieve better results than traditional communication systems. The performance of the modeling method.
  • a neural network is derived from the neuronal structure of brain tissue.
  • Each neuron performs a weighted sum operation on its input values, and passes the weighted sum result through an activation function to generate an output.
  • Fig. 2a it is a schematic diagram of neuron structure.
  • the bias of the weighted sum for b the form of the activation function can be diversified.
  • the output of the neuron is:
  • the output of the neuron is: b can be various possible values such as decimals, integers (including 0, positive integers or negative integers, etc.), or complex numbers.
  • the activation functions of different neurons in a neural network can be the same or different.
  • a neural network generally includes a multi-layer structure, and each layer may include one or more neurons. Increasing the depth and/or width of a neural network can improve the expressive power of the neural network, providing more powerful information extraction and abstract modeling capabilities for complex systems.
  • the depth of the neural network may refer to the number of layers included in the neural network, and the number of neurons included in each layer may be referred to as the width of the layer.
  • Figure 2b it is a schematic diagram of the layer relationship of the neural network.
  • a neural network includes an input layer and an output layer. The input layer of the neural network processes the input received by neurons, and then passes the result to the output layer, and the output layer obtains the output result of the neural network.
  • a neural network in another implementation, includes an input layer, a hidden layer, and an output layer.
  • the input layer of the neural network processes the input received by neurons, and then passes the result to the hidden layer in the middle, and the hidden layer then passes the calculation result to the output layer or the adjacent hidden layer, and finally the output layer obtains the result of the neural network. Output the result.
  • a neural network may include one or more sequentially connected hidden layers without limitation.
  • a loss function can be defined.
  • the loss function describes the gap or difference between the output value of the neural network and the ideal target value, and the application does not limit the specific form of the loss function.
  • the training process of the neural network is to make the value of the loss function less than the threshold gate Limits or the process of meeting target requirements.
  • Figure 2c is a schematic diagram of an application framework of AI.
  • Data source data source
  • the model training host model training host
  • the model training host obtains the AI model by analyzing or training the training data provided by the data source, and deploys the AI model in the model inference host (model inference host).
  • the model training node can also update the AI model deployed on the model inference node.
  • the model inference node can also feed back information about the deployed model to the model training node, so that the model training node can optimize or update the deployed AI model.
  • the AI model represents the mapping relationship between the input and output of the model. Learning the AI model through the model training node is equivalent to using the training data to learn the mapping relationship between the input and output of the model.
  • the model inference node uses the AI model to perform inference based on the inference data provided by the data source, and obtains the inference result. This method can also be described as: the model inference node inputs the inference data into the AI model, and obtains an output through the AI model, and the output is the inference result.
  • the inference result may indicate: configuration parameters used (executed) by the execution object, and/or operations performed by the execution object.
  • the reasoning result can be uniformly planned by the execution (actor) entity, and sent to one or more execution objects (for example, network entities) for execution.
  • Federated learning is a popular AI/ML model training framework that can effectively help multiple organizations in data usage and machine learning modeling while meeting the requirements of user privacy protection, data security, and government regulations.
  • federated learning can effectively solve the problem of data islands, allowing participants to jointly model without sharing data, technically breaking data islands, and realizing AI collaboration.
  • Federated learning includes central nodes and edge nodes.
  • Central nodes such as servers and base stations
  • edge devices such as smartphones and sensors
  • Federated learning is mainly divided into the following three categories: horizontal federated learning, vertical federated learning, and federated transfer learning.
  • This application mainly involves the process of horizontal federated learning.
  • Figure 3 shows the training process of horizontal federated learning.
  • horizontal federated learning consists of a central node and multiple edge nodes. Among them, the original data are distributed in each edge node, the central node does not have the original data, and the edge node is not allowed to send the original data to the central node.
  • the central node In the training process of federated learning, the central node first sends the initialized AI model (which can be called the initial AI model) to each edge node, and then starts iterative training.
  • the initialized AI model which can be called the initial AI model
  • Each iterative training process is:
  • Edge nodes use local data to train the initial AI model and obtain the gradient of the trained AI model
  • Each edge node reports its own trained gradient to the central node
  • the central node After receiving the gradients reported by each edge node, the central node aggregates the gradients and updates the parameters of the AI model according to the aggregated gradients;
  • the central node sends the aggregated gradients to each edge node participating in the training, and the edge nodes update the parameters and gradients of the locally trained AI model according to the aggregated gradients delivered by the central node.
  • the central node calculates the loss function of the AI model after updating the parameters; if the loss function meets the conditions, the model training is terminated; if the loss function does not meet the conditions, repeat the above steps 2-4.
  • the present application provides an AI model training method in a wireless network.
  • terminals participating in federated learning can be allocated the same time-frequency resources and reporting time.
  • the terminals participating in the federated learning use the same time-frequency resources to report the gradient of the trained AI model at the same reporting time, so as to solve the above-mentioned time-frequency problems caused by the allocation of mutually orthogonal multiple time-frequency resources for the terminals participating in the federated learning. Problems such as high frequency resource overhead and high delay.
  • a process of an AI model training method in a wireless network is provided, the process at least includes the following steps:
  • Step 401 The base station sends first configuration information to terminals participating in federated learning, where the first configuration information is used to configure at least one of the following: training duration, time-frequency resources, or reporting time.
  • the terminal receives the first configuration information from the base station.
  • Step 402 The terminal trains the AI model within the aforementioned training duration, and obtains the gradient of the AI model of the current round of model training.
  • Step 403 The terminal reports the gradient of the AI model of the current round of model training to the base station by using the above time-frequency resources at the above reporting time.
  • the base station receives the signal after the gradient reported by the terminal is superimposed in the air.
  • the base station or central node configures the same training duration, time-frequency resources, and reporting time for terminals participating in federated learning.
  • the number of terminals participating in federated learning is n
  • the base station or central node allocates the same time-frequency resources to the n terminals.
  • n orthogonal time-frequency resources are allocated to n terminals participating in federated learning, which can reduce the overhead of time-frequency resources.
  • each terminal uses its own time-frequency resource to report the gradient of the AI model.
  • the base station can receive n radio frequency signals, and process the above n radio frequency signals respectively to recover the gradient reported by each terminal, and the time delay is relatively large.
  • the base station allocates one time-frequency resource to n terminals participating in federated learning, and the n terminals all report gradients on the time-frequency resource. Then, according to the superposition characteristics of the wireless channel, these n gradients will be superimposed together during air transmission. For example, if the value of n is 3, the gradient reported by terminal 1 is Y1, the gradient reported by terminal 2 is Y2, and the gradient reported by terminal 3 is Y3, then the above three gradients are transmitted on the same time-frequency resource.
  • the base station can perform signal processing on the received signal after receiving the above-mentioned signal on the above-mentioned time-frequency resources, and recover the superimposed signal Y, and the terminal uses the superimposed signal Y to perform gradient aggregation That is, the process of gradient aggregation may be an arithmetic mean process, for example, the above-mentioned superimposed signal Y may be divided by 3, and the result may be used as an aggregated gradient.
  • a superimposed signal Y can be obtained by processing one radio frequency signal.
  • adopting the solution of the present application can reduce the time-frequency resource occupation to a certain extent, and can also reduce the delay of gradient aggregation.
  • the above-mentioned training duration configured by the base station (or other devices in the access network device) to the terminal may not be the actual training duration of the terminal, but generally refers to the upper limit of the training time required by the terminal for each round of training. That is, the terminal can complete the current round of model training within the training duration and report the training completion indication to the base station; otherwise, the terminal can terminate the current round of model training and wait for the arrival of the next training duration.
  • the base station sets the training duration to ensure that different terminals can report the gradient of model training to the base station at the same time.
  • the gradient of the model training reported by the terminal to the base station is the gradient of the current round of model training, rather than the gradient of other rounds of model training, such as the gradient of the previous round of model training.
  • the above-mentioned training duration may be a global parameter determined comprehensively by taking into account the computing power of each terminal participating in the federated learning and the complexity of the model, and then configured by the base station for each terminal.
  • the number of terminals participating in federated learning can be n, and each terminal can be considered as an edge node, and the central node can be a base station, or an OAM, or be located independently as a module in the core network, etc., without limitation .
  • the central node is independent from the base station, that is, the central node and the base station are two devices as an example.
  • a flow of a method for training an AI model in a wireless network is provided, and in the flow, at least includes:
  • Step 501 Each of the n terminals reports terminal information to the base station, and the base station collects and aggregates the terminal information and reports it to the central node.
  • the terminal information includes at least one of the following:
  • the communication capability of the terminal including, for example, the maximum transmit power that the terminal can support, the antenna configuration of the terminal, and the like.
  • the computing capability of the terminal including, for example, central processing unit (central processing unit, CPU) performance, graphics processing unit (graphics processing unit, GPU) performance, storage space, and power consumption.
  • central processing unit central processing unit, CPU
  • graphics processing unit graphics processing unit, GPU
  • the characteristics of the data set of the terminal such as the size of the data set, the distribution of the data set, whether the data set is complete, and whether the labels of the data set are complete.
  • the dataset can be further divided by percentage into training, validation and test sets. For example, 60% of the dataset is training set, 20% is validation set, 20% is test set, etc. It can be understood that the training set is used to train the AI model, the verification set is used to evaluate the trained AI model, and the test set is used to test the trained AI model.
  • the base station assigns a terminal temporary identifier to the terminal, and the identifier may be a cell-radio network temporary identifier (cell-radio network temporary identifier, C-RNTI) or other temporary identifiers.
  • C-RNTI cell-radio network temporary identifier
  • other temporary identifiers can be distinguished by encoding, for example, using the serial numbers shown in Table 1 or Table 2 below:
  • the method may further include: the base station sends an instruction to report terminal information to n terminals. According to the instruction, the n terminals respectively report their own terminal information in the above step 501 .
  • Step 502 the central node sends second configuration information to the base station.
  • the second configuration information is used to configure at least one of the following: a list of terminals participating in federated learning, an initial AI model, training duration, a threshold for the number of terminals, a transmission block size, or uplink requirements, and the like.
  • the uplink requirement may include a rate, a bit error rate, or a time delay during uplink transmission of the terminal.
  • the transmission block refers to a data block including a MAC protocol data unit (protocol data unit, PDU), and this data block will be transmitted in a transmission time interval (transmission time interval, TTI).
  • PDU transmission time interval
  • the base station when the base station receives the terminal information reported by the terminal, it may be referred to as first terminal information.
  • the base station allocates a temporary identifier for the terminal, and adds the temporary identifier to the terminal information to form the second terminal information.
  • the second terminal information may also include at least one of the following: communication capability of the terminal, computing capability of the terminal, or data set characteristics of the terminal.
  • the base station reports the second terminal information to the central node.
  • the central node may determine a list of terminals participating in federated learning according to the second terminal information reported by the base station.
  • the second terminal information includes terminal communication capabilities, computing capabilities, characteristics of data sets, temporary identifiers, and the like.
  • the central node formulates a list of terminals participating in federated learning by comprehensively considering terminal communication capabilities, computing capabilities, and characteristics of data sets. For example, when considering comprehensively, priorities can be set for terminal communication capabilities, computing capabilities, and characteristics of data sets. Data set characteristics have priority over terminal communication capabilities, and terminal communication capabilities have priority over terminal computing capabilities.
  • corresponding thresholds are set for terminal communication capabilities, computing capabilities, and data set characteristics, and those that do not meet the threshold conditions will not be considered into the list of terminals participating in federated learning.
  • the central node can use a terminal whose communication capability is greater than or equal to the communication capability threshold, whose computing capability is greater than or equal to the computing capability threshold, and whose data set characteristics meet the requirements of the data set characteristics, as a terminal participating in federated learning.
  • the central node can configure the training time for the terminals participating in the federated learning.
  • the training time needs to consider the computational complexity of the AI model to be trained and the computing power of each terminal, and try to ensure that the terminals participating in the federated learning can be within the training time. Complete the local model training, but it should not be set too long, so as not to affect the overall efficiency of AI model training.
  • the central node can determine the threshold of the number of terminals.
  • the central node can set the threshold of the number of terminals.
  • the average gradient of the current round of model training is calculated and sent to the terminal; otherwise, the average gradient of the previous round of model training is sent to the terminal, or it can also be described as the previous round
  • the average gradient of model training is sent to the terminal as the average gradient of the current round of model training.
  • Step 503 the base station sends the first configuration information to the terminal.
  • the first configuration information is used to configure at least one of the following: training duration, reporting time, time-frequency resource, dedicated bearer radio bearer (radio bear, RB) resource, modulation mode, or initial AI model, etc. .
  • the base station may determine time-frequency resources, dedicated bearer RB resources, modulation schemes, etc. according to the uplink requirements.
  • the base station allocates the same time-frequency resources, reporting time, and training duration to n terminals participating in federated learning.
  • the base station needs to allocate dedicated bearer RB resources to the terminal, and the dedicated bearer RB resources may be signaling radio bearer (signal radio bear, SRB) resources, or data radio bearer (data radio bear, DRB) resources, etc. Gradients used to transfer AI models.
  • the dedicated bearer RB resource may only be used for gradient transmission, but not for other data transmission.
  • the modulation method configured by the base station for the terminal can be phase shift keying (pase sift kying, PSK), quadrature amplitude modulation (quadrature amplitude modulation, QAM) or other modulation methods, etc.
  • the specific order of PSK or QAM modulation can be based on the uplink Requirements, uplink channel quality, communication capabilities of the base station, and communication capabilities of the terminal are determined.
  • Step 504 The terminal performs AI model training within the training period.
  • the terminal when receiving the above-mentioned first configuration information, the terminal may train the AI model. It can be understood that, in the first round of training process, the terminal specifically trains the initial AI model, and the initial AI model is configured to the terminal by the central node. In the subsequent training process, the terminal specifically trains the AI model after the previous training.
  • the training duration in this application can be represented by T. For the terminals participating in the federated learning, if the training of the AI model is completed within the above training time T, a training completion indication is reported to the base station. If the training of the AI model is not completed within the above training duration T, the model training is terminated.
  • Step 505 The base station counts the number of terminals that have completed AI model training within the training duration T according to the training completion instructions reported by the terminals, and measures the uplink channel quality of terminals that have completed the current round of model training.
  • the base station will no longer measure the uplink channel quality of the corresponding terminal, and correspondingly, the terminal that has not completed model training will no longer report to the base station that the current round of model training is in progress The gradient of the AI model.
  • the base station may set a counter to count the terminals reporting the training completion indication at the beginning of each round of training, and reset the counter to 0 at the end of each round of training.
  • the base station measures the uplink channel quality of the terminal; otherwise, the base station triggers the terminal to perform the next round of training.
  • the process of the base station measuring the uplink channel quality of the terminal includes: the base station receives a sounding reference signal (sounding reference signal, SRS) from the terminal; the base station measures the SRS to determine the uplink channel quality of the terminal channel quality.
  • SRS sounding reference signal
  • Step 506 The base station determines the transmit power of the corresponding terminal according to the uplink channel quality of the terminal; the base station sends third configuration information to the terminal, the third configuration information is used to configure the transmit power of the terminal, and report time, etc.
  • the third configuration information and the above-mentioned first configuration information may be collectively referred to as a piece of configuration information.
  • the base station may comprehensively consider the error requirements of over-the-air calculations, the uplink channel quality of each terminal, the maximum transmit power supported by each terminal, and the total power to determine the transmit power of the terminal. Since over-the-air computing requires the synchronization of each node, too much synchronization error will affect the accuracy of over-the-air computing. Therefore, the base station can configure the same reporting time for the n terminals participating in the federated learning. In addition, the channel is always changing, so there is a validity period for the channel quality measurement. Therefore, each terminal needs to report the gradient of the AI model of the current round of model training at the same time at the reporting time, and the reporting time should be within the validity period of the channel quality measurement.
  • the validity period of the channel quality measurement result may be exceeded, resulting in a mismatch between power allocation and uplink channel quality.
  • the base station can predict the uplink channel quality at the current reporting time according to the historical uplink channel quality of the terminal, and optimize the optimal transmit power in advance. The specific prediction method is shown in Figure 6.
  • the base station collects the historical uplink channel state quality of each terminal between t1 and t2 , and starts to predict the uplink channel quality of each terminal at t3 at time t2 based on the historical uplink channel quality of each terminal, and determines The optimal power allocation scheme of each terminal at time t3 , and the optimal power allocation scheme is sent to each terminal at time t3 , and each terminal immediately reports the gradient of the current round of model training after receiving the power allocation scheme.
  • Step 507 The terminal reports the gradient of the current round of model training, the base station calculates the average gradient of the current round of model training, the base station updates the parameters of the AI model according to the average gradient of the current round of model training, and sends the average gradient of the current round of model training to participants n terminals for federated learning.
  • n terminals participating in federated learning can report the gradient of the current round of model training to the base station if the model training is completed within the training duration T at the reporting time.
  • the base station may calculate the average gradient of the current round of model training according to the gradients reported by the terminals. For example, since the above n terminals report the gradients of the current round of model training on the same time-frequency resource, the signal received by the base station is a signal obtained by superimposing the above n gradients in the air.
  • the base station can determine that the average gradient of the current round of model training is equal to Y/n.
  • the gradient of the current round of model training is no longer reported to the base station.
  • each terminal should try to start reporting the gradient of the current round of model training when the reporting time arrives, so as to reduce the reporting time error of each terminal as much as possible.
  • the base station needs to process the received air computing signal to restore the average gradient of the current round of model training. For example, if the counter in the base station is greater than or equal to the threshold of the number of terminals, the base station sends the locally calculated average gradient of the current round of model training to each terminal. Otherwise, the average gradient of the previous round of model training is sent to each terminal as the average gradient of the current round of model training.
  • the base station may send the updated model parameters to each terminal.
  • step 504 to step 507 is a cyclic process.
  • the terminal can update the parameters and gradients of the AI model according to the average gradient of the current round of model training, and within the training duration T, if the model training is completed, the terminal will send a message to the base station Report training completion instructions.
  • the parameters of the AI model can be updated in the current round of model training according to the average gradient of the previous round of model training, and then the gradient of the AI model can be updated according to the updated parameters, that is, gradient update.
  • Step 508 the base station judges the model training termination condition, and sends a model training termination instruction to the terminal.
  • the base station may determine the model training termination condition, and send a model training termination instruction to each terminal.
  • the model training termination condition may be at least one of model parameter convergence, reaching the maximum number of model training requirements, or meeting the maximum model training time requirement.
  • the base station Before gradient reporting, the base station first measures the uplink channel quality of all participating terminals and optimizes the transmit power of each terminal, which can improve the performance of air computing and further improve the training effect of federated learning.
  • this application provides a flow of an AI model training method in a wireless network.
  • the main difference between this flow and the flow shown in Figure 5 above is that in this flow, the terminal itself determines and reports the current round of model training
  • the transmit power of the gradient, the transmit power is no longer configured by the base station, at least including the following steps:
  • Step 701 n terminals report terminal information to the base station, and the base station uniformly reports the terminal information to the central node.
  • the terminal when receiving the terminal information of the terminal, the terminal may assign a temporary identifier to the terminal, add the temporary identifier to the terminal information, and report to the central node. See the description in Figure 5 above.
  • Step 702 The central node sends the second configuration information to the base station, and the second configuration information is used to configure at least one of the following items: a list of terminals participating in federated learning, an initial AI model, a training duration, a threshold for the number of terminals, and channel state information (channel state information, CSI) interval, channel inversion parameter, transport block size or uplink requirement.
  • a list of terminals participating in federated learning an initial AI model
  • a training duration a threshold for the number of terminals
  • CSI channel state information
  • channel inversion parameter transport block size or uplink requirement.
  • CSI refers to channel state information, which includes signal-to-noise ratio, Doppler frequency shift, and multipath delay spread.
  • the CSI interval includes a signal-to-noise ratio interval, a maximum Doppler frequency shift interval, and a maximum delay extension interval.
  • the SNR interval is [ ⁇ min , ⁇ max ], ⁇ min and ⁇ max are the lower limit and upper limit of SNR respectively;
  • the maximum Doppler frequency shift interval is [f min , f max ], f min and f max is the lower limit and upper limit of the maximum Doppler frequency shift;
  • the maximum time delay spread interval is [ ⁇ min , ⁇ max ], and ⁇ min and ⁇ max are the lower limit and upper limit of the maximum time delay spread respectively.
  • the gradient of the current round of model training can be reported to the base station only when the downlink CSI of the terminal satisfies the CSI interval, otherwise, it is not reported.
  • the SNR, maximum Doppler shift and maximum delay spread measured by a certain terminal are ⁇ 1 , f 1 and ⁇ 1 respectively, if ⁇ min ⁇ 1 ⁇ max , f min ⁇ f 1 ⁇ f max and ⁇ min ⁇ 1 ⁇ max , then the terminal can report the gradient of the current round of model training to the base station at the reporting time, otherwise it does not report the gradient of the current round of model training.
  • the channel inversion parameter ⁇ is a parameter used for power control, assuming that the maximum power and channel gain of the kth terminal are P k and h k respectively, and P 1
  • 2 , then ⁇ P 1
  • Step 703 The base station determines time-frequency resources, dedicated bearer RB resources, scheduling methods, etc. according to uplink requirements.
  • the base station sends the first configuration information to the terminal, and the first configuration information is used to configure at least one of the following: training duration, reporting time, time-frequency resource, initial AI model, dedicated bearer RB resource or modulation mode, etc.
  • the dedicated bearer RB resources include dedicated SRB resources, and/or dedicated DRB resources and the like.
  • the base station can determine time-frequency resources, SRB/DRB resources, or modulation schemes, etc. according to uplink requirements.
  • each terminal needs to use the same time-frequency resources.
  • the base station may allocate the same multiple time-frequency resources to n terminals, and use time diversity or frequency diversity to improve the performance of over-the-air computing. It should be noted that when the base station configures the same multiple time-frequency resources for n terminals, at a certain reporting time, the n terminals use the multiple time-frequency resources to report the gradient of the current round of model training to the base station.
  • the base station configures three time-frequency resources for n terminals, then at the time of reporting, the n terminals simultaneously use the first time-frequency resource, the second time-frequency resource, and the third time-frequency resource among the above-mentioned three time-frequency resources.
  • the frequency resource reports the gradient of the current round of model training. That is to say, at the reporting moment, when n terminals report the gradient of the current round of model training, the specific time-frequency resources used are the same.
  • the base station can allocate independent SRB/DRB resources to the terminal for transmitting the gradient of the current round of model training.
  • the base station can configure modulation methods such as PSK or QAM for the terminal.
  • the specific order of PSK or QAM modulation can be determined according to at least one of uplink requirements, uplink channel quality, communication capabilities of the base station, or communication capabilities of the terminal.
  • Step 704 The terminal trains the model locally.
  • the terminal may start to perform model training when receiving the first configuration information. If the terminal completes the model training within the training duration T, it reports a training completion indication to the base station. If the terminal does not complete the model training within the training duration T, the model training is terminated.
  • Step 705 the terminal measures the CSI of the downlink channel, and determines the transmit power of the terminal.
  • the uplink channel and downlink channel of the terminal can be configured with the same frequency resource, and according to channel reciprocity, the terminal can measure the CSI of the downlink channel and obtain the CSI of the uplink channel. The terminal can determine whether the acquired CSI of the uplink channel is in the CSI interval. If it is in the CSI interval, the transmission power is determined according to the channel inversion parameter.
  • the channel inversion parameter ⁇ P 1
  • 2 the transmit power of the kth terminal when reporting the gradient should be
  • the parameter h k can be obtained through CSI, combined with the channel inversion parameter ⁇ , the transmission power p k of the kth terminal can be determined, and the k is greater than or equal to 1, less than or equal to A positive integer of n.
  • Step 706 The terminal reports the gradient of the current round of model training, the base station calculates the average gradient of the current round of model training, updates the parameters of the AI model according to the average gradient of the current round of model training, and sends the average gradient of the current round of model training to participating federated learning each terminal.
  • the base station calculates the average gradient of the current round of model training, updates the parameters of the AI model according to the average gradient of the current round of model training, and sends the average gradient of the current round of model training to participating federated learning each terminal.
  • step 706 refer to the aforementioned step 507.
  • Step 707 The base station judges the model termination condition, and sends a model training termination instruction to each terminal.
  • the terminal can use the configured training duration and reporting time to ensure that the terminal simultaneously reports the gradient of the local training.
  • it can be configured that when the terminal receives the average gradient of the last round of model training sent by the base station, it will start the current round of model training, and report the current round at T seconds after receiving the average gradient of the previous round of model training Local training gradients during model training.
  • the Tth second is the training period
  • the Tth second after receiving the average gradient of the last round of model training is the reporting time.
  • the configured training duration and reporting time can reduce the time error of gradient reporting by each terminal.
  • the base station sends parameters such as training duration, power adjustment scheme, and reporting time to each terminal participating in federated learning in a pre-configured manner, and the terminal periodically trains the model and reports the gradient.
  • the terminal uses channel reciprocity to measure the quality of the downlink channel to optimize its own transmit power, improve the performance of air computing, and then improve the training effect of federated learning.
  • the terminal actively reports the training gradient, which can reduce the scheduling signaling overhead of the base station to the terminal.
  • a flow chart of an AI model training method in a wireless network is provided.
  • the main difference between this flow chart and the process shown in Figure 5 above is that the terminals participating in the federated learning are grouped and then scheduled by the central node The terminals in a certain group uniformly report the gradient of the current round of model training, including at least the following steps:
  • Step 801 the terminal reports terminal information to the base station, and the base station reports the terminal information to the central node in a unified manner.
  • Step 802 the central node sends second configuration information to the base station.
  • the central node can determine the list of terminals participating in the federated learning, the initial AI model, the training duration, the threshold of the number of terminals, the size of the transmission block, the uplink requirement, or the temporary identification of the group, etc.
  • the second configuration information is used to configure at least one of the following: a list of terminals participating in federated learning, an initial AI model, training duration, a threshold for the number of terminals, a transmission block size, uplink requirements, or group temporary identifiers.
  • the central node can comprehensively consider terminal communication capabilities, terminal computing capabilities, and data set characteristics, etc., determine the list of terminals participating in federated learning within the service range of each base station, and use terminals participating in federated learning within the service range of a base station as a group, and assign a temporary identifier to each group, which may be called a group temporary identifier.
  • the central node when it considers comprehensively, it can set priorities for the terminal's communication capabilities, computing capabilities, and data set features. The priority of the data set features is higher than the terminal's communication capabilities, and the terminal's communication capabilities are higher than the terminal. computing power.
  • corresponding thresholds are set for the communication capabilities, computing capabilities, and data set characteristics of the terminals, and those that do not meet the threshold conditions will not be considered to be included in the list of terminals participating in federated learning in this group.
  • the terminals participating in federated learning within the coverage of a base station can be regarded as a group.
  • the m terminals within the coverage of base station 1 are taken as one group
  • the n terminals within the coverage of base station N are taken as another group as an example.
  • Both m and n are positive integers, And the values of m and n can be the same or different.
  • the terminal can allocate a temporary identifier for each group, which is called a group temporary identifier.
  • group temporary identifier a temporary identifier for each group.
  • Step 803 The base station sends the first configuration information to the terminal, the first configuration information is used to configure at least one of the following: initial AI model, group temporary identifier, training duration, time-frequency resource, dedicated bearer RB resource or modulation mode, etc.
  • the base station may determine time-frequency resources, dedicated bearer RB resources, modulation schemes, etc. according to uplink requirements in the first configuration information.
  • the time-frequency resource may be configured by the central node and delivered to each terminal through the base station.
  • the central node can allocate the same time-frequency resources to all terminals in the same group.
  • Step 804 The central node sends a scheduling instruction to the base station, and the base station forwards the scheduling instruction to the terminal.
  • the scheduling instruction includes a group temporary identifier, and the scheduling instruction is used to schedule the terminal corresponding to the group temporary identifier, and is executed in the current round of model training AI model training.
  • the central node specifically schedules which group to perform model training and report the gradient of the current round of model training, and then sends a scheduling instruction to the base station corresponding to the group.
  • the base station may broadcast the scheduling instruction within its coverage area.
  • the terminal may compare the group temporary identifier carried in the scheduling instruction with the group temporary identifier assigned to it by the central node. If the two are the same, the terminal performs AI model training in the current round of model training; otherwise, the terminal does not perform AI model training in the current round of model training.
  • the terminal at the end of the training period, if the terminal completes the training of the AI model, it will report the training completion indication to the base station; training.
  • the base station counts the number of terminals that have completed model training within the training duration according to the training completion indication reported by the terminal, and reports the number of terminals to the central node.
  • Step 805 the base station sends third configuration information to the terminal, where the third configuration information is used to configure the transmit power and reporting time of each terminal.
  • the base station configures the same reporting time for each terminal.
  • the reporting time may be configured by the central node and forwarded to each terminal through the base station, that is, the second configuration information may also be used to configure the reporting time.
  • the third configuration information and the first configuration information may be referred to as a piece of configuration information.
  • the process for the base station to determine the transmit power includes: the base station measures the SRS from the terminal to determine the uplink channel quality of the terminal; the terminal determines the transmit power of the terminal according to the uplink channel quality.
  • the base station measures the SRS from the terminal to determine the uplink channel quality of the terminal; the terminal determines the transmit power of the terminal according to the uplink channel quality.
  • Step 806 The terminal reports the gradient of the current round of model training to the central node, the central node calculates the average gradient of the current round of model training, updates the parameters of the AI model according to the average gradient of the current round of model training, and passes the average gradient of the current round of model training through The base station delivers to the terminals in the current scheduling group.
  • the central node can compare the relationship between the number of terminals that have completed AI model training and the threshold of the number of terminals within the training period; when the number of terminals that have completed AI model training is greater than the threshold of terminals within the training period, then Calculate the average gradient of the current round of model training according to the gradient of the current round of model training reported by the terminal; otherwise, use the average gradient of the previous round of model training as the average gradient of the current round of model training.
  • the central node updates the parameters of the AI model according to the average gradient of the current round of model training, and sends the average gradient of the current round of model training to the terminals in the aforementioned scheduling group through the base station.
  • Step 807 The central node judges the model training termination condition, and sends a model training termination instruction to the corresponding terminal.
  • the central node determines whether the termination condition for model training is met, while the above-mentioned Figure 5 or Figure 7 shows In the process, the base station determines whether the model training termination condition is satisfied. It can be understood that the central node may first send the above instruction to terminate the model training to the base station, and the base station forwards it to the terminal and so on.
  • the central node may schedule all grouped terminals to participate in federated learning training, or may only schedule some grouped terminals to participate in federated learning training.
  • the terminal when the terminal receives the above scheduling instruction, it can perform federated learning training; otherwise, it does not perform federated learning training.
  • the terminal can also configure the monitoring parameters of the terminal, and the parameters include the monitoring time and the monitoring duration.
  • the scheduling instruction is monitored according to the configured monitoring duration.
  • the monitoring duration is reached, it will remain dormant, thereby saving the power of the terminal.
  • the base station, the terminal and the central node include corresponding hardware structures and/or software modules for performing respective functions.
  • the application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed by hardware or computer software drives the hardware depends on the specific application scenario and design constraints of the technical solution.
  • FIG. 9 and FIG. 10 are schematic structural diagrams of possible communication devices provided in the present application. These communication devices can be used to implement the functions of a terminal, a base station, or a central node in the above method, and thus can also realize the beneficial effects of the above method.
  • the communication device when the communication device realizes the function of a terminal, it may be one of the terminals 120a-120j as shown in FIG. 1; when the communication device realizes the function of a base station, it may be the base station 110a or 110b may also be a module (such as a chip) applied to a terminal or a base station.
  • a communication device 900 includes a processing unit 910 and a transceiver unit 920 .
  • the communication device 900 is configured to implement functions of a terminal, a base station, or a central node in the methods shown in FIG. 4 , FIG. 5 , FIG. 7 or FIG. 8 above.
  • the information is at least used for configuration: training duration, time-frequency resources, and reporting time; wherein, the training duration, time-frequency resources, and reporting time configured for different terminals participating in federated learning are the same; and, receiving the terminal participating in federated learning A signal obtained by superimposing the reported gradient in the air, where the gradient is the gradient of the AI model that has been trained within the training duration and reported by the terminal using the time-frequency resource at the reporting moment.
  • the processing unit 910 is configured to generate the first configuration information, and process the gradient reported by the terminal.
  • the transceiver unit 920 is used to receive the first configuration information from the second node, the first configuration information At least for configuration: training duration, time-frequency resources, and reporting time; wherein, the training duration, time-frequency resources, and reporting time configured for different terminals participating in federated learning are the same; the processing unit 910 is used to , train the AI model, and obtain the gradient of the AI model of the current round of model training; the transceiver unit 920 is further configured to report the current round of model training to the second node by using the time-frequency resource at the reporting time The gradient of the AI model.
  • the processing unit 910 is used to determine the second configuration information, and the second configuration information is used to configure At least one of the following: a list of terminals participating in federated learning, an initial AI model, a group temporary identifier, a training duration, a threshold for the number of terminals, a transmission block size, or an uplink requirement; the transceiver unit 920 is configured to send the second configuration to the second node information.
  • processing unit 910 and the transceiver unit 920 can be directly obtained by referring to the relevant descriptions in the methods shown in FIG. 4 , FIG. 5 , FIG. 7 or FIG. 8 , and will not be repeated here.
  • a communication device 1000 includes a processor 1010 and an interface circuit 1020 .
  • the processor 1010 and the interface circuit 1020 are coupled to each other.
  • the interface circuit 1020 may be a transceiver or an input-output interface.
  • the communication device 1000 may further include a memory 1030 for storing instructions executed by the processor 1010 or storing input data required by the processor 1010 to execute the instructions or storing data generated by the processor 1010 after executing the instructions.
  • the processor 1010 is used to implement the functions of the processing unit 910
  • the interface circuit 1020 is used to implement the functions of the transceiver unit 920 .
  • the terminal chip implements the functions of the terminal in the above method.
  • the terminal chip receives information from other modules in the terminal (such as radio frequency modules or antennas), and the information is sent to the terminal by the base station; or, the terminal chip sends information to other modules in the terminal (such as radio frequency modules or antennas), and the The information is sent by the terminal to the base station.
  • the base station module implements the functions of the base station in the above method.
  • the base station module receives information from other modules in the base station (such as radio frequency modules or antennas), and the information is sent by the terminal to the base station; or, the base station module sends information to other modules in the base station (such as radio frequency modules or antennas), the The information is sent by the base station to the terminal.
  • the base station module here may be a baseband chip of the base station, or a DU or other modules, and the DU here may be a DU under an open radio access network (O-RAN) architecture.
  • OF-RAN open radio access network
  • the central node can realize the function of the central node in the above method.
  • the central node receives information from other modules of the central node (such as radio frequency modules or antennas), and the information is sent to the central node by the base station; or, the central node module sends information to other modules (such as radio frequency modules or antennas) in the central node Information, which is sent by the central node to the base station.
  • the central node module here may be a baseband chip or other modules of the central node.
  • processor in this application can be a central processing unit (central processing unit, CPU), and can also be other general processors, digital signal processors (digital signal processor, DSP), application specific integrated circuits (application specific integrated circuit, ASIC), field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof.
  • CPU central processing unit
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • a general-purpose processor can be a microprocessor, or any conventional processor.
  • the memory in this application can be random access memory, flash memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, register, hard disk, mobile hard disk, CD-ROM or any other form of storage media known in the art.
  • An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium.
  • a storage medium may also be an integral part of the processor.
  • the processor and storage medium can be located in the ASIC.
  • the ASIC can be located in the base station or the terminal.
  • the processor and the storage medium may also exist in the base station or the terminal as discrete components.
  • the methods in this application may be fully or partially implemented by software, hardware, firmware or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product comprises one or more computer programs or instructions. When the computer programs or instructions are loaded and executed on the computer, the processes or functions described in this application are executed in whole or in part.
  • the computer may be a general computer, a special computer, a computer network, a network device, a user device, a core network device, an OAM or other programmable devices.
  • the computer program or instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instructions may be downloaded from a website, computer, A server or data center transmits to another website site, computer, server or data center by wired or wireless means.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrating one or more available media.
  • the available medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; it may also be an optical medium, such as a digital video disk; or it may be a semiconductor medium, such as a solid state disk.
  • the computer readable storage medium may be a volatile or a nonvolatile storage medium, or may include both volatile and nonvolatile types of storage media.
  • “at least one” means one or more, and “multiple” means two or more.
  • “And/or” describes the association relationship of associated objects, indicating that there may be three types of relationships, for example, A and/or B, which can mean: A exists alone, A and B exist at the same time, and B exists alone, where A, B can be singular or plural.
  • the character “/” generally indicates that the contextual objects are an “or” relationship; in the formulas of this application, the character “/” indicates that the contextual objects are a “division” Relationship.
  • “Including at least one of A, B or C” may mean: including A; including B; including C; including A and B; including A and C; including B and C; including A, B, and C.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Neurology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

一种无线网络中的人工智能AI模型训练方法及装置，包括：向参与联邦学习的终端发送第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻均相同；接收来自所述终端上报的梯度在空中叠加后的信号，所述梯度是所述终端在所述上报时刻，利用所述时频资源上报的在所述训练时长内训练完成的AI模型的梯度。采用本申请的方法及装置，可解决因参与联邦学习的边缘节点需要使用相互正交的上行时频资源，而导致的时频资源占用率太高，和时延大等问题。

Description

一种无线网络中的人工智能AI模型训练方法及装置
相关申请的交叉引用
本申请要求在2021年12月10日提交中国专利局、申请号为202111505116.7、申请名称为“一种无线网络中的人工智能AI模型训练方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及人工智能(artificial intelligence，AI)领域，尤其涉及一种无线网络中的AI模型训练方法及装置。
背景技术
在无线通信网络中，例如在移动通信网络中，网络支持的业务越来越多样，因此需要满足的需求越来越多样。例如，网络需要能够支持超高速率、超低时延、和/或超大连接。该特点使得网络规划、网络配置、和/或资源调度越来越复杂。这些新需求、新场景和新特性给网络规划、运维和高效运营带来了前所未有的挑战。为了迎接该挑战，可以将人工智能技术引入无线通信网络中，从而实现网络智能化。其中，联邦学习是一种流行的模型训练架构，可以在不损害隐私的前提下调动位于边缘设备的分布式数据参与模型训练。在联邦学习中，中心节点和边缘设备需要交互人工智能(artificial intelligence，AI)模型的参数或梯度。如何在无线网络中部署联邦学习，是一个值得研究的问题。
发明内容
本申请提供一种无线网络中的AI模型训练方法及装置，以解决在无线网络中部署联邦学习时，参与联邦学习的边缘节点需要使用相互正交的上行时频资源，而导致的时频资源占用率太高，和时延大等问题。
第一方面，提供一种无线网络中的人工智能AI模型训练方法，该方法的执行主体为第二节点，还可以为配置于第二节点中的部件(处理器、芯片或其它等)，或者可以为软件模块，包括：向参与联邦学习的终端发送第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻均相同；接收参与联邦学习的终端上报的梯度在空中叠加后的信号，所述梯度是所述终端在所述上报时刻，利用所述时频资源上报的在所述训练时长内训练完成的AI模型的梯度。
应当指出，在本申请的描述中，第二节点还可称为接入网设备，第一节点还可称为中心节点等。通过上述设计，接入网设备(可称为第二节点)或中心节点(可称为第一节点)为参与联邦学习的终端所配置的训练时长、时频资源和上报时刻都相同。示例的，参与联邦学习的终端的数量为n，在本申请中，接入网设备或中心节点为该n个终端分配的时频资源相同。相对于目前方案中，为参与联邦学习的n个终端分配n个正交的时频资源，可以降低时频资源的开销。同时，若为上述参与联邦学习的n个终端分配n个正交的时频资 源，每个终端利用各自的时频资源上报AI模型的梯度。接入网设备可以接收到n个射频信号，对上述n个射频信号分别进行处理，才能恢复出每个终端上报的梯度，时延较大。而在本申请中，接入网设备为参与联邦学习的n个终端分配1个时频资源，该n个终端都在该时频资源上上报梯度。那么，根据无线信道的叠加特性，这n个梯度在空中传输时会叠加到一起。比如，n的取值为3，终端1上报的梯度为Y1，终端2上报的梯度为Y2，终端3上报的梯度为Y3，则将上述3个梯度在同一个时频资源上传输，上述3个梯度会叠加到一起，假设无线信道满足完美的信号叠加(衰落、干扰和噪声等可以忽略不计)，则叠加后的信号为Y＝Y1+Y2+Y3。当信道不满足完美的信号叠加时，接入网设备在上述时频资源上，接收到上述信号后，可对接收到的信号进行信号处理，恢复出叠加的信号Y，终端利用叠加的信号Y进行梯度聚合即可，该梯度聚合的过程可以为求算术平均的过程，例如，可对上述叠加的信号Y除3，结果作为聚合的梯度等。采用本申请的方案，对一个射频信号进行处理，即可得到叠加的信号Y，而在现有方案中，需要对不同时隙的3个射频信号依次进行处理，各自恢复出对应的梯度，再进一步进行聚合，采用本申请的方案，在一定程度上可以减少时频资源占用，也可以降低梯度聚合的时延。
应理解，上述第二节点(或称为接入网设备)配置给终端的训练时长有可能并不是终端的实际训练时长，而是指终端每轮训练所需要的时间上限。即终端在这个训练时长内完成本轮模型训练、并向基站上报训练完成指示即可，否则，终端可以终止本轮模型训练，等待下一个训练时长的到达。在本申请中，设置训练时长是为了保证不同终端可以同时向接入网设备上报模型训练的梯度。可选的，终端向接入网设备上报的模型训练的梯度是本轮模型训练的梯度，而不是其它轮模型训练的梯度，例如上一轮模型训练的梯度等。示例性的，上述训练时长可以是综合考虑了参与联邦学习的各个终端的计算能力和模型的复杂度等情况综合确定的一个全局参数。
在一种可能的设计中，还包括：接收来自所述终端的训练完成指示，所述训练完成指示是所述终端在所述训练时长内，对所述AI模型训练完成时，向第二节点发送的；根据所述终端发送的训练完成指示，统计在所述训练时长内，完成AI模型训练的终端数量。
在一种可能的设计中，还包括：若所述完成AI模型训练的终端数量大于或等于终端数阈值，则根据参与联邦学习的不同终端上报的梯度，确定本轮模型训练的平均梯度；否则，将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度；根据所述本轮模型训练的平均梯度更新所述AI模型的参数，并向所述终端发送所述本轮模型训练的平均梯度。
通过上述设计，第一节点可以确定终端数阈值。将空中计算引入到联邦学习中时，参与联邦学习的终端数量会影响计算本轮模型训练的平均梯度的准确性。第一节点可以设置终端数阈值。当上报梯度的终端数量大于或等于终端数阈值时，才计算本轮模型训练的平均梯度发送给终端；否则将上一轮模型训练的平均梯度发送给终端，或者也可以描述为将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度发送给终端；从而保证所计算的本轮模型训练的平均梯度的准确性满足要求。
在一种可能的设计中，还包括：向第一节点发送在所述训练时长内，完成AI模型训练的终端数量，以及所述终端上报的梯度在空中叠加信号。
在一种可能的设计中，所述第一配置信息还用于配置以下至少一项：专用的承载RB资源、调制方式、初始AI模型、或发射功率。
在一种可能的设计中，确定所述发射功率的过程，包括：根据对来自所述终端的探测 参考信号SRS进行测量，确定所述终端的上行信道质量；根据所述上行信道质量，确定所述终端的发射功率。
通过上述设计，第二节点通过对上行信道测量，确定最佳的发射功率，且将该发射功率配置给终端发送本轮模型训练的梯度，从而提高空中计算的精度，进而提高梯度聚合的准确性。
在一种可能的设计中，所述第一配置信息还用于配置以下至少一项：专用的承载RB资源、调制方式、初始AI模型、信道状态信息CSI区间、或信道反转参数。
在一种可能的设计中，还包括：接收来自第一节点的第二配置信息，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、组临时标识、训练时长、终端数阈值、传输块大小或上行需求。
在一种可能的设计中，还包括：接收来自所述终端的第一终端信息，向第一节点发送第二终端信息；其中，所述第一终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、或终端的数据集特征；所述第二终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、终端的数据集特征、或终端临时标识，所述终端临时标识是所述第二节点分配给所述终端的。
在该设计中，终端的通信能力，例如包括终端能支持的最大发射功率、终端的天线配置等。终端的计算能力，例如包括中央处理器(central processing unit，CPU)性能、图形处理器(graphics processing unit，GPU)性能、存储空间和电量等。终端的数据集特征，例如数据集的大小、数据集的分布、数据集是否完备以及数据集的标签是否完整等。可选的，数据集可以按照百分比进一步划分为训练集、验证集和测试集。例如，数据集中的60％是训练集、20％是验证集，20％是测试集等。可以理解的是，训练集用于对AI模型进行训练，验证集用于评估训练后的AI模型，测试集用于对训练后的AI模型进行测试。终端临时标识，该标识可以是小区无线网络临时标识(cell-radio network temporary identifier，C-RNTI)或其它临时标识等。
在一种可能的设计中，还包括：在满足模型训练终止条件时，向所述终端发送终止模型训练指示；或者，接收来自第一节点的终止模型训练指示，向所述终端转发所述终止模型训练指示。
第二方面，提供一种无线网络中的人工智能AI模型训练方法，该方法与上述第一方面相对应，有益效果可参见上述第一方面的描述，该方法的执行主体为终端，还可以为配置于终端中的部件(处理器、芯片或其它)，或者可以为软件模块等，包括：接收来自第二节点的第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源、以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻相同；在所述训练时长内，对AI模型进行训练，获得本轮模型训练的AI模型的梯度；在所述上报时刻，利用所述时频资源，向所述第二节点上报所述本轮模型训练的AI模型的梯度。
在一种可能的设计中，还包括：在所述训练时长结束时，若所述AI模型完成训练，则向所述第二节点发送训练完成指示。
在一种可能的设计中，还包括：若在所述训练时长内未完成所述AI模型训练，则终止对所述AI模型的训练。
在一种可能的设计中，还包括：接收来自第二节点的上一轮模型训练的平均梯度；根据所述上一轮模型训练的平均梯度，更新本轮模型训练中所述AI模型的梯度；或者，根 据所述本轮模型训练的平均梯度，在所述本轮模型训练中更新所述AI模型的参数和梯度。
在一种可能的设计中，所述第一配置信息还用于配置以下至少一项：专用的承载RB资源、调制方式、初始AI模型或发射功率。
在一种可能的设计中，所述第一配置信息还用于配置以下至少一项：专用的承载RB资源、调制方式、初始AI模型、信道状态信息CSI区间、或信道反转参数。
在一种可能的设计中，所述第一配置信息还用于配置所述信道状态信息CSI区间和信道反转参数时，所述方法还包括：若下行信道和上行信道所配置的频率资源相同，则根据测量的下行信道的CSI，确定所述终端的上行信道的CSI；若所述上行信道的CSI满足CSI区间的要求，则根据所述信道反转参数，确定发射功率；向所述第二节点上报本轮模型训练的AI模型的梯度，包括：基于确定的所述发射功率，向所述第二节点上报所述本轮模型训练的AI模型的梯度。
在一种可能的设计中，所述第一配置信息中还包括组临时标识，所述组临时标识是第一节点为所述终端分配的组临时标识。
在一种可能的设计中，还包括：接收来自所述第二节点的调度指示，所述调度指示中包括组临时标识；当所述调度指示中包括的组临时标识，与所述第一节点为所述终端分配的组临时标识相同时，则在本轮模型训练中执行AI模型训练；否则，在本轮模型训练中不再执行AI模型训练。
在一种可能的设计中，还包括：接收来自所述第二节点的终止模型训练指示；根据所述终止模型训练指示，终止对所述AI模型的训练。
在一种可能的设计中，还包括：向所述第二节点发送第一终端信息，所述第一终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、或终端的数据集特征。
第三方面，提供一种无线网络中的人工智能AI模型训练方法，该方法与上述第一方面对应，有益效果可参见上述第一方面的描述，该方法的执行主体为第一节点，还可以为配置于第一节点中的部件(处理器、芯片或其它)，或者可以为软件模块等，包括：确定第二配置信息，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、组临时标识、训练时长、终端数阈值、传输块大小、或上行需求；向第二节点发送所述第二配置信息。
在一种可能的设计中，还包括：接收来自所述第二节点的第二终端信息，所述第二终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、终端的数据集特征、或终端临时标识，所述终端临时标识是所述第二节点分配给所述终端的；根据所述终端信息，确定参与联邦学习的终端列表。
在一种可能的设计中，还包括：接收来自所述第二节点的参与联邦学习的终端上报的梯度在空中叠加的信号以及在所述训练时长内完成AI模型训练的终端数量；若在所述训练时长内完成模型训练的终端数量大于或等于所述终端数阈值，则根据所述终端上报的AI模型的梯度，确定本轮模型训练的平均梯度；否则，将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度；根据所述本轮模型训练的平均梯度更新所述AI模型的参数，并向所述第二节点发送所述本轮模型训练的平均梯度，以使得所述第二节点将所述本轮模型训练的平均梯度发送给所述终端。
在一种可能的设计中，还包括：向所述第二节点发送调度指示，所述调度指示中包括组临时标识，所述调度指示用于调度组临时标识对应的终端，在本轮模型训练内执行AI 模型训练。
在一种可能的设计中，还包括：在满足模型训练终止条件时，向所述第二节点发送终止模型训练指示，用于指示所述终端在本轮模型训练中，停止对AI模型的训练。
第四方面，提供一种装置，该装置包括执行第一方面、第二方面或第三方面中所描述的方法/操作/步骤/动作一一对应的单元或模块，该单元或模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。
第五方面，提供一种通信装置，该装置包括处理器与存储器。其中，存储器用于存储计算机程序或指令，处理器与存储器耦合；当处理器执行计算机程序或指令时，使得该装置执行上述第一方面、第二方面或第三方面的方法。
第六方面，提供一种装置，包括处理器和接口电路，所述处理器用于通过接口电路与其它装置通信，并执行上述第一方面、第二方面或第三方面中任一方面所描述的方法。该处理器包括一个或多个。
第七方面，提供一种装置，包括与存储器耦合的处理器，该处理器用于执行所述存储器中存储的程序，以执行上述第一方面、第二方面或第三方面中任一方面描述的方法。该存储器可以位于该装置之内，也可以位于该装置之外。且该处理器可以是一个或多个。
第八方面，提供一种芯片系统，包括：处理器或电路，用于执行上述第一方面、第二方面或第三方面中任一方面描述的方法。
第九方面，提供一种计算机可读存储介质，该计算机可读存储介质中存储有计算机程序或指令，当计算机程序或指令被装置执行时，使得该装置执行上述第一方面、第二方面、或第三方面的方法。
第十方面，提供一种计算机程序产品，该计算机程序产品包括计算机程序或指令，当计算机程序或指令被装置执行时，使得该装置执行上述第一方面、第二方面或第三方面的方法。
第十一方面，提供一种系统，包括执行前述第一方面方法的装置和前述第二方面方法的装置。可选的，还可包括执行前述第三方面方法的装置。
附图说明
图1为本申请提供的网络架构示意图；
图2a和图2b为本申请提供的神经网络的示意图；
图2c为本申请提供的AI模型的示意图；
图3为本申请提供的联邦学习训练的示意图；
图4和图5为本申请提供的一流程示意图；
图6为本申请提供的信道测量示意图；
图7和图8为本申请提供的另一流程示意图；
图9和图10为本申请提供的装置示意图。
具体实施方式
图1是本申请能够应用的通信系统1000的架构示意图。如图1所示，该通信系统包括无线接入网100和核心网200，可选的，通信系统1000还可以包括互联网300。其中，无 线接入网100可以包括至少一个接入网设备(如图1中的110a和110b)，还可以包括至少一个终端(如图1中的120a-120j)。终端通过无线的方式与接入网设备相连，接入网设备通过无线或有线方式与核心网连接。核心网设备与接入网设备可以是独立的不同的物理设备，或者可以是将核心网设备的功能与接入网设备的逻辑功能集成在同一个物理设备上，或者可以是一个物理设备上集成了部分核心网设备的功能和部分的接入网设备的功能。终端和终端之间以及接入网设备和接入网设备之间可以通过有线或无线的方式相互连接。图1只是示意图，该通信系统中还可以包括其它网络设备，如还可以包括无线中继设备和无线回传设备等，在图1中未画出。
接入网设备可以是基站(base station)、演进型基站(evolved NodeB，eNodeB)、发送接收点(transmission reception point，TRP)、第五代(5th generation，5G)移动通信系统中的下一代基站(next generation NodeB，gNB)、开放无线接入网(open radio access network，O-RAN)中的接入网设备、第六代(6th generation，6G)移动通信系统中的下一代基站、未来移动通信系统中的基站或无线保真(wireless fidelity，WiFi)系统中的接入节点等；或者可以是完成基站部分功能的模块或单元，例如，可以是集中式单元(central unit，CU)、分布式单元(distributed unit，DU)、集中单元控制面(CU control plane，CU-CP)模块、或集中单元用户面(CU user plane，CU-UP)模块。接入网设备可以是宏基站(如图1中的110a)，也可以是微基站或室内站(如图1中的110b)，还可以是中继节点或施主节点等。本申请中对接入网设备所采用的具体技术和具体设备形态不做限定。
在本申请中，用于实现接入网设备的功能的装置可以是接入网设备；也可以是能够支持接入网设备实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在接入网设备中或可以与接入网设备匹配使用。在本申请中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。为了便于描述，下文以用于实现接入网设备的功能的装置是接入网设备，接入网设备为基站为例，描述本申请提供的技术方案。
(1)协议层结构。
接入网设备和终端之间的通信遵循一定的协议层结构。该协议层结构可以包括控制面协议层结构和用户面协议层结构。例如，控制面协议层结构可以包括无线资源控制(radio resource control，RRC)层、分组数据汇聚层协议(packet data convergence protocol，PDCP)层、无线链路控制(radio link control，RLC)层、媒体接入控制(media access control，MAC)层和物理层等协议层的功能。例如，用户面协议层结构可以包括PDCP层、RLC层、MAC层和物理层等协议层的功能，在一种可能的实现中，PDCP层之上还可以包括业务数据适配协议(service data adaptation protocol，SDAP)层。
可选的，接入网设备和终端之间的协议层结构还可以包括人工智能(artificial intelligence，AI)层，用于传输AI功能相关的数据。
(2)集中单元(central unit，CU)和分布单元(distributed unit，DU)。
接入设备可以包括CU和DU。多个DU可以由一个CU集中控制。作为示例，CU和DU之间的接口可以称为F1接口。其中，控制面(control panel，CP)接口可以为F1-C，用户面(user panel，UP)接口可以为F1-U。本申请不限制各接口的具体名称。CU和DU可以根据无线网络的协议层划分：比如，PDCP层及以上协议层的功能设置在CU，PDCP层以下协议层(例如RLC层和MAC层等)的功能设置在DU；又比如，PDCP层以上协 议层的功能设置在CU，PDCP层及以下协议层的功能设置在DU，不予限制。
上述对CU和DU的处理功能按照协议层的划分仅仅是一种举例，也可以按照其他的方式进行划分。例如可以将CU或者DU划分为具有更多协议层的功能，又例如将CU或DU还可以划分为具有协议层的部分处理功能。在一种设计中，将RLC层的部分功能和RLC层以上的协议层的功能设置在CU，将RLC层的剩余功能和RLC层以下的协议层的功能设置在DU。在另一种设计中，还可以按照业务类型或者其他系统需求对CU或者DU的功能进行划分，例如按时延划分，将处理时间需要满足时延要求的功能设置在DU，不需要满足该时延要求的功能设置在CU。在另一种设计中，CU也可以具有核心网的一个或多个功能。示例性的，CU可以设置在网络侧方便集中管理。在另一种设计中，将DU的无线单元(radio unit，RU)拉远设置。可选的，RU可以具有射频功能。
可选的，DU和RU可以在物理层(physical layer，PHY)进行划分。例如，DU可以实现PHY层中的高层功能，RU可以实现PHY层中的低层功能。其中，用于发送时，PHY层的功能可以包括以下至少一项：添加循环冗余校验(cyclic redundancy check，CRC)码、信道编码、速率匹配、加扰、调制、层映射、预编码、资源映射、物理天线映射、或射频发送功能。用于接收时，PHY层的功能可以包括以下至少一项：CRC校验、信道解码、解速率匹配、解扰、解调、解层映射、信道检测、资源解映射、物理天线解映射、或射频接收功能。其中，PHY层中的高层功能可以包括PHY层的一部分功能，例如该部分功能更加靠近MAC层，PHY层中的低层功能可以包括PHY层的另一部分功能，例如该部分功能更加靠近射频功能。例如，PHY层中的高层功能可以包括添加CRC码、信道编码、速率匹配、加扰、调制、和层映射，PHY层中的低层功能可以包括预编码、资源映射、物理天线映射、和射频发送功能；或者，PHY层中的高层功能可以包括添加CRC码、信道编码、速率匹配、加扰、调制、层映射和预编码，PHY层中的低层功能可以包括资源映射、物理天线映射、和射频发送功能。例如，PHY层中的高层功能可以包括CRC校验、信道解码、解速率匹配、解码、解调、和解层映射，PHY层中的低层功能可以包括信道检测、资源解映射、物理天线解映射、和射频接收功能；或者，PHY层中的高层功能可以包括CRC校验、信道解码、解速率匹配、解码、解调、解层映射、和信道检测，PHY层中的低层功能可以包括资源解映射、物理天线解映射、和射频接收功能。
示例性的，CU的功能可以由一个实体来实现，或者也可以由不同的实体来实现。例如，可以对CU的功能进行进一步划分，即将控制面和用户面分离并通过不同实体来实现，分别为控制面CU实体(即CU-CP实体)和用户面CU实体(即CU-UP实体)。该CU-CP实体和CU-UP实体可以与DU相耦合，共同完成接入网设备的功能。
可选的，上述DU、CU、CU-CP、CU-UP和RU中的任一个可以是软件模块、硬件结构、或者软件模块+硬件结构，不予限制。其中，不同实体的存在形式可以是不同的，不予限制。例如DU、CU、CU-CP、CU-UP是软件模块，RU是硬件结构。这些模块及其执行的方法也在本公开的保护范围内。
一种可能的实现中，接入网设备包括CU-CP、CU-UP、DU和RU。例如，本申请的执行主体包括DU，或者包括DU和RU，或者包括CU-CP、DU和RU，或者包括CU-UP、DU和RU，不予限制。各模块所执行的方法也在本申请的保护范围内。
终端也可以称为终端设备、用户设备(user equipment，UE)、移动台、移动终端等。终端可以广泛应用于各种场景中的通信，例如包括但不限于以下至少一个场景：设备到设 备(device-to-device，D2D)、车物(vehicle to everything，V2X)、机器类通信(machine-type communication，MTC)、物联网(internet of things，IOT)、虚拟现实、增强现实、工业控制、自动驾驶、远程医疗、智能电网、智能家具、智能办公、智能穿戴、智能交通、或智慧城市等。终端可以是手机、平板电脑、带无线收发功能的电脑、可穿戴设备、车辆、无人机、直升机、飞机、轮船、机器人、机械臂、或智能家居设备等。本申请对终端所采用的具体技术和具体设备形态不做限定。
在本申请中，用于实现终端的功能的装置可以是终端；也可以是能够支持终端实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在终端中或可以与终端匹配使用。为了便于描述，下文以用于实现终端的功能的装置是终端为例，描述本申请提供的技术方案。
基站和终端可以是固定位置的，也可以是可移动的。基站和/或终端可以部署在陆地上，包括室内或室外、手持或车载；也可以部署在水面上；还可以部署在空中的飞机、气球和人造卫星上。本申请对基站和终端的应用场景不做限定。基站和终端可以部署在相同的场景或不同的场景，例如，基站和终端同时部署在陆地上；或者，基站部署在陆地上，终端部署在水面上等，不再一一举例。
基站和终端的角色可以是相对的，例如，图1中的直升机或无人机120i可以被配置成移动基站，对于那些通过120i接入到无线接入网100的终端120j来说，终端120i是基站；但对于基站110a来说，120i是终端，即110a与120i之间是通过无线空口协议进行通信的。110a与120i之间也可以是通过基站与基站之间的接口协议进行通信的，此时，相对于110a来说，120i也是基站。因此，基站和终端都可以统一称为通信装置，图1中的110a和110b可以称为具有基站功能的通信装置，图1中的120a-120j可以称为具有终端功能的通信装置。
基站和终端之间、基站和基站之间、终端和终端之间可以通过授权频谱进行通信，也可以通过免授权频谱进行通信，也可以同时通过授权频谱和免授权频谱进行通信；可以通过6千兆赫(gigahertz，GHz)以下的频谱进行通信，也可以通过6GHz以上的频谱进行通信，还可以同时使用6GHz以下的频谱和6GHz以上的频谱进行通信。本申请对无线通信所使用的频谱资源不做限定。
在本申请中，基站向终端发送下行信号或下行信息，下行信息承载在下行信道上；终端向基站发送上行信号或上行信息，上行信息承载在上行信道上。终端为了与基站进行通信，可以与基站控制的小区建立无线连接。与终端建立了无线连接的小区称为该终端的服务小区。当终端与该服务小区进行通信的时候，可能会受到来自邻区的信号的干扰。
在本申请中，可以在前述图1所示的通信系统中引入独立的网元(如称为中心节点、AI网元、或AI节点等)来实现AI相关的操作，中心节点可以和通信系统中的接入网设备之间直接连接，或者可以通过第三方网元和接入网设备实现间接连接，其中，第三方网元可以是认证管理功能(authentication management function，AMF)网元、或用户面功能(user plane function，UPF)网元等核心网网元；或者，可以在通信系统中的其他网元内配置AI功能、AI模块或AI实体来实现AI相关的操作，例如该其他网元可以是接入网设备(如gNB)、核心网设备、或网管(operation,administration and maintenance，OAM)等，在这种情况下，执行AI相关的操作的网元为内置AI功能的网元。本申请中，OAM用于操作、管理和/或维护核心网设备，和/或，用于操作、管理和/或维护接入网设备。
本申请中，AI模型是实现AI功能的具体方法，AI模型表征了模型的输入和输出之间的映射关系。AI模型可以是神经网络或者其他机器学习模型。其中，AI模型可以简称为模型。AI相关的操作可以包括以下至少一项：数据收集、模型训练、模型信息发布、模型推断(模型推理)、或推理结果发布等。
以神经网络为例，神经网络是机器学习技术的一种具体实现形式。根据通用近似定理，神经网络在理论上可以逼近任意连续函数，从而使得神经网络具备学习任意映射的能力。传统的通信系统需要借助丰富的专家知识来设计通信模块，而基于神经网络的深度学习通信系统可以从大量的数据集中自动发现隐含的模式结构，建立数据之间的映射关系，获得优于传统建模方法的性能。
神经网络的思想来源于大脑组织的神经元结构。每个神经元都对其输入值做加权求和运算，将加权求和结果通过一个激活函数产生输出。如图2a所示，为神经元结构示意图。假设神经元的输入为x＝[x 0,x 1,…,x n],与各输入对应的权值分别为w＝[w,w 1,…,w n],加权求和的偏置为b。激活函数的形式可以多样化，假设一个神经元的激活函数为：y＝f(z)＝max(0,z)，则该神经元的输出为：
Figure PCTCN2022137671-appb-000001
Figure PCTCN2022137671-appb-000002
再例如一个神经元的激活函数为：y＝f(z)＝z，则该神经元的输出为：
Figure PCTCN2022137671-appb-000003
Figure PCTCN2022137671-appb-000004
b可以为小数、整数(包括0、正整数或负整数等)、或复数等各种可能的取值。神经网络中不同神经元的激活函数可以相同或不同。
神经网络一般包括多层结构，每层可包括一个或多个神经元。增加神经网络的深度和/或宽度可以提高该神经网络的表达能力，为复杂系统提供更强大的信息提取和抽象建模能力。其中，神经网络的深度可以指神经网络包括的层数，每层包括的神经元个数可以称为该层的宽度。如图2b所示，为神经网络的层关系示意图。一种实现中，神经网络包括输入层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给输出层，由输出层得到神经网络的输出结果。另一种实现中，神经网络包括输入层、隐藏层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给中间的隐藏层，隐藏层再将计算结果传递给输出层或者相邻的隐藏层，最后由输出层得到神经网络的输出结果。一个神经网络可以包括一层或多层依次连接的隐藏层，不予限制。神经网络的训练过程中，可以定义损失函数。损失函数描述了神经网络的输出值和理想目标值之间的差距或差异，本申请不限制损失函数的具体形式。神经网络的训练过程就是通过调整神经网络参数，如神经网络的梯度、层数、宽度、神经元的权值、和/或神经元的激活函数中的参数等，使得损失函数的值小于阈值门限值或者满足目标需求的过程。
如图2c所示为AI的一种应用框架的示意图。数据源(data source)用于存储训练数据和推理数据。模型训练节点(model trainning host)通过对数据源提供的训练数据(training data)进行分析或训练，得到AI模型，且将AI模型部署在模型推理节点(model inference host)中。可选的，模型训练节点还可以对已部署在模型推理节点的AI模型进行更新。模型推理节点还可以向模型训练节点反馈已部署模型的相关信息，以使得模型训练节点对已部署的AI模型进行优化或更新等。
其中，AI模型表征了模型的输入和输出之间的映射关系。通过模型训练节点学习得到AI模型，相当于由模型训练节点利用训练数据学习得到模型的输入和输出之间的映射关系。模型推理节点使用AI模型，基于数据源提供的推理数据进行推理，得到推理结果。该方法还可以描述为：模型推理节点将推理数据输入到AI模型，通过AI模 型得到输出，该输出即为推理结果。该推理结果可以指示：由执行对象使用(执行)的配置参数、和/或由执行对象执行的操作。推理结果可以由执行(actor)实体统一规划，并发送给一个或多个执行对象(例如，网络实体)去执行。
联邦学习(federated learning，FL)是一种流行的AI/ML模型训练框架，能有效帮助多个机构在满足用户隐私保护、数据安全和政府法规的要求下，进行数据使用和机器学习建模。联邦学习作为分布式的机器学习示范，可以有效解决数据孤岛问题，让参与方在不共享数据的基础上联合建模，从技术上打破数据孤岛，实现AI协作。联邦学习包含中心节点和边缘节点，中心节点(例如服务器、基站)可以在不损害隐私的前提下调动位于边缘设备(例如智能手机和传感器)的分布式设备参与模型训练。
联邦学习主要分为以下三种类别：横向联邦学习、纵向联邦学习和联邦迁移学习。本申请主要涉及横向联邦学习的过程。图3所示为横向联邦学习的训练过程，可以看到横向联邦学习由一个中心节点和多个边缘节点组成。其中，原始数据都分布在各个边缘节点，中心节点不具有原始数据，并且边缘节点不允许将原始数据发送给中心节点。
在联邦学习的训练过程中，中心节点首先将初始化的AI模型(可称为初始AI模型)发送给各个边缘节点，然后开始迭代训练，每一次迭代训练过程为：
1、边缘节点利用本地数据，对初始AI模型进行训练，获取训练后AI模型的梯度；
2、每个边缘节点向中心节点上报各自训练出的梯度；
3、中心节点收到各个边缘节点上报的梯度后，对梯度进行聚合，并根据聚合后的梯度更新AI模型的参数；
4、中心节点将聚合后的梯度下发给每个参与训练的边缘节点，边缘节点根据中心节点下发的聚合后的梯度，更新在本地训练的AI模型的参数和梯度。
5、中心节点计算更新参数后的AI模型的损失函数；如果损失函数满足条件，则终止模型训练；如果损失函数未满足条件，则重复上述步骤2-4。
在目前的联邦学习方案中，需要为各个边缘节点分配相互正交的上行时频资源，各个边缘节点在相互正交的上行时频资源上传输梯度，中心节点需要依次恢复出参与联邦学习训练的所有终端上报的梯度才能进行梯度聚合，这会带来较大的时频资源开销和时延，不适用于对带宽受限和/或时延要求较高的联邦学习场景。
本申请提供一种无线网络中的AI模型训练方法，在该方法，可以为参与联邦学习的终端分配相同的时频资源和上报时刻。参与联邦学习的终端，在同一上报时刻，采用相同的时频资源上报训练后AI模型的梯度，从而解决上述由于为参与联邦学习的终端分配相互正交的多个时频资源，所引起的时频资源开销大和时延高等问题。
如图4所示，提供一种无线网络中的AI模型训练方法的流程，该流程至少包括以下步骤：
步骤401：基站向参与联邦学习的终端发送第一配置信息，该第一配置信息用于配置以下至少一项：训练时长、时频资源、或上报时刻。相应的，终端接收来自基站的第一配置信息。
步骤402：终端在上述训练时长内，对AI模型进行训练，获得本轮模型训练的AI模型的梯度。
步骤403：终端在上述上报时刻，利用上述时频资源，向基站上报本轮模型训练的AI模型的梯度。相应的，基站接收终端上报的梯度在空中叠加后的信号。
在本申请中，基站或中心节点为参与联邦学习的终端所配置的训练时长、时频资源和上报时刻都相同。示例的，参与联邦学习的终端的数量为n，在本申请中，基站或中心节点为该n个终端分配的时频资源相同。相对于目前方案中，为参与联邦学习的n个终端分配n个正交的时频资源，可以降低时频资源的开销。同时，若为上述参与联邦学习的n个终端分配n个正交的时频资源，每个终端利用各自的时频资源上报AI模型的梯度。基站可以接收到n个射频信号，对上述n个射频信号分别进行处理，才能恢复出每个终端上报的梯度，时延较大。而在本申请中，基站为参与联邦学习的n个终端分配1个时频资源，该n个终端都在该时频资源上上报梯度。那么，根据无线信道的叠加特性，这n个梯度在空中传输时会叠加到一起。比如，n的取值为3，终端1上报的梯度为Y1，终端2上报的梯度为Y2，终端3上报的梯度为Y3，则将上述3个梯度在同一个时频资源上传输，上述3个梯度会叠加到一起，假设无线信道满足完美的信号叠加(衰落、干扰和噪声等可以忽略不计)，则叠加后的信号为Y＝Y1+Y2+Y3。当信道不满足完美的信号叠加时，基站在上述时频资源上，接收到上述信号后，可对接收到的信号进行信号处理，恢复出叠加的信号Y，终端利用叠加的信号Y进行梯度聚合即可，该梯度聚合的过程可以为求算术平均的过程，例如，可对上述叠加的信号Y除3，结果作为聚合的梯度等。采用本申请的方案，对一个射频信号进行处理，即可得到叠加的信号Y，而在现有方案中，需要对不同时隙的3个射频信号依次进行处理，各自恢复出对应的梯度，再进一步进行聚合，采用本申请的方案，在一定程度上可以减少时频资源占用，也可以降低梯度聚合的时延。
上述基站(或者接入网设备中的其他设备)配置给终端的训练时长有可能并不是终端的实际训练时长，而通常是指终端每轮训练所需要的时间上限。即终端在这个训练时长内完成本轮模型训练、并向基站上报训练完成指示即可，否则，终端可以终止本轮模型训练，等待下一个训练时长的到达。在本申请中，基站设置训练时长是为了保证不同终端可以同时向基站上报模型训练的梯度。可选的，可以较好保证终端向基站上报的模型训练的梯度是本轮模型训练的梯度，而不是其它轮模型训练的梯度，例如上一轮模型训练的梯度等。示例性的，上述训练时长可以是综合考虑了参与联邦学习的各个终端的计算能力和模型的复杂度等情况综合确定的一个全局参数，然后由基站配置给各个终端。
在本申请中，参与联邦学习的终端的数量可以为n，每个终端可认为是一个边缘节点，中心节点可以是基站，或者可以是OAM，或者独立作为一个模块位于核心网等中，不作限定。在后续描述中，是以中心节点独立于基站，即中心节点和基站为两个设备为例，描述的。
如图5所示，提供一种无线网络中的AI模型训练的方法的流程，在该流程中，至少包括：
步骤501：n个终端中的每个终端分别向基站上报终端信息，基站将该终端信息统一收集合，上报给中心节点。在本申请中，所述终端信息中包括以下至少一项：
1、终端的通信能力，例如包括终端能支持的最大发射功率、终端的天线配置等。
2、终端的计算能力，例如包括中央处理器(central processing unit，CPU)性能、图形处理器(graphics processing unit，GPU)性能、存储空间和电量等。
3、终端的数据集特征，例如数据集的大小、数据集的分布、数据集是否完备以及数据集的标签是否完整等。可选的，数据集可以按照百分比进一步划分为训练集、验证集和测试集。例如，数据集中的60％是训练集、20％是验证集，20％是测试集等。可以理解的 是，训练集用于对AI模型进行训练，验证集用于评估训练后的AI模型，测试集用于对训练后的AI模型进行测试。
可选的，基站为终端分配终端临时标识，该标识可以是小区无线网络临时标识(cell-radio network temporary identifier，C-RNTI)或其它临时标识等。可选的，其它临时标识可以采用编码的方式进行区分，比如，采用下述表1或表2所示的序号：
表1
终端 终端1 终端2 终端n
终端临时标识 1 2 n
表2
终端 终端1 终端2 终端n
终端临时标识 000001 000010 111111
可选的，在步骤501之前，还可以包括：基站向n个终端发送上报终端信息的指示。n个终端根据该指示，在上述步骤501中分别上报各自的终端信息。
步骤502：中心节点向基站发送第二配置信息。可选的，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、训练时长、终端数阈值、传输块大小、或上行需求等。其中，所述上行需求可以包括终端上行传输时的速率、误码率或时延等。所述传输块就是指包含MAC协议数据单元(protocol data unit,PDU)的一个数据块，这个数据块会在一个传输时间间隔(transmission time interval，TTI)上传输。
应当指出，在本申请中，基站在接收到终端上报的终端信息时，可称为第一终端信息。基站为终端分配临时标识，将该临时标识加入到上述终端信息中，组成第二终端信息。所述第二终端信息中除包括终端临时标识外，还可包括以下至少一项：终端的通信能力、终端的计算能力、或终端的数据集特征。基站向中心节点上报第二终端信息。
具体的，中心节点可以根据基站上报的第二终端信息，确定参与联邦学习的终端列表。例如，所述第二终端信息包括终端通信能力、计算能力、数据集的特征、和临时标识等。例如，中心节点综合考虑终端通信能力、计算能力、数据集的特征等，制定参与联邦学习的终端列表。例如，在综合考虑时，可以对终端通信能力、计算能力、数据集的特征等设置优先级，数据集的特征的优先级高于终端通信能力，终端通信能力优先级高于终端计算能力。进一步地，对终端通信能力、计算能力、数据集的特征分别设置相应的阈值，不满足阈值条件的则不会被考虑放进参与联邦学习的终端列表。例如，中心节点可以将通信能力大于或等于通信能力阈值，计算能力大于或等于计算能力阈值，数据集特征满足数据集特征要求的终端，作为参与联邦学习的终端。
在本申请中，中心节点可以为参与联邦学习的终端配置训练时长，训练时长需要考虑待训练AI模型的计算复杂度和各终端的计算能力，尽量保证参与联邦学习的终端都能够在训练时长内完成本地模型训练，但也不应设置的太长，以免影响AI模型训练的整体效率等。
在本申请中，中心节点可以确定终端数阈值。将空中计算引入到联邦学习中时，参与联邦学习的终端数量会影响计算本轮模型训练的平均梯度的准确性。所以中心节点可以设 置终端数阈值。当上报梯度的终端数量大于或等于终端数阈值时，才计算本轮模型训练的平均梯度发送给终端；否则将上一轮模型训练的平均梯度发送给终端，或者也可以描述为将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度发送给终端。
步骤503：基站向终端发送第一配置信息。
在本申请中，所述第一配置信息用于配置以下至少一项：训练时长、上报时刻、时频资源、专用的承载无线承载(radio bear，RB)资源、调制方式、或初始AI模型等。
示例的，基站可以根据所述上行需求，确定时频资源、专用的承载RB资源和调制方式等。在本申请中，基站为参与联邦学习的n个终端分配的时频资源、上报时刻、与训练时长均相同。
示例的，基站需要给终端分配专用的承载RB资源，所述专用的承载RB资源可以为信令无线承载(signal radio bear，SRB)资源，或数据无线承载(data radio bear，DRB)资源等，用于传输AI模型的梯度。该专用的承载RB资源可以仅用于传输梯度，而不能用于其它数据传输。基站为终端配置的调制方式可以为相移键控(pase sift kying，PSK)、正交振幅调制(quadrature amplitude modulation，QAM)或其它调制方式等，具体采用多少阶的PSK或QAM调制可以根据上行需求、上行信道质量、基站的通信能力以及终端的通信能力等确定。
步骤504：终端在训练时长内，进行AI模型训练。
在本申请中，终端在接收到上述第一配置信息时，可以对AI模型进行训练。可以理解的是，在首轮的训练过程中，终端具体对初始AI模型进行训练，该初始AI模型是中心节点配置给终端的。在后续的训练过程中，终端具体对前一次训练后的AI模型进行训练。本申请中的训练时长可以用T表示。针对参与联邦学习的终端，如果上述训练时长T内，完成对AI模型的训练，则向基站上报训练完成指示。如果在上述训练时长T内，没有完成对AI模型的训练，则终止模型训练。
步骤505：基站根据终端上报的训练完成指示，统计在训练时长T内，完成AI模型训练的终端数量，并测量完成本轮模型训练的终端的上行信道质量。可选的，对于在本轮模型训练中未完成模型训练的终端，基站不再测量对应终端的上行信道质量，相应的，未完成模型训练的终端，也不再向基站上报本轮模型训练中AI模型的梯度。
示例的，基站可以设置计数器，在每轮训练开始时，对上报训练完成指示的终端进行计数，并在每轮训练结束时计数器重置为0。在训练时长T结束时，如果计数器的计数大于或等于终端数阈值时，则基站测量终端的上行信道质量；否则，基站则触发终端进行下一轮训练。
可选的，针对一个终端，基站测量该终端的上行信道质量的过程，包括：基站接收来自该终端的探测参考信号(sounding reference signal，SRS)；基站对该SRS进行测量，确定该终端的上行信道质量。
步骤506：基站根据终端的上行信道质量，确定对应终端的发射功率；基站向终端发送第三配置信息，该第三配置信息用于配置终端的发射功率，以及上报时刻等。在本申请中，该第三配置信息与上述第一配置信息可统称为一个配置信息。
示例的，基站可以综合考虑空中计算的误差要求，各终端的上行信道质量、各终端支持的最大发射功率以及总功率等条件，确定终端的发射功率。由于空中计算对于各节点的同步有要求，同步误差太大会影响空中计算的准确性。因此，基站可为参与联邦学习的n 个终端所配置的上报时刻相同。此外，信道是一直在变化的，所以信道质量测量存在有效期。因此，各终端需要在上报时刻同时上报本轮模型训练的AI模型的梯度，且上报时刻应处于信道质量测量的有效期内。
可选的，考虑到基站确定终端的发射功率的时间开销，可能会超出信道质量测量结果的有效期，导致功率分配和上行信道质量不匹配等。在一种设计中，基站可以根据终端的历史上行信道质量，预测当前上报时刻的上行信道质量，并提前优化最佳发射功率。具体的预测方式，如图6所示。基站收集各终端在t 1到t 2时刻之间的各终端的历史上行信道状态质量，在t 2时刻开始根据各终端的历史上行信道质量，预测各终端在t 3时刻的上行信道质量，确定各终端在t 3时刻的最佳功率分配方案，并在t 3时刻将最佳功率分配方案下发给各终端，各终端在收到功率分配方案后立刻上报本轮模型训练的梯度。
步骤507：终端上报本轮模型训练的梯度，基站计算本轮模型训练的平均梯度，基站根据本轮模型训练的平均梯度更新AI模型的参数，并将本轮模型训练的平均梯度下发给参与联邦学习的n个终端。
例如，参与联邦学习的n个终端，在上报时刻，若在训练时长T内完成模型训练，则可以向基站上报本轮模型训练的梯度。基站在接收到上述n个终端上报的本轮模型训练的梯度时，可以根据该终端上报的梯度，计算本轮模型训练的平均梯度。例如，由于上述n个终端在同一个时频资源向上报本轮模型训练的梯度，基站接收到的信号是上述n个梯度在空中叠加后的信号。比如，上述n个梯度叠加后的梯度为Y，则基站可以确定本轮模型训练的平均梯度等于Y/n。可选的，在本申请中，对于在训练时长T内，未完成模型训练的终端，不再向基站上报本轮模型训练的梯度。
在本申请中，各终端应尽量在上报时刻到达时，便开始上报本轮模型训练的梯度，尽可能的减少各终端的上报时间误差。基站需要对接收到的空中计算信号作处理以恢复出本轮模型训练的平均梯度。示例的，如果基站中的计数器大于或等于终端数阈值，则基站将本地计算的本轮模型训练的平均梯度下发给各个终端。否则，将上一轮模型训练的平均梯度，作为本轮模型训练的平均梯度，下发给各个终端。
可选的，如果参与联邦学习的终端是变化的，基站可以将更新后的模型参数发送给各终端。
可以理解的是，在上述步骤504至步骤507是一个循环过程。在步骤507中下发本轮模型训练的平均梯度后，终端可以根据该本轮模型训练的平均梯度，更新AI模型的参数和梯度，且在训练时长T内，如果完成模型训练，则向基站上报训练完成指示。应当指出，在机器学习中，为了使损失函数下降，需要让AI模型的参数沿着梯度的负方向下降，即梯度下降。在本申请中，可以先根据上一轮模型训练的平均梯度，在本轮模型训练中更新所述AI模型的参数，再根据更新的参数更新AI模型的梯度，即梯度更新。
步骤508：基站判断模型训练终止条件，并向终端发送终止模型训练指示。
示例的，基站可以判断模型训练终止条件，并向各终端发送终止模型训练指示。模型训练终止条件可以为模型参数收敛、达到模型训练最多次数要求或达到模型训练最高时间要求等中的至少一项。
在本申请中，通过引入空中计算，降低了联邦学习过程中的通信时延、时频资源开销以及信令开销，配置的训练时长和上报时刻可以降低各终端上报梯度的时间同步误差。在梯度上报之前，基站先测量所有参与终端的上行信道质量，并优化出每个终端的发射功率， 可以提升空中计算的性能，进而提升联邦学习的训练效果。
如图7所示，本申请提供一种无线网络中的AI模型训练方法的流程，该流程与上述图5所示流程的主要区别在于，在该流程中，由终端自行确定上报本轮模型训练的梯度的发射功率，该发射功率不再由基站配置，至少包括以下步骤：
步骤701：n个终端向基站上报终端信息，基站将该终端信息统一上报给中心节点。
具体的，终端在接收到终端的终端信息时，可为终端分配临时标识，将临时标识加入到终端信息中，上报给中心节点。可参见上述图5中的说明。
步骤702：中心节点向基站发送第二配置信息，该第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、训练时长、终端数阈值、信道状态信息(channel state information，CSI)区间、信道反转参数、传输块大小或上行需求。关于联邦学习的终端列表、初始AI模型、训练时长、终端数阈值、传输块大小和上行需求等，可参见上述图5中的说明。该流程重点介绍，CSI区间和信道反转参数。
在本申请中，CSI是指信道状态信息，它包含信噪比、多普勒频移和多径时延扩展等。CSI区间包含信噪比区间、最大多普勒频移区间和最大时延扩展区间。其中，信噪比区间为[γ min,γ max]，γ min和γ max分别为信噪比的下限和上限；最大多普勒频移区间为[f min,f max],f min和f max分别为最大多普勒频移的下限和上限；最大时延扩展区间为[τ min,τ max],τ min和τ max分别为最大时延扩展的下限和上限。
在本申请中，当终端的下行CSI满足CSI区间时才可以将本轮模型训练的梯度上报给基站，否则不上报。例如，某终端测得的信噪比、最大多普勒频移和最大时延扩展分别为γ 1、f 1和τ 1，如果γ min≤γ 1≤γ max、f min≤f 1≤f max且τ min≤τ 1≤τ max时，则该终端可以在上报时刻，向基站上报本轮模型训练的梯度，否则不上报本轮模型训练的梯度。
在本申请中，信道反转参数α是用于功率控制的参数，假设第k个终端的最大功率和信道增益分别为P k和h k，且P 1|h 1| 2≤…≤P k|h k| 2≤…≤P K|h K| 2，则α＝P 1|h 1| 2，则第k个终端在上报梯度时的发射功率应为
Figure PCTCN2022137671-appb-000005
步骤703：基站根据上行需求，确定时频资源、专用的承载RB资源和调度方式等。基站向终端发送第一配置信息，该第一配置信息用于配置以下至少一项：训练时长、上报时刻、时频资源、初始AI模型、专用的承载RB资源或调制方式等。
在本申请中，所述专用的承载RB资源包括专用的SRB资源，和/或专用的DRB资源等。基站可以根据上行需求，确定时频资源、SRB/DRB资源、或调制方式等。为了满足空中计算的要求，各终端需要采用相同的时频资源。可选的，基站可以为n终端分配相同的多个时频资源，利用时间分集或频率分集提升空中计算的性能。需要说明的是，在基站为n个终端配置相同的多个时频资源时，在某一个上报时刻，n个终端利用上述多个时频资源向基站上报本轮模型训练的梯度。例如，基站为n个终端配置3个时频资源，则在上报时刻，n个终端同时利用上述3个时频资源中的第一个时频资源、第二个时频资源，和第三时频资源上报本轮模型训练的梯度。也就是说，在上报时刻，n终端上报本轮模型训练的梯度时，所具体采用的时频资源是相同的。为了避免其他数据的干扰，基站可以给终端分配独立的SRB/DRB资源用于传输本轮模型训练的梯度。基站可以为终端配置PSK或QAM等调制方式，具体采用多少阶的PSK或QAM调制可以根据上行需求、上行信道质量、基站的通信能力或终端的通信能力等中的至少一项确定。
步骤704：终端在本地训练模型。
在本申请中，终端可以在接收到第一配置信息时，开始执行模型训练。如果在训练时长T内，终端完成模型训练，则向基站上报训练完成指示。如果在训练时长T内，终端没有完成模型训练，则终止模型训练。
步骤705：终端测量下行信道的CSI，并确定终端的发射功率。
在本申请中，终端上行信道和下行信道可配置相同的频率资源，根据信道互易性，终端可以测量下行信道的CSI，获得上行信道的CSI。终端可以判断获取的上行信道的CSI是否在CSI区间。如果在CSI区间，则根据信道反转参数，确定发射功率。通过前述记载可知，信道反转参数α＝P 1|h 1| 2，第k个终端在上报梯度时的发射功率应为
Figure PCTCN2022137671-appb-000006
在本申请中，针对第k个终端，通过CSI可以获取参数h k，结合信道反转参数α，可以确定第k个终端的发射功率p k，所述k为大于或等于1，小于或等于n的正整数。
步骤706：终端上报本轮模型训练的梯度，基站计算本轮模型训练的平均梯度，根据本轮模型训练的平均梯度更新AI模型的参数，并将本轮模型训练的平均梯度发送给参与联邦学习的各个终端。关于步骤706的具体过程，可参见前述步骤507。
步骤707：基站判断模型终止条件，并向各终端发送终止模型训练指示。
应当指出，在一种设计中，终端可以利用配置的训练时长和上报时刻，保证终端同时上报本地训练的梯度。例如，可以配置当终端接收到基站下发的上一轮模型训练的平均梯度后，就开始进行本轮模型训练，并在接收到上一轮模型训练的平均梯度后的第T秒上报本轮模型训练中的本地训练梯度。其中，第T秒是训练周期，接收到上一轮模型训练的平均梯度后的第T秒是上报时刻。
在本申请中，通过引入空中计算，降低了联邦学习过程中的通信时延和带宽开销，配置的训练时长和上报时刻可以降低各终端上报梯度的时间误差。基站将训练时长、功率调整方案、上报时刻等参数通过预配置的方式发送给各个参与联邦学习的终端，终端周期性的训练模型并上报梯度。在梯度上报之前，终端利用信道互易性，通过测量下行信道质量，优化出自己的发射功率，提升空中计算的性能，进而提升联邦学习的训练效果。此外，终端在训练时长结束时，主动上报训练梯度，可以减少基站对终端的调度信令开销。
如图8所示，提供一种无线网络中的AI模型训练方法的流程图，该流程图与上述图5所示流程的主要区别在于，对参与联邦学习的终端进行分组，之后由中心节点调度某个组中的终端统一进行本轮模型训练的梯度的上报，至少包括以下步骤：
步骤801：终端向基站上报终端信息，基站将该终端信息统一上报给中心节点。
具体的介绍，可参见上述图5中的说明。
步骤802：中心节点向基站发送第二配置信息。
在本申请中，中心节点可以确定参与联邦学习的终端列表、初始AI模型、训练时长、终端数阈值、传输块大小、上行需求、或组临时标识等。所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、训练时长、终端数阈值、传输块大小、上行需求、或组临时标识等。
在本申请中，中心节点可以综合考虑终端通信能力、终端计算能力、数据集的特征等，确定每个基站服务范围内参与联邦学习的终端列表，将一个基站服务范围内参与联邦学习的终端作为一个分组，并为每个分组分配临时标识，该标识可称为组临时标识。例如，中 心节点在综合考虑时，可以对终端的通信能力、计算能力、和数据集特征等设置优先级，数据集特征的优先级高于终端的通信能力、终端的通信能力优先级高于终端的计算能力。进一步地，对终端的通信能力、计算能力、数据集特征等分别设置相应的阈值，不满足阈值条件的，则不会被考虑放进该组参与联邦学习的终端列表。在本申请中，可以将一个基站覆盖范围内参与联邦学习的终端，作为一个分组。例如，在图8的流程中，将基站1覆盖范围内的m个终端为一个分组，将基站N覆盖范围内的n个终端作为另一个分组为例，所述m与n均为正整数，且m与n的取值可以相同或不同。针对每个分组，终端可以为每个组分配临时标识，称为组临时标识。在一种可能的实现方式中，以两个分组为例，每个组内参与联邦学习的终端列表，可参见下述表3所示：
表3
Figure PCTCN2022137671-appb-000007
步骤803：基站向终端发送第一配置信息，该第一配置信息用于配置以下至少一项：初始AI模型、组临时标识、训练时长、时频资源、专用的承载RB资源或调制方式等。
在本申请中，基站可以根据第一配置信息中的上行需求，确定时频资源、专用的承载RB资源和调制方式等。或者，时频资源可以由中心节点配置，并通过基站下发给各个终端。中心节点可以为同一组内的所有终端分配相同的时频资源。
步骤804：中心节点向基站发送调度指示，基站向终端转发该调度指示，所述调度指示中包括组临时标识，所述调度指示用于调度组临时标识对应的终端，在本轮模型训练内执行AI模型训练。
在本申请中，以将一个基站覆盖范围内的n个终端作为一个分组为例。中心节点具体调度哪一个分组进行模型训练和本轮模型训练的梯度上报，则向该分组对应的基站发送调度指示。基站在接收到该调度指示时，可以在其覆盖范围内广播该调度指示。终端在接收在该调度指示时，可将该调度指示中携带的组临时标识，与中心节点为其分配的组临时标识进行对比。若两者相同，则终端在本轮模型训练中执行AI模型训练；否则，在本轮模型训练中，不再执行AI模型训练。
与上述图5所示流程相似，在训练时长结束时，如果终端完成AI模型的训练，则向基站上报训练完成指示；在训练时长结束时，对于没有完成AI模型训练的终端，也终止AI模型的训练。基站根据终端上报的训练完成指示，统计在训练时长内完成模型训练的终端数量，且将该终端数量上报给中心节点。
步骤805：基站向终端发送第三配置信息，该第三配置信息用于配置各个终端的发射功率，以及上报时刻。可选的，基站为各个终端配置相同的上报时刻。或者，该上报时刻可以是中心节点配置，并通过基站转发给各个终端的，即上述第二配置信息中还可以用于配置上报时刻。在本申请中，上述第三配置信息和第一配置信息可称为一个配置信息。
例如，基站确定发射功率的过程，包括：基站对来自终端的SRS进行测量，确定终端的上行信道质量；终端根据上行信道质量，确定终端的发射功率，具体可参见上述图5中的介绍，在此不再赘述。
步骤806：终端向中心节点上报本轮模型训练的梯度，中心节点计算本轮模型训练的平均梯度，根据本轮模型训练的平均梯度更新AI模型的参数，并将本轮模型训练的平均 梯度通过基站下发给当前调度组中的终端。
具体的，中心节点可以比较在训练时长内，完成AI模型训练的终端数量，与终端数量阈值两者的大小关系；当在训练时长内，完成AI模型训练的终端数量大于终端数阈值时，则根据终端上报的本轮模型训练的梯度，计算本轮模型训练的平均梯度；否则，将上一轮模型训练的平均梯度，作为本轮模型训练的平均梯度。中心节点根据本轮模型训练的平均梯度更新AI模型的参数，并通过基站向前述调度组中的终端，发送本轮模型训练的平均梯度。
步骤807：中心节点判断模型训练终止条件，并向对应终端发送终止模型训练指示。
关于模型训练终止条件，可以参见上述图5或图7流程的记载，不同的是，在图8的流程中，由中心节点确定是否满足模型训练终止条件，而上述图5或图7所示的流程中，由基站确定是否满足模型训练终止条件。可以理解的是，中心节点可将上述终止模型训练指示先发送给基站，由基站转发给终端等。
在上述图8所示的流程中，中心节点可以调度所有分组的终端参与联邦学习训练，或者也可以仅调度部分分组的终端参与联邦学习训练。在上述描述中，终端在接收到上述调度指示时，可以进行联邦学习训练，否则不进行联邦学习训练。可选的，终端还可以配置终端的监听参数，该参数中包括监听时刻和监听时长。在到达所述监听时刻时，按照配置的监听时长，监听调度指示。在到达监听时长时，则保持休眠，从而节省终端的功率。
在上述方案中，通过对终端分组，可以仅调度部分终端进行联邦学习训练，相对于必须所有终端参与联邦学习的方案，可以降低终端的功耗。
可以理解的是，为了实现上述方法中的功能，基站、终端和中心节点包括了执行各个功能相应的硬件结构和/或软件模块。本领域技术人员应该很容易意识到，结合本申请描述的各示例的单元及方法步骤，本申请能够以硬件或硬件和计算机软件相结合的形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行，取决于技术方案的特定应用场景和设计约束条件。
图9和图10为本申请提供的可能的通信装置的结构示意图。这些通信装置可以用于实现上述方法中终端、基站或中心节点的功能，因此也能实现上述方法所具备的有益效果。在本申请中，该通信装置实现终端的功能时，可以是如图1所示的终端120a-120j中的一个；该通信装置实现基站的功能时，可以是如图1所示的基站110a或110b，还可以是应用于终端或基站的模块(如芯片)。
如图9所示，通信装置900包括处理单元910和收发单元920。通信装置900用于实现上述图4、图5、图7或图8中所示的方法中终端、基站或中心节点的功能。
当通信装置900用于实现图4、图5、图7或图8所示的方法中基站的功能时：收发单元920用于向参与联邦学习的终端发送第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻均相同；以及，接收所述参与联邦学习的终端上报的梯度在空中叠加后的信号，所述梯度是所述终端在所述上报时刻，利用所述时频资源上报的在所述训练时长内训练完成的AI模型的梯度。处理单元910用于生成第一配置信息，和对终端上报的梯度进行处理。
当通信装置900用于实现图4、图5、图7或图8所示的方法中终端的功能时：收发单元920用于接收来自第二节点的第一配置信息，所述第一配置信息至少用于配置：训练 时长、时频资源、以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻相同；处理单元910用于在所述训练时长内，对AI模型进行训练，获得本轮模型训练的AI模型的梯度；收发单元920还用于在所述上报时刻，利用所述时频资源，向所述第二节点上报所述本轮模型训练的AI模型的梯度。
当通信装置900用于实现图4、图5、图7或图8所示的方法中的中心节点的功能时：处理单元910用于确定第二配置信息，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、组临时标识、训练时长、终端数阈值、传输块大小、或上行需求；收发单元920用于向第二节点发送所述第二配置信息。
有关上述处理单元910和收发单元920更详细的描述可以直接参考图4、图5、图7或图8所示的方法中相关描述直接得到，这里不加赘述。
如图10所示，通信装置1000包括处理器1010和接口电路1020。处理器1010和接口电路1020之间相互耦合。可以理解的是，接口电路1020可以为收发器或输入输出接口。可选的，通信装置1000还可以包括存储器1030，用于存储处理器1010执行的指令或存储处理器1010运行指令所需要的输入数据或存储处理器1010运行指令后产生的数据。
当通信装置1000用于实现上述方法时，处理器1010用于实现上述处理单元910的功能，接口电路1020用于实现上述收发单元920的功能。
当上述通信装置为应用于终端的芯片时，该终端芯片实现上述方法中终端的功能。该终端芯片从终端中的其它模块(如射频模块或天线)接收信息，该信息是基站发送给终端的；或者，该终端芯片向终端中的其它模块(如射频模块或天线)发送信息，该信息是终端发送给基站的。
当上述通信装置为应用于基站的模块时，该基站模块实现上述方法中基站的功能。该基站模块从基站中的其它模块(如射频模块或天线)接收信息，该信息是终端发送给基站的；或者，该基站模块向基站中的其它模块(如射频模块或天线)发送信息，该信息是基站发送给终端的。这里的基站模块可以是基站的基带芯片，也可以是DU或其他模块，这里的DU可以是开放式无线接入网(open radio access network，O-RAN)架构下的DU。
当上述装置为应该于中心节点的模块时，该中心节点实现上述方法中的中心节点的功能。该中心节点从中心节点的其它模块(如射频模块或天线)接收信息，该信息是基站发送给中心节点的；或者，该中心节点模块向中心节点中的其它模块(如射频模块或天线)发送信息，该信息是中心节点发送给基站的。这里的中心节点模块可以是中心节点的基带芯片或其他模块。
可以理解的是，本申请中的处理器可以是中央处理单元(central processing unit，CPU)，还可以是其它通用处理器、数字信号处理器(digital signal processor，DSP)、专用集成电路(application specific integrated circuit，ASIC)、现场可编程门阵列(field programmable gate array，FPGA)或者其它可编程逻辑器件、晶体管逻辑器件，硬件部件或者其任意组合。通用处理器可以是微处理器，也可以是任何常规的处理器。
本申请中的存储器可以是随机存取存储器、闪存、只读存储器、可编程只读存储器、可擦除可编程只读存储器、电可擦除可编程只读存储器、寄存器、硬盘、移动硬盘、CD-ROM或者本领域熟知的任何其它形式的存储介质。
一种示例性的存储介质耦合至处理器，从而使处理器能够从该存储介质读取信息，且可向该存储介质写入信息。存储介质也可以是处理器的组成部分。处理器和存储介质可以 位于ASIC中。另外，该ASIC可以位于基站或终端中。当然，处理器和存储介质也可以作为分立组件存在于基站或终端中。
本申请中的方法可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机程序或指令。在计算机上加载和执行所述计算机程序或指令时，全部或部分地执行本申请所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、网络设备、用户设备、核心网设备、OAM或者其它可编程装置。所述计算机程序或指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机程序或指令可以从一个网站站点、计算机、服务器或数据中心通过有线或无线方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是集成一个或多个可用介质的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，例如，软盘、硬盘、磁带；也可以是光介质，例如，数字视频光盘；还可以是半导体介质，例如，固态硬盘。该计算机可读存储介质可以是易失性或非易失性存储介质，或可包括易失性和非易失性两种类型的存储介质。
在本申请中，如果没有特殊说明以及逻辑冲突，不同的实施例之间的术语和/或描述具有一致性、且可以相互引用，不同的实施例中的技术特征根据其内在的逻辑关系可以组合形成新的实施例。
本申请中，“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B的情况，其中A，B可以是单数或者复数。在本申请的文字描述中，字符“/”，一般表示前后关联对象是一种“或”的关系；在本申请的公式中，字符“/”，表示前后关联对象是一种“相除”的关系。“包括A，B或C中的至少一个”可以表示：包括A；包括B；包括C；包括A和B；包括A和C；包括B和C；包括A、B和C。
可以理解的是，在本申请中涉及的各种数字编号仅为描述方便进行的区分，并不用来限制本申请的范围。上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定。

Claims (37)

  1. 一种无线网络中的人工智能AI模型训练方法，其特征在于，包括：
    向参与联邦学习的终端发送第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻均相同；
    接收所述参与联邦学习的终端上报的梯度在空中叠加后的信号，所述梯度是所述终端在所述上报时刻，利用所述时频资源上报的在所述训练时长内训练完成的AI模型的梯度。
  2. 如权利要求1所述的方法，其特征在于，还包括：
    接收来自所述终端的训练完成指示，所述训练完成指示是所述终端在所述训练时长内，对所述AI模型训练完成时，向第二节点发送的；
    根据所述终端发送的训练完成指示，统计在所述训练时长内，完成AI模型训练的终端数量。
  3. 如权利要求2所述的方法，其特征在于，还包括：
    若所述完成AI模型训练的终端数量大于或等于终端数阈值，则根据参与联邦学习的终端上报的梯度，确定本轮模型训练的平均梯度；否则，将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度；
    根据所述本轮模型训练的平均梯度更新所述AI模型的参数，并向所述终端发送所述本轮模型训练的平均梯度。
  4. 如权利要求2所述的方法，其特征在于，还包括：
    向第一节点发送在所述训练时长内，完成AI模型训练的终端数量，以及所述终端上报的梯度在空中叠加的信号。
  5. 如权利要求1至4中任一项所述的方法，其特征在于，所述第一配置信息还用于配置以下至少一项：
    专用的承载RB资源、调制方式、初始AI模型、或发射功率。
  6. 如权利要求5所述的方法，其特征在于，确定所述发射功率的过程，包括：
    根据对来自所述终端的探测参考信号SRS进行测量，确定所述终端的上行信道质量；
    根据所述上行信道质量，确定所述终端的发射功率。
  7. 如权利要求1至4中任一项所述的方法，其特征在于，所述第一配置信息还用于配置以下至少一项：
    专用的承载RB资源、调制方式、初始AI模型、信道状态信息CSI区间、或信道反转参数。
  8. 如权利要求1至7中任一项所述的方法，其特征在于，还包括：
    接收来自第一节点的第二配置信息，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、组临时标识、训练时长、终端数阈值、传输块大小、或上行需求。
  9. 如权利要求1至8中任一项所述的方法，其特征在于，还包括：
    接收来自所述终端的第一终端信息，向第一节点发送第二终端信息；
    其中，所述第一终端信息中包括以至少一项：终端的通信能力、终端的计算能力、或终端的数据集特征；所述第二终端信息中包括以下至少一项：终端的通信能力、终端的计 算能力、终端的数据集特征、或终端临时标识，所述终端临时标识是所述第二节点分配给所述终端的。
  10. 如权利要求1至9中任一项所述的方法，其特征在于，还包括：
    在满足模型训练终止条件时，向所述终端发送终止模型训练指示；或者，
    接收来自第一节点的终止模型训练指示，向所述终端转发所述终止模型训练指示。
  11. 一种无线网络中的人工智能AI模型训练方法，其特征在于，包括：
    接收来自第二节点的第一配置信息，所述第一配置信息至少用于配置：训练时长、时频资源、以及上报时刻；其中，为参与联邦学习的不同终端所配置的训练时长、时频资源和上报时刻相同；
    在所述训练时长内，对AI模型进行训练，获得本轮模型训练的AI模型的梯度；
    在所述上报时刻，利用所述时频资源，向所述第二节点上报所述本轮模型训练的AI模型的梯度。
  12. 如权利要求11所述的方法，其特征在于，还包括：
    在所述训练时长结束时，若所述AI模型完成训练，则向所述第二节点发送训练完成指示。
  13. 如权利要求11或12所述的方法，其特征在于，还包括：
    若在所述训练时长内未完成所述AI模型训练，则终止对所述AI模型的训练。
  14. 如权利要求11至13中任一项所述的方法，其特征在于，还包括：
    接收来自第二节点的上一轮模型训练的平均梯度；
    根据所述上一轮模型训练的平均梯度，更新本轮模型训练中所述AI模型的梯度；或者，
    根据所述本轮模型训练的平均梯度，在所述本轮模型训练中更新所述AI模型的参数和梯度。
  15. 如权利要求11至14中任一项所述的方法，其特征在于，所述第一配置信息还用于配置以下至少一项：
    专用的承载RB资源、调制方式、初始AI模型或发射功率。
  16. 如权利要求11至14中任一项所述的方法，其特征在于，所述第一配置信息还用于配置以下至少一项：
    专用的承载RB资源、调制方式、初始AI模型、信道状态信息CSI区间、或信道反转参数。
  17. 如权利要求16所述的方法，其特征在于，所述第一配置信息还用于配置所述信道状态信息CSI区间和信道反转参数时，所述方法还包括：
    若下行信道和上行信道所配置的频率资源相同，则根据测量的下行信道的CSI，确定所述终端的上行信道的CSI；
    若所述上行信道的CSI满足CSI区间的要求，则根据所述信道反转参数，确定发射功率；
    向所述第二节点上报所述本轮模型训练的AI模型的梯度，包括：
    基于确定的所述发射功率，向所述第二节点上报所述本轮模型训练的AI模型的梯度。
  18. 如权利要求11至17中任一项所述的方法，其特征在于，所述第一配置信息中还包括组临时标识，所述组临时标识是第一节点为所述终端分配的组临时标识。
  19. 如权利要求18所述的方法，其特征在于，还包括：
    接收来自所述第二节点的调度指示，所述调度指示中包括组临时标识；
    当所述调度指示中包括的组临时标识，与所述第一节点为所述终端分配的组临时标识相同时，则在本轮模型训练中执行AI模型训练；否则，在本轮模型训练中不再执行AI模型训练。
  20. 如权利要求11至19中任一项所述的方法，其特征在于，还包括：
    接收来自所述第二节点的终止模型训练指示；
    根据所述终止模型训练指示，终止对所述AI模型的训练。
  21. 如权利要求11至20中任一项所述的方法，其特征在于，还包括：
    向所述第二节点发送第一终端信息，所述第一终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、或终端的数据集特征。
  22. 一种无线网络中的人工智能AI模型训练方法，其特征在于，包括：
    确定第二配置信息，所述第二配置信息用于配置以下至少一项：参与联邦学习的终端列表、初始AI模型、组临时标识、训练时长、终端数阈值、传输块大小、或上行需求；
    向第二节点发送所述第二配置信息。
  23. 如权利要求22所述的方法，其特征在于，还包括：
    接收来自所述第二节点的第二终端信息，所述第二终端信息中包括以下至少一项：终端的通信能力、终端的计算能力、终端的数据集特征、或终端临时标识，所述终端临时标识是所述第二节点分配给所述终端的；
    根据所述终端信息，确定参与联邦学习的终端列表。
  24. 如权利要求22或23所述的方法，其特征在于，还包括：
    接收来自所述第二节点的所述参与联邦学习的终端上报的梯度在空中叠加的信号以及在所述训练时长内完成AI模型训练的终端数量；
    若在所述训练时长内完成模型训练的终端数量大于或等于所述终端数阈值，则根据所述终端上报的AI模型的梯度，确定本轮模型训练的平均梯度；否则，将上一轮模型训练的平均梯度作为本轮模型训练的平均梯度；
    根据所述本轮模型训练的平均梯度更新所述AI模型的参数，并向所述第二节点发送所述本轮模型训练的平均梯度，以使得所述第二节点将所述本轮模型训练的平均梯度发送给所述终端。
  25. 如权利要求22至24中任一项所述的方法，其特征在于，还包括：
    向所述第二节点发送调度指示，所述调度指示中包括组临时标识，所述调度指示用于调度组临时标识对应的终端，在本轮模型训练内执行AI模型训练。
  26. 如权利要求22至25中任一项所述的方法，其特征在于，还包括：
    在满足模型训练终止条件时，向所述第二节点发送终止模型训练指示，用于指示所述终端在本轮模型训练中，停止对AI模型的训练。
  27. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括用于实现权利要求1至10中任一项所述方法的单元。
  28. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括处理器和存储器，所述处理器用于实现权利要求1至10中任一项所述的方法。
  29. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括用于实现权利要 求11至21中任一项所述的方法的单元。
  30. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括处理器和存储器，所述处理器用于实现权利要求11至21中任一项所述的方法。
  31. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括用于实现权利要求22至26中任一项所述的方法的单元。
  32. 一种无线网络中的人工智能AI模型训练装置，其特征在于，包括处理器和存储器，所述处理器用于实现权利要求22至26中任一项所述的方法。
  33. 一种系统，其特征在于，包括权利要求27或28所述的装置，和权利要求29或30所述的装置。
  34. 如权利要求33所述的系统，其特征在于，还包括权利要求31或32所述的装置。
  35. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有指令，当所述指令在计算机上运行时，使得计算机执行权利要求1至10中任一项所述的方法，或者权利要求11至21中任一项所述的方法，或者权利要求22至26中任一项所述的方法。
  36. 一种计算机程序产品，其特征在于，包括指令，当所述指令在计算机上运行时，使得计算机执行权利要求1至10中任一项所述的方法，或者权利要求11至21中任一项所述的方法，或者权利要求22至26中任一项所述的方法。
  37. 一种芯片系统，其特征在于，包括：处理器或电路，用于执行权利要求1至10中任一项所述的方法，或者用于执行权利要求11至21中任一项所述的方法，或者用于执行权利要求22至26中任一项所述的方法。
PCT/CN2022/137671 2021-12-10 2022-12-08 一种无线网络中的人工智能ai模型训练方法及装置 WO2023104169A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111505116.7A CN116264684A (zh) 2021-12-10 2021-12-10 一种无线网络中的人工智能ai模型训练方法及装置
CN202111505116.7 2021-12-10

Publications (1)

Publication Number Publication Date
WO2023104169A1 true WO2023104169A1 (zh) 2023-06-15

Family

ID=86721643

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/137671 WO2023104169A1 (zh) 2021-12-10 2022-12-08 一种无线网络中的人工智能ai模型训练方法及装置

Country Status (2)

Country Link
CN (1) CN116264684A (zh)
WO (1) WO2023104169A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116546567A (zh) * 2023-07-06 2023-08-04 深圳市大数据研究院 基于贝叶斯联邦学习的数据处理方法、系统及电子设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111522669A (zh) * 2020-04-29 2020-08-11 深圳前海微众银行股份有限公司 横向联邦学习系统优化方法、装置、设备及可读存储介质
CN113139663A (zh) * 2021-04-23 2021-07-20 深圳市大数据研究院 联邦边缘学习配置信息获取方法、装置、设备及介质
CN113139662A (zh) * 2021-04-23 2021-07-20 深圳市大数据研究院 联邦学习的全局及局部梯度处理方法、装置、设备和介质
CN113516250A (zh) * 2021-07-13 2021-10-19 北京百度网讯科技有限公司 一种联邦学习方法、装置、设备以及存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111522669A (zh) * 2020-04-29 2020-08-11 深圳前海微众银行股份有限公司 横向联邦学习系统优化方法、装置、设备及可读存储介质
CN113139663A (zh) * 2021-04-23 2021-07-20 深圳市大数据研究院 联邦边缘学习配置信息获取方法、装置、设备及介质
CN113139662A (zh) * 2021-04-23 2021-07-20 深圳市大数据研究院 联邦学习的全局及局部梯度处理方法、装置、设备和介质
CN113516250A (zh) * 2021-07-13 2021-10-19 北京百度网讯科技有限公司 一种联邦学习方法、装置、设备以及存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116546567A (zh) * 2023-07-06 2023-08-04 深圳市大数据研究院 基于贝叶斯联邦学习的数据处理方法、系统及电子设备
CN116546567B (zh) * 2023-07-06 2023-11-24 深圳市大数据研究院 基于贝叶斯联邦学习的数据处理方法、系统及电子设备

Also Published As

Publication number Publication date
CN116264684A (zh) 2023-06-16

Similar Documents

Publication Publication Date Title
Wu et al. Common throughput maximization in UAV-enabled OFDMA systems with delay consideration
CN109547174A (zh) 一种时间配置的方法、网络设备及ue
WO2018072905A1 (en) Resource allocation and scheduling for wireless networks with self-backhauled links
WO2023036268A1 (zh) 一种通信方法及装置
US20230077500A1 (en) Communication Method and Apparatus
WO2023104169A1 (zh) 一种无线网络中的人工智能ai模型训练方法及装置
Zhao et al. Dynamic and non-centric networking approach using virtual gateway platforms for low power wide area systems
WO2021068259A1 (zh) 联合调度的方法和装置
WO2023040971A1 (zh) 一种人工智能模型下载方法、装置及系统
US20200404743A1 (en) Method and apparatus for controlling packet duplication transmission in wireless communication system
Wang et al. Graph-based file dispatching protocol with D2D-aided UAV-NOMA communications in large-scale networks
WO2023036280A1 (zh) 一种模型测试方法及装置
WO2023030221A1 (zh) 一种通信方法及装置
WO2022247739A1 (zh) 一种数据传输方法和相关装置
WO2023125331A1 (zh) 通信方法和通信装置
WO2023088396A1 (zh) 通信方法及装置
WO2024031615A1 (en) Techniques for binary wake-up signal sequences
WO2024067245A1 (zh) 模型匹配的方法和通信装置
WO2024103973A1 (zh) 通信方法和通信装置
WO2024082159A1 (en) Paging techniques for wireless devices with mixed capabilities
WO2023060417A1 (en) Ue clustering in fl model update reporting
CN110351026A (zh) 数据传输方法、终端和网络设备
US20240020513A1 (en) Data processing method and apparatus
WO2024032230A1 (zh) 下行控制信道传输方法、装置及系统
WO2024036594A1 (en) Methods for discovery and signaling procedure for clustered and peer-to-peer federated learning

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: 22903587

Country of ref document: EP

Kind code of ref document: A1