WO2023088465A1 - 一种模型训练方法及相关装置 - Google Patents

一种模型训练方法及相关装置 Download PDF

Info

Publication number
WO2023088465A1
WO2023088465A1 PCT/CN2022/133214 CN2022133214W WO2023088465A1 WO 2023088465 A1 WO2023088465 A1 WO 2023088465A1 CN 2022133214 W CN2022133214 W CN 2022133214W WO 2023088465 A1 WO2023088465 A1 WO 2023088465A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
communication device
training
parameter
resources
Prior art date
Application number
PCT/CN2022/133214
Other languages
English (en)
French (fr)
Inventor
乔云飞
李榕
王坚
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 WO2023088465A1 publication Critical patent/WO2023088465A1/zh

Links

Images

Classifications

    • 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
    • G06N3/045—Combinations of networks
    • 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
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • G06N3/098—Distributed learning, e.g. federated learning

Definitions

  • the present application relates to the field of neural networks, in particular to a model training method and related devices.
  • 3GPP has introduced artificial intelligence (AI) capabilities in the 5th generation mobile communication (5G) network by adding a new network data analysis function (NWDAF).
  • NWDAF is responsible for the training of AI models.
  • the AI model trained by NWDAF can be applied to the network's own fields such as mobility management, session management and network automation.
  • federated learning FL
  • FL federated learning
  • each distributed node participates in each round of training of the central node, it needs to send the local neural network model updated in the previous round to the central node.
  • the central node fuses the neural network models of the distribution nodes to obtain the global neural network model. If the global neural network model does not converge, the central node broadcasts the global neural network model to each distribution node.
  • Each distribution node updates the local neural network model according to the global neural network model, and then uses the updated local neural network model to participate in the next round of training of the central node's neural network model.
  • Embodiments of the present application provide a model training method and a related device, which can reduce signaling overhead.
  • the embodiment of the present application provides a model training method.
  • the second communication device receives the first neural network parameter of the first communication device, and when the correlation coefficient between the first neural network parameter and the second neural network parameter of the second communication device is less than a first threshold, sends The first communication device sends first indication information.
  • the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the first communication device.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is smaller than the first threshold, indicating that the second neural network parameter makes a greater contribution to the convergence of the first neural network model. Therefore, the second communication device determines whether to participate in the training of the first neural network model according to the contribution of the second neural network parameters to the convergence of the first neural network model, which can prevent the second communication device from affecting the first neural network model by the second neural network parameters. When the convergence contribution of the network model is small, it still participates in the training of the first neural network model, thereby reducing the signaling overhead of the second communication device.
  • the first neural network parameter is the model parameter of the first neural network or the gradient of the first neural network
  • the second neural network parameter is the model parameter of the second neural network or the gradient of the second neural network .
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network.
  • the first neural network parameter is the gradient of the neural network of the first communication device
  • the second neural network parameter is the gradient of the neural network of the second communication device. Therefore, the second communication device determines the correlation coefficient between the first neural network parameter and the second neural network parameter according to the type of the received first neural network parameter.
  • the first neural network parameter is received on a cooperative discovery resource, and the cooperative discovery resource is configured in sidelink configuration information. That is to say, the second communication device receives the first neural network parameters from the first communication device by using the cooperative discovery resource in the sidelink configuration information.
  • the second communication device may also send the second neural network parameters to the first communication device, so that the first communication device Updating the first neural network model, so that the first communication device updates the first neural network model by using the second neural network parameters that contribute more to the convergence of the first neural network model, is beneficial to speed up the convergence of the first neural network model.
  • the second communication device may also receive a control signal from the first communication device, where the control signal is used to indicate a time-frequency resource, and the indicated time-frequency resource is used by the second communication device to send the second Neural Network Parameters. It can be seen that the second communication device obtains the time-frequency resource for sending the second neural network parameters to the first communication device by receiving the control signal from the first communication device, and then the second communication device can send the second neural network parameter on the time-frequency resource. Neural Network Parameters.
  • the resource receiving the control signal is a cooperative control resource
  • the cooperative control resource is configured in the above-mentioned sidelink configuration information. That is to say, the second communication device receives the above control signal by using the coordinated control resource in the sidelink configuration information.
  • the second communication device may also receive a synchronization signal on the coordinated synchronization resource, and perform synchronization with the first communication device according to the synchronization signal. Therefore, after the second communication device is synchronized with the first communication device, it can communicate with the first communication device.
  • the cooperative synchronization resource may be configured in the above sidelink configuration information.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information may be pre-configured, or dynamically indicated, or unlicensed spectrum resources.
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network
  • the correlation coefficient between is determined according to the first parameter and the second parameter.
  • the first parameter is a parameter output by the first neural network model when the second communication device inputs training data to the first neural network model; the first neural network model is determined according to the model parameters of the first neural network ;
  • the second parameter is a parameter output by the second neural network model when the second communication device inputs the training data to the second neural network model of the second communication device. That is to say, the first parameter and the second parameter are respectively output by the first neural network model and the second neural network model when the second communication device inputs the same training data to the first neural network model and the second neural network model respectively. parameter.
  • the relationship between the first neural network parameter and the second neural network parameter is determined according to the probability density distribution of the first neural network parameters and the probability density distribution of the second neural network parameters.
  • the second communication device can flexibly determine the correlation coefficient between the first neural network parameter and the second neural network parameter in a corresponding manner according to the type of the received first neural network parameter.
  • the present application also provides a model training method.
  • the model training method in this aspect corresponds to the model training method in the first aspect, and the model training method in this aspect is explained from the side of the first communication device.
  • the first communication device sends the first neural network parameters of the first communication device.
  • the first communication device receives the first indication information from the second communication device, the first indication information is that the correlation coefficient between the first neural network parameter of the second communication device and the second neural network parameter of the second communication device is smaller than the first
  • the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the first communication device.
  • the first indication information received by the first communication device is sent by the second communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter is less than the first threshold, so that the second communication The device determines whether to participate in the training of the first neural network model according to the contribution of the second neural network parameters to the convergence of the first neural network model, so that the first communication device does not follow up the second neural network parameters based on all the second neural network parameters of the second communication device. Updating the first neural network model, but updating the first neural network model according to the second neural network parameters that contribute more to the convergence of the first neural network model can reduce the signaling overhead of the first communication device.
  • the first neural network parameter is a model parameter of the first neural network or the gradient of the first neural network
  • the second neural network parameter is a model parameter of the second neural network or the gradient of the first neural network. Gradients of the second neural network.
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network.
  • the first neural network parameter is the gradient of the neural network of the first communication device
  • the second neural network parameter is the gradient of the neural network of the second communication device.
  • the first neural network parameter is sent on a cooperative discovery resource, and the cooperative discovery resource is configured in sidelink configuration information. That is to say, the first communication device sends the first neural network parameters to the first communication device by using the cooperative discovery resource in the sidelink configuration information.
  • the first communication device may also receive second neural network parameters from the second communication device, and update the first neural network model according to the second neural network parameters. It can be seen that the first communication device updates the first neural network model according to the second neural network parameters of the second communication device fed back the first indication information, so that the signaling overhead of the first communication device can be saved.
  • the first communication device may also send a control signal to the second communication device, where the control signal is used to indicate time-frequency resources, and the indicated time-frequency resources are used by the second communication device to send the second neural network parameter. It can be seen that the first communication device indicates to the second communication device the time-frequency resources for sending the second neural network parameters through the control signal, which is beneficial for the second communication device to use the time-frequency resources to send the second neural network parameters.
  • the resource for sending the control signal is a cooperative control resource
  • the cooperative control resource is configured in the above sidelink configuration information. That is to say, the first communication device uses the coordinated control resource in the sidelink configuration information to send the foregoing control signal.
  • the first communication device may also send a synchronization signal on the coordinated synchronization resource, so that the second communication device synchronizes with the first communication device according to the synchronization signal.
  • the coordinated synchronization resource may be configured in the above sidelink configuration information.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information may be pre-configured, or dynamically indicated, or unlicensed spectrum resources.
  • the present application also provides a model training method.
  • the first communication device sends cooperation request information, and the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained.
  • the first communication device receives second instruction information from the second communication device, the second instruction information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is one or more of the plurality of training tasks indivual.
  • the first communication device divides the neural network model to be trained into multiple training tasks, and broadcasts the multiple training tasks to the surrounding second communication devices through the cooperation request information, so as to request each The second communication device participates in training for a plurality of training tasks.
  • the first communication device learns the training tasks that the second communication device itself can participate in by receiving the second indication information. In this manner, the first communication device learns that the surrounding second communication device assists in training the neural network model to be trained, thereby reducing the requirement on the capability of the first communication device.
  • the above-mentioned cooperation request information is sent on the cooperation discovery resource, and the cooperation discovery resource is configured in the sidelink configuration information. It can be seen that the first communication device uses the cooperation discovery resource in the sidelink configuration information to send the cooperation request information.
  • the first training task indicated by the second indication information sent by the second communication device includes multiple training tasks.
  • the first communication device may also send third indication information, and the third indication information uses to indicate one of the training tasks in the first training tasks.
  • the first communication device determines the third indication information according to the training tasks indicated by the received second indication information, so as to ensure that the training tasks trained by each second communication device participating in the training are not repeated.
  • the first training task indicated by the second indication information sent by multiple second communication devices is the same training task among the multiple training tasks.
  • the first communication device may also pass the third The instruction information indicates to one of the second communication devices the training task to participate in the training. Therefore, the second communication device that has received the third indication information knows the training task that needs to be trained, while the second communication device that has not received the third indication information does not participate in the training.
  • the first communication device may also send fourth indication information to the second communication device, and the fourth indication information is used to indicate the first output to be received by the second communication device and the time corresponding to the first output.
  • the first output is the output of the neural network model trained by the first communication device, or the output of the neural network model trained by other second communication devices except the second communication device; the second output is the neural network trained by the second communication device The output of the model.
  • the first communication device notifies the second communication device participating in the training of the parameters to be received, the location of the time-frequency resource corresponding to the parameters to be received, and/or the parameters to be sent and the corresponding The location of the time-frequency resource, so that any second communication device participating in the training can receive and/or send the corresponding output during the training process of the training task, so as to ensure the cooperative training of other second communication devices.
  • the resource for sending the fourth indication information is a cooperative control resource
  • the cooperative control resource is configured in sidelink configuration information. It can be seen that the first communication device sends the fourth indication information by using the cooperative control resource in the sidelink configuration information.
  • the first communication device may also send a synchronization signal on the coordinated synchronization resource, so that the second communication device synchronizes with the first communication device according to the synchronization signal.
  • the coordinated synchronization resource is configured in sidelink configuration information.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information may be pre-configured, or dynamically indicated, or unlicensed spectrum resources.
  • the present application also provides a model training method.
  • the model training method in this aspect corresponds to the model training method described in the third aspect.
  • the model training method in this aspect is explained from the side of the second communication device .
  • the second communication device receives the cooperation request information, and the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained.
  • the second communication device determines to participate in the training of the first training task, it sends the second indication information, the second indication information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is one of the multiple training tasks one or more of them.
  • the second communication device when it determines to participate in the training of the first training task among the multiple training tasks requested by the first communication device, it sends indication information indicating participation in the first training task to inform the second communication device
  • a communication device itself can assist the first communication device to participate in the training of the first training task, thereby reducing the demand on the capabilities of the first communication device.
  • the cooperation request information is received on the cooperation discovery resource, and the cooperation discovery resource is configured in the sidelink configuration information. It can be seen that the second communication device receives the cooperation request information by using the cooperation discovery resource in the sidelink configuration information.
  • the second communication device may also receive third indication information, and the third indication information is used to indicate one of the training tasks in the first training tasks, so that the second communication device learns about the training tasks participating in the training .
  • the second communication device may also receive fourth indication information, where the fourth indication information is used to indicate the first output received by the second communication device, the location of the time-frequency resource corresponding to the first output, and/or Or the sent second output, and the time-frequency resource position corresponding to the second output.
  • the first output is the output of the neural network model trained by the first communication device, or the output of the neural network model trained by other second communication devices except the second communication device; the second output is the neural network trained by the second communication device The output of the model. Therefore, during the training process of the training task, the second communication device receives and/or sends the corresponding output, so as to ensure the cooperative training of other second communication devices.
  • the resources receiving the fourth indication information are cooperative control resources, and the cooperative control resources are configured in sidelink configuration information. It can be seen that the second communication device receives the fourth indication information by using the cooperative control resource in the sidelink configuration information.
  • the second communication device may also receive a synchronization signal on the coordinated synchronization resource, and perform synchronization with the first communication device according to the synchronization signal. Therefore, after the second communication device is synchronized with the first communication device, it can communicate with the first communication device.
  • the coordinated synchronization resources are configured in the above sidelink configuration information.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information may be pre-configured, or dynamically indicated, or unlicensed spectrum resources.
  • the present application further provides a communication device.
  • the communication device has part or all of the functions of the second communication device described in the first aspect above, or part or all of the functions of the first communication device described in the second aspect above, or has the function of realizing the above third aspect.
  • the function of the communication device may have the functions of some or all embodiments of the second communication device described in the first aspect of the present application, or may have the function of independently implementing any one embodiment of the present application.
  • the functions described above may be implemented by hardware, or may be implemented by executing corresponding software on the hardware.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the structure of the communication device may include a processing unit and a communication unit, and the processing unit is configured to support the communication device to perform corresponding functions in the foregoing method.
  • the communication unit is used to support communication between the communication device and other communication devices.
  • the communication device may further include a storage unit, which is used to be coupled with the processing unit and the communication unit, and stores necessary program instructions and data of the communication device.
  • the communication device includes: a processing unit and a communication unit, the processing unit is used to control the communication unit to send and receive data/signaling; the communication unit is used to receive the first neural network parameters of the first communication device; The unit is further configured to send first indication information to the first communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter of the communication device is less than the first threshold; the first indication information is used to indicate the The communication device participates in the training of the first neural network model of the first communication device.
  • the communication device includes: a processing unit and a communication unit, the processing unit is used to control the communication unit to send and receive data/signaling; the communication unit is used to send the first neural network parameters of the communication device; The communication unit is also used to receive first indication information from the second communication device; the first indication information is that the correlation coefficient between the first neural network parameter of the second communication device and the second neural network parameter of the second device is less than The first threshold is sent; the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the communication device.
  • the communication device includes: a processing unit and a communication unit, the processing unit is used to control the communication unit to send and receive data/signaling; the communication unit is used to send cooperation request information, and the cooperation request information includes a plurality of Training tasks, multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained; the communication unit is also used to receive second instruction information from the second communication device, the second instruction information is used to The second communication device is instructed to participate in the training of the first training task, where the first training task is one or more of the plurality of training tasks.
  • the communication device includes: a processing unit and a communication unit, the processing unit is used to control the communication unit to send and receive data/signaling; the communication unit is used to receive cooperation request information, and the cooperation request information includes a plurality of Training tasks, multiple training tasks are obtained by splitting the neural network model to be trained by the first communication device; the communication unit is also used to send second instruction information when determining to participate in the training of the first training task, the second The instruction information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is one or more of the multiple training tasks.
  • the communication unit may be a transceiver or a communication interface
  • the storage unit may be a memory
  • the processing unit may be a processor
  • the communication device includes: a processor and a transceiver, the processor is used to control the transceiver to transmit and receive data/signaling; the transceiver is used to receive the first neural network parameters of the first communication device; the transceiver , is also used to send first indication information to the first communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter of the communication device is less than the first threshold; the first indication information is used to indicate the communication The device participates in the training of the first neural network model of the first communication device.
  • the communication device includes: a processor and a transceiver, the processor is used to control the transceiver to send and receive data/signaling; the transceiver is used to send the first neural network parameters of the communication device; The device is also used to receive first indication information from the second communication device; the first indication information is that the correlation coefficient between the first neural network parameter of the second communication device and the second neural network parameter of the second device is less than the first The threshold is sent; the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the communication device.
  • the communication device includes: a processor and a transceiver, the processor is used to control the transceiver to send and receive data/signaling; the transceiver is used to send cooperation request information, and the cooperation request information includes multiple training tasks , the multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained; the transceiver is also used to receive second instruction information from the second communication device, the second instruction information is used to indicate the second The communication device participates in the training of the first training task, and the first training task is one or more of the plurality of training tasks.
  • the communication device includes: a processor and a transceiver, the processor is used to control the transceiver to perform data/signaling transmission and reception; the transceiver is used to receive cooperation request information, and the cooperation request information includes a plurality of training task, a plurality of training tasks are obtained by splitting the neural network model to be trained by the first communication device; the transceiver is also used to send the second indication information when it is determined to participate in the training of the first training task, the second indication information It is used to instruct the second communication device to participate in the training of the first training task, where the first training task is one or more of the multiple training tasks.
  • the communication device is a chip or a chip system.
  • the processing unit may also be embodied as a processing circuit or a logic circuit; the transceiver unit may be an input/output interface, interface circuit, output circuit, input circuit, pin or related circuit on the chip or chip system.
  • the processor may be used to perform, for example but not limited to, baseband related processing
  • the transceiver may be used to perform, for example but not limited to, radio frequency transceiving.
  • the above-mentioned devices may be respectively arranged on independent chips, or at least partly or all of them may be arranged on the same chip.
  • processors can be further divided into analog baseband processors and digital baseband processors.
  • the analog baseband processor can be integrated with the transceiver on the same chip, and the digital baseband processor can be set on an independent chip.
  • a digital baseband processor can be integrated with various application processors (such as but not limited to graphics processors, multimedia processors, etc.) on the same chip.
  • application processors such as but not limited to graphics processors, multimedia processors, etc.
  • SoC System on a Chip
  • the present application further provides a processor configured to execute the above various methods.
  • the process of sending the above information and receiving the above information in the above method can be understood as the process of outputting the above information by the processor and the process of receiving the input of the above information by the processor.
  • the processor When outputting the above information, the processor outputs the above information to the transceiver for transmission by the transceiver. After the above information is output by the processor, other processing may be required before reaching the transceiver.
  • the processor receives the above-mentioned input information
  • the transceiver receives the above-mentioned information and inputs it to the processor. Furthermore, after the transceiver receives the above information, the above information may need to be processed before being input to the processor.
  • the receiving of the first neural network parameters of the first communication device mentioned in the aforementioned method may be understood as inputting the first neural network parameters of the first communication device by the processor.
  • the above-mentioned processor may be a processor dedicated to performing these methods, or may be a processor that executes computer instructions in a memory to perform these methods, such as a general-purpose processor.
  • the above-mentioned memory can be a non-transitory (non-transitory) memory, such as a read-only memory (Read Only Memory, ROM), which can be integrated with the processor on the same chip, or can be respectively arranged on different chips.
  • ROM read-only memory
  • the embodiment does not limit the type of the memory and the arrangement of the memory and the processor.
  • the present application further provides a communication system, which includes at least one first communication device and at least two second communication devices according to the above aspect.
  • the system may further include other devices that interact with the first communication device and the second communication device in the solutions provided in this application.
  • the present application provides a computer-readable storage medium for storing instructions, and when the instructions are executed by a computer, the method described in any one of the first aspect to the fourth aspect above is realized.
  • the present application further provides a computer program product including instructions, which, when run on a computer, implement the method described in any one of the first aspect to the fourth aspect above.
  • the present application provides a chip system
  • the chip system includes a processor and an interface, the interface is used to acquire programs or instructions, and the processor is used to call the programs or instructions to implement or support the second communication
  • the device implements the functions involved in the first aspect, or is used to invoke the program or instruction to implement or support the first communication device to implement the function involved in the second aspect, and is used to invoke the program or instruction to implement or support the first communication device.
  • a communication device implements the functions involved in the third aspect, and is used to call the program or instruction to implement or support the second communication device to implement the functions involved in the fourth aspect. For example, at least one of the data and information involved in the above methods is determined or processed.
  • the chip system further includes a memory, and the memory is configured to store necessary program instructions and data of the terminal.
  • the system-on-a-chip may consist of chips, or may include chips and other discrete devices.
  • FIG. 1 is a schematic structural diagram of a communication system provided by an embodiment of the present application.
  • FIG. 2 is a schematic structural diagram of another communication system provided by an embodiment of the present application.
  • FIG. 3 is a schematic diagram of a federated learning system provided by an embodiment of the present application.
  • Fig. 4 is a schematic diagram of a segmentation learning provided by an embodiment of the present application.
  • FIG. 5 is a schematic diagram of an interaction process of a model training method provided in an embodiment of the present application.
  • FIG. 6 is a schematic diagram of an interaction process of another model training method provided by an embodiment of the present application.
  • Fig. 7 is a schematic diagram of an interaction flow of another model training method provided by the embodiment of the present application.
  • Fig. 8 is a schematic diagram of splitting a neural network model to be trained provided by the embodiment of the present application.
  • FIG. 9 is a schematic diagram of an interaction process of another model training method provided by the embodiment of the present application.
  • FIG. 10 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • Fig. 11 is a schematic structural diagram of another communication device provided by an embodiment of the present application.
  • FIG. 12 is a schematic structural diagram of a chip provided by an embodiment of the present application.
  • a wireless communication system may include one or more network devices and one or more terminal devices.
  • the wireless communication system can also perform point-to-point communication, such as communication between multiple terminal devices.
  • the wireless communication systems mentioned in the embodiments of the present application include but are not limited to: narrow band-internet of things system (narrow band-internet of things, NB-IoT), long term evolution system (long term evolution, LTE), 5G mobile communication system
  • NB-IoT narrow band-internet of things
  • LTE long term evolution
  • 5G mobile communication system Three application scenarios of the system: enhanced mobile broadband (eMBB), ultra reliable low latency communication (URLLC) and massive machine type of communication (mMTC), wireless security True (wireless fidelity, WiFi) system, or mobile communication system after 5G, etc.
  • eMBB enhanced mobile broadband
  • URLLC ultra reliable low latency communication
  • mMTC massive machine type of communication
  • WiFi wireless security True
  • mobile communication system after 5G etc.
  • FIG. 2 is a schematic structural diagram of another communication system provided by an embodiment of the present application.
  • the communication system may include but not limited to one first communication device 201 and two second communication devices 202 .
  • the number and form of equipment shown in FIG. 2 are for example and do not constitute a limitation to the embodiment of the present application. In practical applications, two or more second communication devices 202 and more than three second communication devices 202 may be included.
  • the first communication device may be a network device or a terminal device
  • the second communication device may be a terminal device.
  • the first communication device and the second communication device are both terminal devices as an example for illustration.
  • the second communication device is a peripheral terminal device of the first communication device, that is, each second communication device and the first communication device are located in the same cell. Both the first communication device and the second communication device are provided with a neural network model, and the second communication device can cooperate with the first communication device to participate in the training of the first neural network model of the first communication device.
  • the network device is a device with wireless transceiver function, which is used to communicate with the terminal device, and may be an evolved base station (evolved Node B, eNB or eNodeB) in LTE; or a base station in a 5G network or a future Evolved public land mobile network (public land mobile network, PLMN) base station, broadband network gateway (broadband network gateway, BNG), aggregation switch or non-third generation partnership project (3rd generation partnership project, 3GPP) access equipment etc.
  • eNB evolved Node B
  • eNodeB evolved base station
  • PLMN public land mobile network
  • BNG broadband network gateway
  • aggregation switch or non-third generation partnership project (3rd generation partnership project, 3GPP) access equipment etc.
  • the network equipment in the embodiment of the present application may include various forms of base stations, such as: macro base stations, micro base stations (also called small stations), relay stations, access points, devices that will implement base station functions in the future, and WiFi systems
  • the access node in the transmission and receiving point (transmitting and receiving point, TRP), the transmitting point (transmitting point, TP), mobile switching center and device-to-device (Device-to-Device, D2D), vehicle outreach (vehicle- to-everything, V2X), machine-to-machine (machine-to-machine, M2M) communications, etc., which undertake the function of a base station, etc., which are not specifically limited in this embodiment of the present application.
  • Network devices can communicate and interact with core network devices to provide communication services to terminal devices.
  • the core network device is, for example, a device in a 5G network core network (core network, CN).
  • core network CN
  • the core network provides an interface to the data network, providing communication connections, authentication, management, policy control, and carrying data services for terminals.
  • the terminal devices involved in the embodiments of the present application may include various handheld devices with wireless communication functions, vehicle-mounted devices, wearable devices, computing devices or other processing devices connected to wireless modems.
  • Terminal equipment may also refer to user equipment (user equipment, UE), access terminal, subscriber unit (subscriber unit), user agent, cellular phone (cellular phone), smart phone (smart phone), wireless data card, personal digital assistant (PDA) Personal digital assistant (PDA) computer, tablet computer, wireless modem (modem), handheld device (handset), laptop computer (laptop computer), machine type communication (machine type communication, MTC) terminal, high-altitude aircraft Communication devices, wearable devices, drones, robots, terminals in device-to-device (D2D), terminals in vehicle to everything (V2X), virtual reality (virtual reality, VR) terminal equipment, augmented reality (augmented reality, AR) terminal equipment, wireless terminals in industrial control, wireless terminals in self driving, wireless terminals in remote medical, Wireless terminals in smart grid, wireless terminals in transportation safety, wireless terminals
  • Embodiments disclosed in the application will present various aspects, embodiments or features of the application around a system including a plurality of devices, components, modules, and the like. It is to be understood and appreciated that the various systems may include additional devices, components, modules, etc. and/or may not include all of the devices, components, modules etc. discussed in connection with the figures. Additionally, combinations of these schemes can also be used.
  • Federated learning is a learning method that efficiently completes the model by promoting the cooperation of various peripheral devices and central servers under the premise of fully ensuring the privacy and security of user data.
  • the FL algorithm is as follows:
  • the terminal device m uses the local data set to train the local neural network model, and the local gradient It is transmitted to the central server through the air interface.
  • ⁇ represents the number of gradient parameters returned
  • T is the threshold number of times
  • T is greater than or equal to 2
  • M is the total number of terminal devices, Represents the gradient corresponding to the ⁇ -th gradient parameter of the terminal device m in the i-th round of training;
  • the central server aggregates and collects the gradients from all (part) terminal devices, and weights and averages them to obtain a new global gradient:
  • the central terminal updates the local neural network model according to the new global gradient to obtain the updated neural network model. If the updated neural network model does not converge and the training times do not reach the threshold, the new global gradient is broadcast to each Terminal Equipment. After receiving the new global gradient, the terminal device updates its own local neural network model according to the new global gradient until the neural network model at the center converges or the number of training rounds reaches a threshold number of times.
  • FIG. 3 An exemplary system schematic diagram of federated learning is shown in FIG. 3 .
  • the central end is a network device
  • the peripheral devices are various terminal devices.
  • each terminal device uploads the gradient calculated locally to the network device through a wireless channel.
  • the network device summarizes multiple local gradients, that is, performs weighted average processing on multiple received gradients to obtain the global gradient, and updates the local neural network model according to the global gradient. If the updated neural network model still does not converge, and the number of training times does not reach the threshold number of times, the global gradient is broadcast to each terminal device.
  • any terminal device After any terminal device receives the global gradient, it uses the global gradient to update the neural network model of the terminal device, and uploads the gradient of the updated neural network model to the network device for the next round of neural network model training until the network The neural network model of the device converges, or the number of training times reaches the threshold number.
  • Segmentation learning is shown in Figure 4.
  • segmentation learning the complete neural network model is divided into two parts (ie, two sub-networks), part of the sub-network of the neural network is deployed on the distributed nodes, and the other part of the sub-network is deployed on the central node.
  • the place where the full neural network is divided is called a "segmentation layer”.
  • the distributed nodes input the local data into the local sub-network, reason to the segmentation layer, and send the result F1 of the segmentation layer to the central node through the communication link, and the central node inputs the received F1 Another sub-network deployed by itself, and continue to perform forward reasoning to obtain the final reasoning result.
  • the gradient is reversely transferred to the segmentation layer through the sub-network of the central node, and the reverse transfer result G1 is obtained, and then the central node sends G1 to the distributed nodes, so that G1 continues to be distributed in the distributed nodes.
  • Gradient backpropagation is performed on the sub-network of .
  • the trained sub-network on the distributed node can be saved locally on the distributed node or on a specific model storage server.
  • the new distributed node can first download the trained distributed node sub-network, and then use local data for further training.
  • the central node summarizes the local models reported by each distributed node, fuses the neural network models of each distributed node, and then sends them to each distributed node for the next round Train until the neural network model of the central node converges.
  • the contribution of some distributed nodes to the convergence of the neural network model of the central node will gradually decrease. Node signaling overhead.
  • the embodiment of the present application provides a model training method 100 .
  • the first communication device sends the first neural network parameters of the first communication device.
  • the second communication device receives the first neural network parameter of the first communication device, and when the correlation coefficient between the first neural network parameter and the second neural network parameter of the second communication device is less than the first threshold, the second communication device sends The first communication device sends first indication information, where the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the first communication device. Therefore, the second communication device receives the first indication information.
  • the second communication device feeds back to the first communication device to participate in the training of the first neural network model, which can prevent the second communication device from
  • the correlation coefficient between a neural network parameter and a second neural network parameter is equal to or greater than the first threshold, they still participate in the training of the first neural network model, thereby reducing the signaling overhead of the second communication device.
  • the embodiment of the present application also provides a model training method 200 .
  • the first communication device sends cooperation request information, and the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained.
  • the second communication device receives the cooperation request information.
  • the second communication device determines to participate in the first training task, it sends second instruction information, the second instruction information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is one of the multiple training tasks or more.
  • the first communication device receives the second indication information from the second communication device.
  • the first communication device divides the neural network model to be trained into multiple training tasks, and broadcasts the multiple training tasks to the surrounding second communication devices through the cooperation request information, and the surrounding second communication devices feed back the second
  • the instruction information is used to notify the first communication device that it can participate in the training task of the training through the second instruction information. That is to say, each second communication device in the surrounding assists in the training of the neural network model to be trained of the first communication device, thereby reducing the requirement on the capabilities of the first communication device.
  • FIG. 5 is a schematic diagram of an interaction process of the model training method 100 .
  • the model training method 100 is described from the perspective of interaction between the first communication device and the second communication device.
  • the model training method 100 includes but not limited to the following steps:
  • the first communication device sends first neural network parameters of the first communication device.
  • the first neural network parameter is a neural network model, or a gradient of the neural network, or may also be training data for training the neural network model. That is to say, the first neural network parameter is the first neural network model of the first communication device, or the gradient of the first neural network, or the training data for training the first neural network model.
  • the neural network model includes neurons included in the neural network, and weights between neurons in each layer.
  • the first communication device trains the first neural network model based on its own local data set. After the first communication device has been trained for a threshold number of rounds, if the first neural network model still fails to meet the preset convergence condition, the cooperation mechanism may be triggered. Triggering the cooperation mechanism by the first communication device may include: the first communication device sends a request message to the network device, so as to request the network device to configure the first communication device with relevant resources for cooperative training.
  • the request message may be an on demand system information block (on demand system information block, on demand SIB).
  • the network device After receiving the request message from the first communication device, the network device sends sidelink configuration information to the first communication device and peripheral devices (each second communication device) of the first communication device.
  • the sidelink configuration information may be SIB_AI_sidelink, and the sidelink configuration information is used to configure cooperative synchronization resources, cooperative discovery resources or cooperative control resources.
  • the cooperation synchronization resource may be an artificial intelligence cooperation synchronization (AI-cooperation-sync) resource, and the cooperation synchronization resource is used for each second communication device to synchronize with the first communication device;
  • the cooperation discovery resource may be an artificial intelligence cooperation discovery (AI-cooperation-sync) resource.
  • the cooperation discovery resource is used for the first communication device to send the first neural network parameter, and is also used for the second communication device to monitor the first neural network parameter of the first communication device;
  • the cooperative control resource can be artificial intelligence
  • a control discovery (AI-cooperation-control) resource is used for the first communication device to instruct each second communication device to send a resource of the second neural network parameter.
  • the first neural network parameter is sent on the cooperation discovery resource, for example, the first neural network parameter is sent by the first communication device on the AI-cooperation-discover resource.
  • the cooperative discovery resource is configured in the above sidelink configuration information.
  • the above sidelink configuration information may be pre-configured by itself, and delivered to each second communication device.
  • the first communication device may also send a synchronization signal on the cooperative synchronization resource, so that the second communication device communicates with the first communication device according to the synchronization signal. Synchronize.
  • the second communication device receives the first neural network parameters of the first communication device.
  • the network device after receiving the request message from the first communication device, the network device sends sidelink configuration information to each second communication device, so that the second communication device configures A first neural network parameter is received on the collaborative discovery resource configured by the information.
  • the second communication device may also monitor a synchronization signal on the cooperative synchronization resources configured in the sidelink configuration information, and, according to the synchronization signal and the first The communication device is synchronized. Therefore, after the second communication device completes the synchronization with the first communication device, it can communicate with the first communication device, for example, receive the first neural network parameters from the first communication device.
  • the second communication device sends the first indication information to the first communication device, and the first indication information is used for Instructing the second communication device to participate in the training of the first neural network model of the first communication device.
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network
  • the first neural network is a neural network of the first communication device
  • the second neural network parameter is a model parameter of the second neural network.
  • the neural network is the neural network of the second communication device; when the first neural network parameter is the gradient of the first neural network model, the second neural network parameter is the gradient of the second neural network model; when the first neural network parameter is the gradient of the first neural network model;
  • the second neural network parameters are training data for training the second neural network model.
  • the model parameters include a neural network structure, weights between neurons in the neural network structure, and the like.
  • the first threshold is preset by the second communication device.
  • the second neural network parameters of the second communication device are the neural network parameters of the neural network model after the second communication device receives the first neural network parameters and updates the local neural network model of the second communication device according to the first neural network parameters. Therefore, the second communication device compares the neural network parameters of the updated neural network model with the first neural network parameters to determine the correlation coefficient between the first neural network parameters and the second neural network parameters.
  • the local neural network model before the second communication device receives the first neural network parameters is the neural network model X
  • the neural network model after the second communication device receives the first neural network parameters is updated according to the first neural network parameters X is updated to obtain the neural network model Y.
  • the neural network model Y is the second neural network model, and then the second communication device compares the neural network parameters of the neural network model Y with the first neural network parameters to determine the relationship between the first neural network parameters and the second neural network parameters correlation coefficient.
  • the methods of determining the correlation coefficient between them are also different.
  • the method of determining the correlation coefficient between the first neural network parameter and the second neural network parameter is described in combination with the types to which the first neural network parameter and the second neural network parameter belong:
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is based on the first parameters and the second parameter is determined.
  • the first parameter is a parameter output by the first neural network model when the second communication device inputs training data to the first neural network model; the first neural network model is determined according to the model parameters of the first neural network; the second parameter is a parameter output by the second neural network model when the second communication device also inputs the training data to the second neural network model of the second communication device.
  • the second communication device determines the first neural network model according to the model parameters of the first neural network, and then determines the first parameter and the second neural network model according to the first neural network model and the second neural network model. the second parameter, and determine the correlation coefficient between the first neural network parameter and the second neural network parameter according to the first parameter and the second parameter.
  • the first parameter and the second parameter are parameters respectively output by the first neural network model and the second neural network model when the second communication device inputs the same training data to the first neural network model and the second neural network model.
  • the second communication device divides the covariance of the first parameter and the second parameter by the product of the standard deviation of the first parameter and the standard deviation of the second parameter as the first neural network parameter and the second Evaluation Criteria for Correlations Between Neural Network Parameters. For example, when the second communication device inputs the same training data to the first neural network model and the second neural network model, the first neural network model and the second neural network model output X and Y respectively, that is, X is the first parameter, and Y is the second parameter, then the correlation coefficient between the first neural network parameter and the second neural network parameter is:
  • Cov(X,Y) represents the covariance of X and Y
  • ⁇ X and ⁇ Y represent the standard deviations of X and Y, respectively.
  • the first neural network parameter is the gradient of the first neural network model, and the second neural network parameter is the gradient of the second neural network model; or, the first neural network parameter is the training data for training the first neural network model, and the second neural network parameter is the gradient of the second neural network model;
  • the neural network parameters are training data for training the second neural network model.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is based on the first The probability distribution of gradients of the neural network model is determined with the probability distribution of gradients of the second neural network model.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is determined according to the probability distribution of the training data for training the first neural network model and the probability distribution of the training data for training the second neural network model.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is based on the probability distribution of the first neural network parameter, the probability distribution of the second neural network parameter, and the Hellinger distance (Hellinger The definition of distance) is determined. That is, the correlation coefficient between the first neural network parameter and the second neural network parameter is:
  • Z a , Z b represent the first neural network parameter and the second neural network parameter respectively
  • S(Z a ), S(Z b ) respectively represent the probability distribution of the first neural network parameter and the probability distribution of the second neural network parameter distributed.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is less than the first threshold, which means that the correlation between the second neural network parameter and the first neural network parameter is relatively large, so that the second neural network parameter Contributes more to the convergence of the first neural network model. That is to say, the second communication device determines whether to participate in the training of the first neural network model according to the contribution of the second neural network parameters to the convergence of the first neural network model. When the second neural network parameter has a large contribution to the convergence of the first neural network model, the second communication device determines the parameters for the training of the first neural network model, and informs the first communication device through the first indication information.
  • the second communication device still participates in the training of the first neural network model when the contribution of the second neural network parameters to the convergence of the first neural network model is low, thereby reducing the signaling overhead of the second communication device, that is, reducing Unnecessary transmission overhead by the first telecommunications device.
  • the above-mentioned sidelink configuration information is also configured with a cooperation response resource for the second communication device to send the first indication information, so that the second communication device sends the first communication device Send the first indication information.
  • the second communication device may also send the first indication information to the first communication device on the cooperative control resource.
  • the second communication device sends a fifth indication to the first communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter is equal to or greater than the first threshold information, the fifth indication information is used to instruct the second communication device not to participate in the training of the first neural network model. That is to say, when the second communication device determines not to participate in the training of the first neural network model, it indicates to the first communication device not to participate in the training of the first neural network model through the fifth indication information, so that the first communication device learns that The second communication device does not participate in this round of neural network parameter training.
  • the first communication device and the second communication device agree in advance that the second communication device does not Feedback any information to the first communication device, that is, when the second communication device determines not to participate in the training of the first neural network model in this round, it does not do any processing. Therefore, when the first communication device does not receive feedback information from the second communication device within a preset time, it is determined that the second communication device does not participate in the current round of training of the first neural network model, thereby saving system signaling overhead.
  • the second communication device determines not to participate in the training of the current round of the first neural network model, it waits for the next round of the first communication device to send the neural network parameters, and again performs the training with the neural network parameters of the local neural network model updated by the second communication device. Compare to determine whether to participate in the next round of training.
  • the first communication device receives the first indication information from the second communication device.
  • the first communication device may receive the first indication information on the cooperation response resource configured by the sidelink configuration information, or may receive the first indication information on the cooperation discovery resource.
  • the embodiment of this application is not limited.
  • the first communication device By receiving the first instruction information, the first communication device knows the second communication device that is willing to participate in the training of the first neural network model, so that the second communication device Updating the first neural network model is beneficial to saving signaling overhead of the second communication device, that is, saving transmission overhead of the second communication device.
  • the first communication device after the first communication device obtains the second communication device willing to participate in the training of the first neural network model through the first indication information, it sends a control signal to the second communication device in this part, and the control signal uses For indicating the time-frequency resource, the indicated time-frequency resource is used for the second communication device to send the second neural network parameters.
  • the resource for sending the control signal by the first communication device is a cooperative control resource
  • the cooperative control resource is configured in the above-mentioned sidelink configuration information.
  • the first communication device determines the second communication device willing to participate in the training of the first neural network model, it sends a control signal to the second communication device of this part, and instructs the second communication device of this part to send the second communication device.
  • the time-frequency resources of the neural network parameters so that the part of the second communication device sends the second neural network parameters to the first communication device on their corresponding time-frequency resources.
  • the time-frequency resources for sending the second neural network parameters by the second communication device are dynamically scheduled by the network equipment to the first communication device.
  • the first communication device schedules the time-frequency resource to the second communication device that feeds back the first indication information through the control signal.
  • the time-frequency resources scheduled by the first communication device to different second communication devices are different. In this way, the first communication device dynamically schedules the time-frequency resources to the second communication device that feeds back the first instruction information each time, which can make the utilization rate of resources higher.
  • the time-frequency resource for sending the second neural network parameters by the second communication device is semi-statically configured by the network device to the first communication device.
  • the semi-static resource occurs periodically, so that the first communication device does not need to schedule time-frequency resources for the second communication device.
  • the first communication device still needs to indicate the semi-static resource in the second communication device to the second communication device that has fed back the first indication information through a control signal, so as to activate the semi-static resource.
  • the second communication device may use the semi-static resource to send the second neural network parameters to the first communication device. In this manner, the first communication device does not need to schedule time-frequency resources for the second communication device, which can reduce signaling overhead.
  • the second communication device uses the time-frequency resource indicated by the control signal to send the second neural network parameters to the first communication device.
  • the first communication device receives the second neural network parameters from the second communication device, and updates the first neural network model according to the second neural network parameters. It can be understood that the first communication device receives the second neural network parameters of multiple second communication devices, and the multiple second communication devices are all second communication devices that feed back the first indication information, so that the first communication device A second neural network parameter updates the first neural network model.
  • the updating of the first neural network model by the first communication device according to a plurality of second neural network parameters refers to: the first communication device averages and sums the parameters of each second neural network model to obtain the processed second neural network model. network parameters, and then update the first neural network model according to the processed second neural network parameters. If the updated first neural network model still does not converge, and the number of training times does not reach the threshold number of times, the first communication device broadcasts the first neural network parameters of the updated first neural network model to the surrounding second communication devices , so that each second communication device decides whether to participate in the next round of training of the updated first neural network model according to its own local neural network parameters and the received neural network parameters.
  • the coordinated synchronization resources, coordinated discovery resources, and coordinated control resources configured in the sidelink configuration information can be dynamically indicated by the network device after receiving the request message.
  • the cooperative synchronization resources, cooperative discovery resources, and cooperative control resources configured in the sidelink configuration information may also be pre-configured by the network device, and may also be unlicensed spectrum resources. This embodiment of the present application does not limit it.
  • FIG. 6 is a schematic diagram of an interaction process of a model training method in which the first communication device is terminal device A and the second communication device includes terminal device B and terminal device C according to an example of an embodiment of the present application. As shown in Figure 6:
  • terminal device A If terminal device A performs N rounds of training using local training data, and the first neural network model of terminal device A has not yet converged, then terminal device A sends synchronization to surrounding terminal devices (terminal device B and terminal device C). Signal.
  • the value of N is less than the threshold number of times.
  • terminal device B and the terminal device C monitor the synchronization signal, they respectively synchronize with the terminal device A according to the synchronization signal, and then monitor the first neural network parameter of the terminal device A.
  • the terminal device B, the terminal device C and the terminal device A are synchronized to ensure that they can communicate with the terminal device A subsequently.
  • Both the resource for terminal device A to send the synchronization signal and the resources for terminal device B and terminal device C to monitor the synchronization signal may be the cooperative synchronization resource configured by the above sidelink configuration information, and details will not be repeated here.
  • the terminal device A broadcasts the first neural network parameters of the terminal device A on the aforementioned collaborative discovery resource.
  • terminal device B and terminal device C monitor the first neural network parameters on the collaborative discovery resource, they compare the first neural network parameters with their own second neural network parameters, and judge the first neural network parameters and the second neural network parameters. Whether the correlation coefficient between the network parameters is smaller than the first threshold.
  • the correlation coefficient between the terminal device B and the terminal device C is less than the first threshold, send to the terminal device A first indication information for indicating participation in the training of the first neural network model.
  • terminal device A After terminal device A receives the first instruction information from terminal device B and terminal device C, it sends control signals to terminal device B and terminal device C, and instructs terminal device B and terminal device C to send its own first instruction information through the control signals, respectively.
  • Two Time-Frequency Sources of Neural Network Parameters Two Time-Frequency Sources of Neural Network Parameters.
  • the terminal device B and the terminal device C send the second respective second neural network parameters to the terminal device A according to the indicated time-frequency resources respectively.
  • control signal #b sent by the terminal device A to the terminal device B is used to indicate the time-frequency resource #b
  • the control signal #c sent to the terminal device C is used to indicate the time-frequency resource #c. Therefore, the terminal device B sends the second neural network parameters of the terminal device B to the terminal device A on the time-frequency resource #b, and the terminal device C sends the second neural network parameters of the terminal device C to the terminal device A on the time-frequency resource #c .
  • terminal device A After receiving the second neural network parameters from terminal device B and the second neural network parameters from terminal device C, terminal device A fuses the two second neural network parameters, that is, weights and averages the two neural network parameters , to obtain the global neural network parameters.
  • the terminal device A starts the N+1 round of training by using the global neural network parameters, that is, updates the first neural network model according to the global neural network parameters, and obtains the updated first neural network model.
  • terminal device A broadcasts the first neural network parameters of the updated first neural network model, and terminal device B and terminal device C again According to the received neural network parameters and the current local neural network parameters, determine whether to participate in the next round of training of terminal device A until the neural network model of terminal device A converges, or the number of training times reaches the threshold number of times.
  • the second communication device compares its own second neural network parameters with the received first neural network parameters, and the correlation coefficient between the first neural network parameters and the second neural network parameters is smaller than the first neural network parameter.
  • a threshold value is reached, feed back to the first communication device the first instruction information used to indicate participation in the training of the first neural network model, so that the first communication device can , to update the first neural network model.
  • the correlation coefficient between the first neural network parameter and the second neural network parameter is smaller than the first threshold, indicating that the second neural network parameter contributes more to the convergence of the first neural network model.
  • the second communication device determines whether to participate in the training of the first neural network model according to the contribution of the second neural network parameters to the convergence of the first neural network model, which can prevent the second communication device from affecting the first neural network model by the second neural network parameters.
  • the convergence contribution of the network model is small, it still participates in the training of the first neural network model, thereby reducing the signaling overhead of the second communication device.
  • the first communication device no longer updates the first neural network model according to the second neural network parameters of all the surrounding second communication devices, but updates the first neural network model according to the second neural network parameters of the second communication device that fed back the first indication information
  • the first neural network model can reduce the signaling overhead of the first communication device.
  • FIG. 7 is a schematic diagram of an interaction process of the model training method 200 .
  • the model training method 200 is also described from the perspective of interaction between the first communication device and the second communication device.
  • the model training method 200 includes but not limited to the following steps:
  • the first communication device sends cooperation request information, where the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by the first communication device by splitting the neural network model to be trained.
  • the first communication device when the first communication device has limited resources or capabilities and cannot independently complete the training of the neural network model to be trained, it splits the neural network model to be trained into multiple training tasks, and uses collaborative The way of requesting information is to broadcast the plurality of training tasks to the surrounding second communication devices, so as to request each second communication device to assist the first communication device to participate in the training of the plurality of training tasks.
  • each training task includes a simple neural network model
  • each simple neural network model is a sub-network of the neural network model to be trained.
  • the first communication device splits the neural network model to be trained based on delays of sub-networks after splitting.
  • the first communication device splits the neural network model to be trained based on the computing power of each second communication device.
  • the first communication device splits the neural network model to be trained based on the amount of data generated by the splitting location.
  • the first communication device divides the neural network model to be trained into multiple training tasks on average according to the structure of the neural network model to be trained, so that each training task includes the neural network model to be trained
  • the number of neural network layers is the same. This application does not limit the number of neural network layers included in each training task.
  • the first communication device splits the neural network model to be trained into three training tasks, namely, training task A, training task B, training task C, training task A, and training task B , and the training task C both include two layers of neural networks.
  • the first communication device takes the training task C as its own training part, takes the training task A and the training task B as two training tasks, and broadcasts them to the surrounding second communication devices in the form of a cooperation request.
  • the first communication device splits the neural network model to be trained into multiple training tasks with uneven training levels, that is, the number of neural network layers included in each training task is different.
  • the second communication device with less remaining resources or its own computing power to participate in the training of the training task with fewer neural network layers, and the second communication device with larger remaining resources or its own computing power to participate in the training of the training task with more neural network layers. Training for training tasks.
  • the first communication device when it determines that the neural network model to be trained cannot be completed, it sends a request message (for example, on demand SIB) to the network device to request the network device to configure the first communication device.
  • a request message for example, on demand SIB
  • the network device After receiving the request message from the first communication device, the network device sends sidelink configuration information to the first communication device and peripheral devices (each second communication device) of the first communication device.
  • the implementation manner of the sidelink configuration information may refer to the description in S101 above, which will not be repeated here.
  • the cooperation request information can be sent on the cooperation discovery resource configured by the sidelink configuration information.
  • the above-mentioned cooperation request information may also include the estimated cost corresponding to each training task, etc., so that after the second communication device receives the cooperation request information, it can determine according to its own remaining resources. Whether to participate in the training of some of these training tasks.
  • the second communication device receives cooperation request information.
  • the network device sends sidelink configuration information to the second communication device, and the sidelink configuration information configures some cooperative discovery resources for receiving request information. Therefore, the second communication device may receive the cooperation request information on the cooperation discovery resource configured by the above-mentioned sidelink configuration information.
  • the second communication device determines to participate in the first training task, it sends second instruction information, the second instruction information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is one of multiple training tasks one or more of them.
  • the second communication device determines one or more training tasks that it can participate in from the multiple training tasks in the cooperation request information according to its own remaining resources, its own computing power, and the like. Then, the second communication device notifies the first communication device of the first training task that it can participate in through the second indication information.
  • the first training task includes one or more training tasks that the second communication device can train.
  • the first communication device and each second communication device pre-agreed: after the second communication device receives the cooperation request information, if it feeds back the training tasks that it can participate in to the first communication device, it will feedback that it can participate in the training tasks.
  • a training task for training That is to say, when the second communication device can participate in the training of one or more training tasks, one of the above-mentioned one or more training tasks is fed back to the first communication device through the second indication information, that is, the first training task includes One of the training tasks carried in the cooperation request information.
  • the first communication device and each second communication device agree in advance: after the second communication device receives the cooperation request information, if it feeds back the training tasks that it can participate in to the first communication device, it can feed back its own training tasks. All training tasks that can participate in training. That is to say, the second communication transposition can feed back all training tasks that can participate in the training to the first communication device through the second indication information, that is, the first training task includes one or more training tasks carried in the cooperation request information. training tasks.
  • the first communication device receives second indication information from the second communication device.
  • the first communication device learns the second communication devices that are willing to participate in the training and the training tasks that each second communication device can participate in.
  • the second communication device feeds back a training task that can be participated in, or feeds back all training tasks that can be participated in to the first communication device.
  • each second communication device feeds back a training task that can be participated in, some of the training tasks fed back by the second communication devices may be the same training task.
  • the first communication device needs to determine the second communication device that trains the same training task from the multiple second communication devices that feed back the same training task. It can be understood that the first communication device negotiates, or the first communication device is based on the remaining resources of multiple second communication devices, or the amount of computing power, or the channel quality between each second communication device and the first communication device, etc. , determining a second communication device participating in the same training task, and sending third indication information to the determined second communication device, so as to inform the second communication device that training for the training task indicated by the third indication information can be performed.
  • the negotiation by the first communication device may mean that the first communication device negotiates with multiple second communication devices, and determines that the second communication device closest to the first communication device participates in the training of the same training task.
  • the embodiment of the present application does not limit a specific negotiation manner.
  • the first communication device determines the second communication device participating in the training of the same training task according to the computing power of each second communication device in the plurality of second communication devices, which may refer to the first communication device combining multiple second communication devices
  • the second communication device with the largest computing power among the communication devices participates in the training of the same training task, so as to ensure that the same training task is completely trained.
  • the cooperation request information sent by the first communication device carries training task A and training task B, and the surrounding second communication device A, second communication device B, and second communication device C have all received the cooperation request information.
  • the second instruction information sent by the second communication device A indicates the training task A
  • the second instruction information sent by the second communication device B and the second communication device C both indicate the training task B. It can be seen that the second communication device B and the second communication device C have fed back the same training task.
  • the first communication device needs to determine a second communication device from the second communication device B and the second communication device C to perform the training task B training. communication device.
  • the first communication device determines to let the second communication device B participate in the training of the training task B through negotiation with the second communication device B and the second communication device C, the first communication device sends the second communication device B to the second communication device B.
  • the third indication information is used to indicate the training task B. Therefore, the second communication device B can perform the training of the training task B after receiving the third indication information.
  • the cooperation request information sent by the first communication device carries training task A and training task B, and the surrounding second communication device A, second communication device B, and second communication device C have all received the cooperation request information.
  • the second instruction information sent by the second communication device A indicates the training task A and the training task C
  • the second instruction information sent by the second communication device B and the second communication device C both indicate the training task B.
  • the first communication device determines that the second communication device A participates in the training of the training task A and the training task C, and determines that the second communication device C participates in the training of the training task B through negotiation. Therefore, the third instruction information sent by the first communication device to the second communication device A is used to indicate the training task A and the training task C, and the third instruction information sent to the second communication device C is used to indicate the training task B.
  • the first communication device when each second communication device feeds back a training task that can be participated in, and each second communication device feeds back a different training task, the first communication device also feeds back the second indication information to each second communication device.
  • the second communication device sends third indication information, where the third indication information is used to indicate the training task fed back by the second communication device, so as to inform each second communication device that has fed back the second indication information that training of the training task can be performed.
  • the second communication device feeds back all training tasks that it can participate in training to the first communication device, then the first communication device determines each training task according to each training task that the second communication device can participate in training.
  • the training tasks that each second communication device needs to participate in and inform each second communication device of the training tasks that need to participate in in the form of third indication information. That is to say, the first communication device sends third indication information to each second communication device, and the third indication information is used to indicate one of the first training tasks.
  • the cooperation request information sent by the first communication device carries training task A, training task B, and training task C, and the surrounding second communication device A, second communication device B, and second communication device C have all received the cooperation request information.
  • Collaboration request information The training task C is indicated in the second instruction information sent by the second communication device A, the training task B is indicated in the second instruction information sent by the second communication device B, and the training task B is indicated in the second instruction information sent by the second communication device C. Training task A and training task B.
  • the first communication device determines that the second communication device can participate in the training of the training task A.
  • the third instruction information sent by the first communication device to the second communication device A is used to indicate the training task C
  • the third instruction information sent to the second communication device B is used to indicate the training task B
  • the third instruction information sent to the second communication device C is used to indicate the training task B.
  • the sent third indication information is used for training task A.
  • the cooperation request information sent by the first communication device carries training task A, training task B, and training task C, and the surrounding second communication device A, second communication device B, and second communication device C have all received the cooperation request information.
  • Collaboration request information The second instruction information sent by the second communication device A indicates the training task A and the training task C, the second instruction information sent by the second communication device B indicates the training task B, and the second instruction information sent by the second communication device C The information indicates training task A and training task B.
  • the second communication device determines that the second communication device A can participate in the training of the training task C according to the training tasks that the second communication device A, the second communication device B, and the second communication device C can respectively participate in, and the second communication device B can participate in the training of the training task C.
  • the second communication device C can participate in the training of the training task A. Therefore, the third instruction information sent by the first communication device to the second communication device A is used to indicate the training task C, the third instruction information sent to the second communication device B is used to indicate the training task B, and the third instruction information sent to the second communication device C is used to indicate the training task B.
  • the third indication information of is used to indicate the training task A.
  • each second communication device that is willing to participate in the training task knows the training task that it needs to train through the third indication information. Therefore, the second communication device in this part trains the corresponding neural network model according to the local training data, and stops training the neural network model until the first communication device determines that the neural network model converges. Understandably, the first communication device determines whether the neural network model is converged through the input and output of the neural network model to be trained.
  • the first communication device splits the neural network model to be trained into the three training tasks shown in FIG. The training of sub-networks output by the network.
  • the first communication device negotiates with each second communication device on the input (X) of the neural network, and through the output (output of the neural network) Y of the training task C trained by itself, determines the neural network trained by itself and each second communication device. Whether the network model converges.
  • the first communication device sends a fourth instruction to each second communication device after determining the training task that each second communication device participates in among the second communication devices that have fed back the second instruction information information.
  • the fourth indication information is used to indicate the first output to be received by the second communication device, the time-frequency position corresponding to the first output, and/or the second output to be transmitted, and the time-frequency resource position corresponding to the second output.
  • the first output is the output of the neural network model trained by the first communication device, or the output of the neural network model trained by other second communication devices except the second communication device;
  • the second output is the output of the neural network model trained by the second communication device The output of the trained neural network model.
  • the first communication device indicates the parameters to be received and/or the parameters to be received for each second communication device participating in the training task through the fourth indication information, and schedules the parameters to be received and/or the parameters to be received for each second communication device.
  • the time-frequency resources of the parameters to be sent so that the second communication device participating in the training knows which parameters it needs to receive and/or send on which time-frequency resources.
  • the first communication device splits the neural network model to be trained into the above training tasks as shown in FIG. 8 , and the first communication device performs the training of training task A, and the second communication device A performs training task B Training, the second communication device B performs the training of the training task C.
  • the first output is the output (output A) of the neural network model trained by the first communication device
  • the second output is the output of the neural network model in the training task B trained by the second communication device A (output B).
  • the first output is the output (output B) of the neural network model in the training task B trained by the second communication device A
  • the second output is the neural network model in the training task C trained by the second communication device B.
  • the first communication device determines that it needs to send the output of the neural network model in the training task A on the time-frequency resource #a.
  • the first communication device indicates to the second communication device A through the fourth indication information that the second communication device needs to receive the output A on part of the time-frequency resources in the time-frequency resource #b, and receive the output A on another part of the time-frequency resource #b.
  • the first communication device indicates to second communication device A through the fourth indication information that the second communication device needs to receive output A and send output B on different frequency domain resources corresponding to time-frequency resource #b Output B; and the first communication device indicates to the second communication device B through the fourth indication information that the second communication device B needs to receive the output B on a part of the time-frequency resources in the time-frequency resource #c, and in the time-frequency resource #c or, the first communication device indicates to the second communication device B through the fourth indication information that the second communication device B needs to receive on a different frequency domain resource corresponding to the time-frequency resource #c output B and send output C. It can be seen that the second communication device A and the second communication device B transmit their own training results in a time-division manner.
  • the resource for sending the fourth indication information may be the cooperative control resource configured by the above-mentioned side row configuration information.
  • the first communication device before sending the cooperation request information, sends a synchronization signal on the coordinated synchronization resource, so that the second communication device receives the synchronization signal on the coordinated synchronization resource, and according to the synchronization signal and the first A communication device is synchronized for subsequent communication with the first communication device.
  • the coordinated synchronization resource is configured in the above sidelink configuration information.
  • FIG. 9 is a schematic diagram of an interaction process of another model training method in which the first communication device is terminal device A and the second communication device includes terminal device B and terminal device C according to the embodiment of the present application. As shown in Figure 9:
  • terminal device A sends a synchronization signal to surrounding terminal devices (terminal device B and terminal device C).
  • the terminal device B and the terminal device C respectively synchronize with the terminal device A according to the synchronization signal, and monitor the cooperation request information of the terminal device A.
  • the terminal device B, the terminal device C and the terminal device A are synchronized to ensure that they can communicate with the terminal device A subsequently.
  • Both the resource for terminal device A to send the synchronization signal and the resources for terminal device B and terminal device C to monitor the synchronization signal may be the cooperative synchronization resource configured by the above sidelink configuration information, and details will not be repeated here.
  • the terminal device A broadcasts the cooperation request information on the above-mentioned cooperation discovery resource, so as to request the surrounding terminal devices to assist in the training of multiple training tasks.
  • terminal device B and terminal device C listen to the cooperation request information on the cooperation discovery resource, they determine one or more training tasks that they can participate in from the multiple training tasks carried in the cooperation request information according to their own remaining resources .
  • Terminal device B and terminal device C send second indication information to terminal device A, so as to feed back to terminal device A one or more training tasks that they can participate in through the second indication information.
  • Terminal device A determines the training tasks that terminal device B and terminal device C need to participate in respectively according to the training tasks that terminal device B and terminal device C can participate in, and sends the third instruction information to notify The training tasks that terminal device B and terminal device C need to participate in respectively.
  • terminal device A sends fourth indication information to terminal device B and terminal device C to inform terminal device B and terminal device C of the parameters to be received and/or sent respectively, and the time-frequency resource location corresponding to each parameter.
  • Terminal device B and terminal device C perform training on the training task indicated by the third indication information, and send or receive corresponding parameters according to the time-frequency resource position of each parameter.
  • terminal device A determines that the neural network model is converged, it determines that the training of the neural network model is completed.
  • the first communication device splits the neural network model to be trained into multiple simple neural network models, that is, multiple training tasks, and broadcasts the multiple training tasks to the Peripheral second communication means.
  • the second communication device determines training tasks that can be participated in according to its own remaining resources.
  • the second communication device indicates to the first communication device the training tasks that it can participate in by means of the second indication information. Therefore, the first communication device determines the training task of each second communication device according to the training tasks that each second communication device can participate in, and notifies each second communication device with the third instruction information.
  • each second communication device participating in the cooperative training trains the training task indicated by the third indication information.
  • One or more second communication devices cooperate with the first communication device to complete the training of the neural network model to be trained by training one or more training tasks among the multiple training tasks carried in the cooperation request information, thereby reducing the need for the first communication The capability requirements of the device.
  • the training of the neural network model to be trained is completed without the participation of network devices, and only the cooperation of each terminal device participates.
  • the first communication device completes the training of the neural network model to be trained according to the local data of the surrounding second communication devices, so that the trained neural network model is more accurate.
  • the first communication device or the second communication device may include a hardware structure and/or a software module in the form of a hardware structure, a software module, or a hardware structure plus a software module. Realize the above functions. Whether one of the above-mentioned functions is executed in the form of a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.
  • the embodiment of the present application provides a communication device 1000 .
  • the communication device 1000 may be a component of the first communication device (eg, integrated circuit, chip, etc.), or a component of the second communication device (eg, integrated circuit, chip, etc.).
  • the communication device 1000 may also be another communication unit, configured to implement the method in the method embodiment of the present application.
  • the communication device 1000 may include: a communication unit 1001 and a processing unit 1002 .
  • a storage unit 1003 may also be included.
  • one or more units in Figure 10 may be implemented by one or more processors, or by one or more processors and memory; or by one or more processors and a transceiver; or by one or more processors, memories, and a transceiver, which is not limited in this embodiment of the present application.
  • the processor, memory, and transceiver can be set independently or integrated.
  • the communication device 1000 has the function of realizing the second communication device described in the embodiment of the present application.
  • the communication device 1000 has the function of realizing the first communication device described in the embodiment of the present application.
  • the communication device 1000 includes a second communication device that executes the modules or units or means (means) corresponding to the steps described in the embodiment of the present application. It may be implemented by hardware, or by executing corresponding software by hardware, or by a combination of software and hardware. For details, further reference may be made to the corresponding descriptions in the aforementioned corresponding method embodiments.
  • a communication device 1000 may include: a processing unit 1002 and a communication unit 1001, where the processing unit 1002 is configured to control the communication unit 1001 to perform data/signaling transceiving;
  • a communication unit 1001, configured to receive the first neural network parameters of the first communication device
  • the communication unit 1001 is further configured to send first indication information to the first communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter of the device is smaller than a first threshold;
  • the first indication information is used to instruct the device to participate in the training of the first neural network model of the first communication device.
  • the first neural network parameter is a model parameter of the first neural network or the gradient of the first neural network;
  • the second neural network parameter is a model parameter of the second neural network or Gradients of the second neural network.
  • the first neural network parameter is received on a cooperative discovery resource; the cooperative discovery resource is configured in sidelink configuration information.
  • the communication unit 1001 is further configured to: send the second neural network parameters to the first communication device.
  • the communication unit 1001 is further configured to: receive a control signal from the first communication device; the control signal is used to indicate a time-frequency resource; the indicated time-frequency resource is used for the The device sends the second neural network parameters.
  • the resource for receiving the control signal is a cooperative control resource; the cooperative control resource is configured in the sidelink configuration information.
  • the communication unit 1001 is further configured to receive a synchronization signal on a coordinated synchronization resource; the processing unit 1002 is configured to perform synchronization with the first communication device according to the synchronization signal ;
  • the coordinated synchronization resource is configured in the sidelink configuration information.
  • one or more of the cooperative discovery resources, the cooperative control resources, and the cooperative synchronization resources configured in the sidelink configuration information are pre-configured, Either dynamically indicated, or unlicensed spectrum resources.
  • the first neural network parameter is a model parameter of the first neural network
  • the second neural network parameter is a model parameter of the second neural network
  • the first neural network parameter is a model parameter of the second neural network
  • the correlation coefficient between the network parameter and the second neural network parameter is determined according to the first parameter and the second parameter
  • the first parameter is the training data input by the second communication device to the first neural network model , the parameters output by the first neural network model
  • the first neural network model is determined according to the model parameters of the first neural network
  • the second neural network model of the communication device inputs the training data, the parameters output by the second neural network model.
  • the first neural network parameter is the gradient of the first neural network
  • the second neural network parameter is the gradient of the second neural network
  • the first neural network The correlation coefficient between the parameter and the second neural network parameter is determined according to the probability density distribution of the first neural network parameter and the probability density distribution of the second neural network parameter.
  • a communication device 1000 may include: a processing unit 1002 and a communication unit 1001, where the processing unit 1002 is configured to control the communication unit 1001 to perform data/signaling transceiving;
  • the communication unit 1001 is further configured to receive first indication information from the second communication device;
  • the first indication information is sent by the second communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter of the second device is smaller than a first threshold;
  • the first indication information is used to instruct the second communication device to participate in the training of the first neural network model of the device.
  • the first neural network parameter is a model parameter of the first neural network or the gradient of the first neural network;
  • the second neural network parameter is a model parameter of the second neural network or Gradients of the second neural network.
  • the first neural network parameter is sent on a cooperative discovery resource; the cooperative discovery resource is configured in sidelink configuration information.
  • the communication unit 1001 is further configured to receive a second neural network parameter from the second communication device; the processing unit 1002 is configured to receive the second neural network parameter according to the second neural network parameter , updating the first neural network model.
  • the communication unit 1001 is further configured to send a control signal to the second communication device; the control signal is used to indicate a time-frequency resource; the indicated time-frequency resource is used for the The second communication device sends the second neural network parameters.
  • the resource for sending the control signal is a cooperative control resource; the cooperative control resource is configured in the sidelink configuration information.
  • the communication unit 1001 is further configured to send a synchronization signal on a coordinated synchronization resource; the coordinated synchronization resource is configured in the sidelink configuration information.
  • one or more of the cooperative discovery resources, the cooperative control resources, and the cooperative synchronization resources configured in the sidelink configuration information are pre-configured, Either dynamically indicated, or unlicensed spectrum resources.
  • a communication device 1000 may include: a processing unit 1002 and a communication unit 1001, where the processing unit 1002 is configured to control the communication unit 1001 to perform data/signaling transceiving;
  • the communication unit 1001 is configured to send cooperation request information, where the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by splitting the neural network model to be trained by the first communication device;
  • the communication unit 1001 is further configured to receive second instruction information from the second communication device, the second instruction information is used to instruct the second communication device to participate in the training of the first training task, the first training task is one of the multiple training tasks one or more.
  • the above-mentioned cooperation request information is sent on the cooperation discovery resource, and the cooperation discovery resource is configured in the sidelink configuration information.
  • the first training task indicated by the second indication information sent by the second communication device includes multiple training tasks.
  • the first communication device may also send third indication information, and the third indication information uses to indicate one of the training tasks in the first training tasks.
  • the first training task indicated by the second indication information sent by multiple second communication devices is the same training task among the multiple training tasks.
  • the first communication device may also pass the third The instruction information indicates to one of the second communication devices the training task to participate in the training.
  • the first communication device may also send fourth indication information to the second communication device, and the fourth indication information is used to indicate the first output to be received by the second communication device and the time corresponding to the first output.
  • the first output is the output of the neural network model trained by the first communication device, or the output of the neural network model trained by other second communication devices except the second communication device; the second output is the neural network trained by the second communication device The output of the model.
  • the resource for sending the fourth indication information is a cooperative control resource
  • the cooperative control resource is configured in sidelink configuration information
  • the first communication device may also send a synchronization signal on the coordinated synchronization resource, so that the second communication device synchronizes with the first communication device according to the synchronization signal.
  • the coordinated synchronization resource is configured in sidelink configuration information.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information are pre-configured, or dynamically indicated, or are unlicensed spectrum resources.
  • a communication device 1000 may include: a processing unit 1002 and a communication unit 1001, where the processing unit 1002 is configured to control the communication unit 1001 to perform data/signaling transceiving;
  • the communication unit 1001 is configured to receive cooperation request information, where the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by splitting the neural network model to be trained by the first communication device;
  • the communication unit 1001 is further configured to send second indication information when determining to participate in the training of the first training task, the second indication information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is a plurality of training tasks One or more of the tasks.
  • the cooperation request information is received on the cooperation discovery resource, and the cooperation discovery resource is configured in the sidelink configuration information.
  • the second communication device may further receive third indication information, where the third indication information is used to indicate one of the training tasks in the first training tasks.
  • the second communication device may also receive fourth indication information, where the fourth indication information is used to indicate the first output received by the second communication device, the location of the time-frequency resource corresponding to the first output, and/or Or the sent second output, and the time-frequency resource position corresponding to the second output.
  • the first output is the output of the neural network model trained by the first communication device, or the output of the neural network model trained by other second communication devices except the second communication device; the second output is the neural network trained by the second communication device The output of the model.
  • the resources receiving the fourth indication information are cooperative control resources, and the cooperative control resources are configured in sidelink configuration information.
  • the second communication device may also receive a synchronization signal on the coordinated synchronization resource, and perform synchronization with the first communication device according to the synchronization signal.
  • the cooperative discovery resources, cooperative control resources, and cooperative synchronization resources configured in the above sidelink configuration information are pre-configured, or dynamically indicated, or are unlicensed spectrum resources.
  • FIG. 11 is a schematic structural diagram of the communication device 1100 .
  • the communication device 1100 may be a first communication device or a second communication device, or may be a chip, a chip system, or a processor that supports the first communication device to implement the above method, or may be a second communication device that supports the above method chips, chip systems, or processors.
  • the device can be used to implement the methods described in the above method embodiments, and for details, refer to the descriptions in the above method embodiments.
  • the communication device 1100 may include one or more processors 1101 .
  • the processor 1101 may be a general-purpose processor or a special-purpose processor. For example, it may be a baseband processor, a digital signal processor, an application specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, or a central processing unit (Central Processing Unit, CPU).
  • the baseband processor can be used to process communication protocols and communication data
  • the central processor can be used to process communication devices (such as base stations, baseband chips, terminals, terminal chips, distributed units (DU) or centralized units (centralized units). unit, CU) etc.) to control, execute the software program, and process the data of the software program.
  • DU distributed units
  • centralized units centralized units
  • the communication device 1100 may include one or more memories 1102, on which instructions 1104 may be stored, and the instructions may be executed on the processor 1101, so that the communication device 1100 executes the above method Methods described in the Examples.
  • data may also be stored in the memory 1102 .
  • the processor 1101 and the memory 1102 can be set separately or integrated together.
  • the memory 1102 may include but not limited to hard disk (hard disk drive, HDD) or solid-state drive (solid-state drive, SSD) and other non-volatile memory, random access memory (Random Access Memory, RAM), erasable and programmable Read-only memory (Erasable Programmable ROM, EPROM), ROM or portable read-only memory (Compact Disc Read-Only Memory, CD-ROM), etc.
  • the communication device 1100 may further include a transceiver 1105 and an antenna 1106 .
  • the transceiver 1105 may be called a transceiver unit, a transceiver, or a transceiver circuit, etc., and is used to realize a transceiver function.
  • the transceiver 1105 may include a receiver and a transmitter, and the receiver may be called a receiver or a receiving circuit for realizing a receiving function; the transmitter may be called a transmitter or a sending circuit for realizing a sending function.
  • the communication device 1100 is a second communication device: the transceiver 1105 is used to execute S102 and S103 in the model training method 100 described above, and to execute S202 and S203 in the model training method 200 .
  • the communication device 1100 is a second communication device: the transceiver 1105 is used for S101 and S104 in the model training method 100 , and for executing S201 and S204 in the model training method 200 .
  • the processor 1101 may include a transceiver for implementing receiving and sending functions.
  • the transceiver may be a transceiver circuit, or an interface, or an interface circuit.
  • the transceiver circuits, interfaces or interface circuits for realizing the functions of receiving and sending can be separated or integrated together.
  • the above-mentioned transceiver circuit, interface or interface circuit may be used for reading and writing code/data, or the above-mentioned transceiver circuit, interface or interface circuit may be used for signal transmission or transfer.
  • the processor 1101 may store instructions 1103, and the instructions 1103 run on the processor 1101, and may cause the communication device 1100 to execute the methods described in the foregoing method embodiments.
  • the instruction 1103 may be fixed in the processor 1101, in this case, the processor 1101 may be implemented by hardware.
  • the communication device 1100 may include a circuit, and the circuit may implement the function of sending or receiving or communicating in the foregoing method embodiments.
  • the processor and the transceiver described in the embodiment of the present application can be implemented in an integrated circuit (integrated circuit, IC), an analog IC, a radio frequency integrated circuit (radio frequency integrated circuit, RFIC), a mixed signal IC, an application specific integrated circuit (application specific integrated circuit) circuit, ASIC), printed circuit board (printed circuit board, PCB), electronic equipment, etc.
  • the processor and transceiver can also be fabricated using various IC process technologies such as complementary metal oxide semiconductor (CMOS), nMetal-oxide-semiconductor (NMOS), P-type Metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (Bipolar Junction Transistor, BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.
  • CMOS complementary metal oxide semiconductor
  • NMOS nMetal-oxide-semiconductor
  • PMOS P-type Metal oxide semiconductor
  • BJT bipolar junction transistor
  • BiCMOS bipolar CMOS
  • SiGe silicon germanium
  • GaAs gallium arsenide
  • the communication device described in the above embodiments may be the first communication device or the second communication device, but the scope of the communication device described in the embodiments of the present application is not limited thereto, and the structure of the communication device may not be limited by FIG. 11 .
  • a communication device may be a stand-alone device or may be part of a larger device.
  • the communication device may be:
  • a set of one or more ICs may also include storage components for storing data and instructions;
  • ASIC such as a modem (modulator)
  • the communication device may be a chip or a chip system
  • the chip 1200 shown in FIG. 12 includes a processor 1201 and an interface 1202 .
  • the number of processors 1201 may be one or more, and the number of interfaces 1202 may be more than one.
  • the processor 1201 may be a logic circuit, and the interface 1202 may be an input-output interface, an input interface or an output interface.
  • the chip 1200 may further include a memory 1203 .
  • the processor 1201 is used to control the interface 1202 to output or receive.
  • the interface 1202 is configured to receive the first neural network parameters of the first communication device
  • the interface 1202 is further configured to output first indication information when the correlation coefficient between the first neural network parameter and the second neural network parameter of the device is less than a first threshold; the first indication information is used for Instructing the device to participate in the training of the first neural network model of the first communication device.
  • the interface 1202 is used to output the first neural network parameters of the device
  • the interface 1202 is further configured to receive first indication information from the second communication device
  • the first indication information is output by the second communication device when the correlation coefficient between the first neural network parameter and the second neural network parameter of the second device is smaller than a first threshold; the first The instruction information is used to instruct the second communication device to participate in the training of the first neural network model of the device.
  • the interface 1202 is configured to output cooperation request information, where the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by splitting the neural network model to be trained by the first communication device;
  • the interface 1202 is also used to receive second instruction information from the second communication device, the second instruction information is used to instruct the second communication device to participate in the training of the first training task, the first training task is one of the multiple training tasks one or more of them.
  • the interface 1202 is configured to receive cooperation request information, where the cooperation request information includes multiple training tasks, and the multiple training tasks are obtained by splitting the neural network model to be trained by the first communication device;
  • the interface 1202 is further configured to output second indication information when it is determined to participate in the training of the first training task, the second indication information is used to instruct the second communication device to participate in the training of the first training task, and the first training task is a plurality of One or more of the training tasks.
  • the communication device 1100 and the chip 1200 may also execute the implementation described in the communication device 1000 above.
  • Those skilled in the art can also understand that various illustrative logical blocks and steps listed in the embodiments of the present application can be implemented by electronic hardware, computer software, or a combination of both. Whether such functionality is implemented as hardware or software depends upon the particular application and overall system design requirements. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the protection scope of the embodiments of the present application.
  • model training method 100 and model training method 200 are based on the same idea, and the technical effects they bring are also the same.
  • the implementation shown in the above-mentioned model training method 100 and model training method 200 The description of the example is omitted.
  • the present application also provides a computer-readable storage medium for storing computer software instructions, and when the instructions are executed by a communication device, the functions of any one of the above method embodiments are realized.
  • the present application also provides a computer program product, which is used for storing computer software instructions, and when the instructions are executed by a communication device, the functions of any one of the above method embodiments are realized.
  • the present application also provides a computer program, which, when running on a computer, can realize the functions of any one of the above method embodiments.
  • the present application also provides a communication system, which includes at least one first communication device and at least two second communication devices according to the above aspects.
  • the system may further include other devices that interact with the first communication device and the second communication device in the solutions provided in this application.
  • all or part may be 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 includes one or more computer instructions. When the computer instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part.
  • the computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server or data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.).
  • 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 integrated with one or more available media.
  • the available medium may be a magnetic medium (for example, a floppy disk, a hard disk, or a magnetic tape), an optical medium (for example, a high-density digital video disc (digital video disc, DVD)), or a semiconductor medium (for example, SSD).
  • a magnetic medium for example, a floppy disk, a hard disk, or a magnetic tape
  • an optical medium for example, a high-density digital video disc (digital video disc, DVD)
  • a semiconductor medium for example, SSD

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请提供了一种模型训练方法及相关装置。该方法中，第二通信装置接收第一通信装置的第一神经网络参数，并在第一神经网络参数与第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息。第一指示信息用于指示第二通信装置参与第一通信装置的第一神经网络模型的训练。第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值，表明第二神经网络参数对第一神经网络模型收敛的贡献较大。从而第二通信装置是根据第二神经网络参数对第一神经网络模型收敛的贡献大小，确定是否参与第一神经网络模型的训练，可减少第二通信装置的信令开销。

Description

一种模型训练方法及相关装置
本申请要求于2021年11月22日提交中国国家知识产权局、申请号为202111386640.7、申请名称为“一种模型训练方法及相关装置”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
本申请涉及神经网络领域，尤其涉及一种模型训练方法及相关装置。
背景技术
3GPP在第五代移动通信(5th generation mobile communication，5G)网络中通过新增网络数据分析功能(network data analysis function，NWDAF)，引入了人工智能(artificial intelligence，AI)能力。NWDAF负责AI模型的训练。NWDAF训练的AI模型可应用于移动性管理、会话管理和网络自动化等网络自身领域。
目前常采用联邦学习(federated learning，FL)进行AI模型的训练。FL中，各分布节点参与中心节点的每轮训练时，需将上一轮更新的本地神经网络模型发送至中心节点。然后，中心节点将各分布节点的神经网络模型进行融合，获得全局神经网络模型。若全局神经网络模型不收敛，则中心节点将全局神经网络模型广播给各分布节点。各分布节点根据全局神经网络模型更新本地神经网络模型，再采用更新后的本地神经网络模型参与下一轮中心节点的神经网络模型的训练。
然而，经过多轮训练后，某些分布节点的本地神经网络模型对全局神经网络模型收敛的贡献逐渐降低，此时若这些分布节点的神经网络模型仍继续参与中心节点的神经网络模型的训练，将会造成信令开销的浪费。
发明内容
本申请实施例提供了一种模型训练方法及相关装置，可减少信令开销。
第一方面，本申请实施例提供一种模型训练方法。该方法中，第二通信装置接收第一通信装置的第一神经网络参数，并在第一神经网络参数与第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息。该第一指示信息用于指示第二通信装置参与第一通信装置的第一神经网络模型的训练。
本申请实施例中，第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值，表明第二神经网络参数对第一神经网络模型收敛的贡献较大。从而第二通信装置是根据第二神经网络参数对第一神经网络模型收敛的贡献大小，确定是否参与第一神经网络模型的训练，可避免第二通信装置在第二神经网络参数对第一神经网络模型收敛贡献较小时，仍参与第一神经网络模型的训练，进而可减少第二通信装置的信令开销。
一种可选的实施方式中，第一神经网络参数是第一神经网络的模型参数或第一神经网络的梯度；第二神经网络参数是第二神经网络的模型参数或第二神经网络的梯度。
也就是说，第一神经网络参数是第一神经网络的模型参数，第二神经网络参数是第二神经网络的模型参数。可选的，第一神经网络参数是第一通信装置的神经网络的梯度，第二神经网络参数是第二通信装置的神经网络的梯度。从而第二通信装置根据接收的第一神经网络 参数所属的类型，确定第一神经网络参数与第二神经网络参数之间的相关系数。
一种可选的实施方式中，第一神经网络参数是在协作发现资源上接收的，该协作发现资源是在侧行链路配置信息中配置的。也就是说，第二通信装置采用侧行链路配置信息中的协作发现资源接收来自第一通信装置的第一神经网络参数。
一种可选的实施方式中，在上述相关系数小于第一阈值时，第二通信装置还可向第一通信装置发送第二神经网络参数，以使得第一通信装置根据该第二神经网络参数更新第一神经网络模型，从而第一通信装置是采用对第一神经网络模型收敛贡献度较高的第二神经网络参数更新第一神经网络模型的，有利于加快第一神经网络模型的收敛。
一种可选的实施方式中，第二通信装置还可接收来自第一通信装置的控制信号，该控制信号用于指示时频资源，且指示的时频资源用于第二通信装置发送第二神经网络参数。可见，第二通信装置通过接收来自第一通信装置的控制信号，获知到向第一通信装置发送第二神经网络参数的时频资源，进而第二通信装置可在该时频资源上发送第二神经网络参数。
一种可选的实施方式中，接收该控制信号的资源是协作控制资源，协作控制资源是在上述侧行链路配置信息中配置的。也就是说，第二通信装置采用侧行链路配置信息中的协作控制资源接收上述控制信号。
一种可选的实施方式中，第二通信装置还可在协作同步资源上接收同步信号，并根据该同步信号，与第一通信装置进行同步。从而第二通信装置与第一通信装置同步后，可与第一通信装置通信。其中，协作同步资源可是在上述侧行链路配置信息中配置的。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源可是预先配置的，或是动态指示的，或是非授权频谱资源。
一种可选的实施方式中，第一神经网络参数是第一神经网络的模型参数，第二神经网络参数是第二神经网络的模型参数时，上述第一神经网络参数与第二神经网络参数之间的相关系数是根据第一参数和第二参数确定的。
其中，第一参数是第二通信装置对第一神经网络模型输入训练数据时，第一神经网络模型输出的参数；所述第一神经网络模型是根据所述第一神经网络的模型参数确定的；第二参数是第二通信装置对第二通信装置的第二神经网络模型输入该训练数据时，第二神经网络模型输出的参数。也就是说，第一参数和第二参数是第二通信装置分别对第一神经网络模型和第二神经网络模型输入相同的训练数据时，第一神经网络模型和第二神经网络模型分别输出的参数。
另一种可选的实施方式中，第一神经网络参数是第一神经网络的梯度，第二神经网络参数是第二神经网络的梯度时，第一神经网络参数与第二神经网络参数之间的相关系数是根据第一神经网络参数的概率密度分布和第二神经网络参数的概率密度分布确定的。
可见，第二通信装置可根据接收的第一神经网络参数所属的类型，灵活采用相应的方式确定第一神经网络参数与第二神经网络参数之间的相关系数。
第二方面，本申请还提供了一种模型训练方法。该方面的模型训练方法与第一方面所述的模型训练方法相对应，该方面的模型训练方法是从第一通信装置侧进行阐述的。该方法中，第一通信装置发送该第一通信装置的第一神经网络参数。第一通信装置接收来自第二通信装置的第一指示信息，第一指示信息是第二通信装置在第一神经网络参数与第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时发送的，第一指示信息用于指示第二通信装置参与第一通信装置的第一神经网络模型的训练。
可见，本申请实施例中，第一通信装置接收的第一指示信息是第二通信装置在第一神经 网络参数与第二神经网络参数的相关系数小于第一阈值时发送的，从而第二通信装置是根据第二神经网络参数对第一神经网络模型收敛的贡献大小，确定是否参与第一神经网络模型的训练，进而使得第一通信装置后续不是根据所有第二通信装置的第二神经网络参数更新第一神经网络模型，而是根据对第一神经网络模型收敛贡献较大的第二神经网络参数更新第一神经网络模型，可减少第一通信装置的信令开销。
一种可选的实施方式中，第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
第一神经网络参数是第一神经网络的模型参数，第二神经网络参数是第二神经网络的模型参数。可选的，第一神经网络参数是第一通信装置的神经网络的梯度，第二神经网络参数是第二通信装置的神经网络的梯度。
一种可选的实施方式中，第一神经网络参数是在协作发现资源上发送的，该协作发现资源是在侧行链路配置信息中配置的。也就是说，第一通信装置采用侧行链路配置信息中的协作发现资源向第一通信装置发送第一神经网络参数。
一种可选的实施方式中，第一通信装置还可接收来自第二通信装置的第二神经网络参数，并根据第二神经网络参数，更新第一神经网络模型。可见，第一通信装置是根据反馈了第一指示信息的第二通信装置的第二神经网络参数，更新第一神经网络模型的，从而可节省第一通信装置的信令开销。
一种可选的实施方式中，第一通信装置还可向第二通信装置发送控制信号，该控制信号用于指示时频资源，指示的时频资源用于第二通信装置发送第二神经网络参数。可见，第一通信装置通过控制信号，向第二通信装置指示了发送第二神经网络参数的时频资源，有利于第二通信装置采用该时频资源发送第二神经网络参数。
一种可选的实施方式中，发送控制信号的资源是协作控制资源，协作控制资源是在上述侧行链路配置信息中配置的。也就是说，第一通信装置采用侧行链路配置信息中的协作控制资源发送上述控制信号。
一种可选的实施方式中，第一通信装置还可在协作同步资源上发送同步信号，以使得第二通信装置根据该同步信号与第一通信装置进行同步。另外，该协作同步资源可是在上述侧行链路配置信息中配置的。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源可是预先配置的，或是动态指示的，或是非授权频谱资源。
第三方面，本申请还提供了一种模型训练方法。该方法中，第一通信装置发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的。第一通信装置接收来自第二通信装置的第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
可见，本申请实施例中，第一通信装置将待训练的神经网络模型拆分为了多个训练任务，并通过协作请求信息将多个训练任务广播给周边的各第二通信装置，以请求各第二通信装置参与多个训练任务的训练。第一通信装置通过接收第二指示信息获知第二通信装置自身可参与的训练任务。该方式中，第一通信装置获知到周边的第二通信装置协助参与待训练的神经网络模型的训练，从而可降低对第一通信装置能力的需求。
一种可选的实施方式中，上述协作请求信息在协作发现资源上发送的，协作发现资源是在侧行链路配置信息中配置的。可见，第一通信装置采用侧行链路配置信息中的协作发现资 源发送协作请求信息。
一种可选的实施方式中，第二通信装置发送的第二指示信息指示的第一训练任务包括多个训练任务，此时第一通信装置还可发送第三指示信息，第三指示信息用于指示第一训练任务中的其中一个训练任务。
可理解的，第一通信装置是根据接收的各第二指示信息指示的训练任务，确定的该第三指示信息，以保证每个参与训练的第二通信装置训练的训练任务不重复。
另一种可选的实施方式中，多个第二通信装置发送的第二指示信息指示的第一训练任务是多个训练任务中的相同训练任务，此时第一通信装置也可通过第三指示信息，向其中的一个第二通信装置指示参与训练的训练任务。从而接收到第三指示信息的第二通信装置获知到需进行训练的训练任务，而未接收到第三指示信息的第二通信装置不参与训练。
一种可选的实施方式中，第一通信装置还可向第二通信装置发送第四指示信息，第四指示信息用于指示第二通信装置需接收的第一输出、第一输出对应的时频资源位置，和/或需发送的第二输出、第二输出对应的时频资源位置。第一输出是第一通信装置训练的神经网络模型的输出，或者是除第二通信装置外的其他第二通信装置训练的神经网络模型的输出；第二输出是第二通信装置训练的神经网络模型的输出。
可见，第一通信装置通过第四指示信息告知了参与训练的第二通信装置需接收的参数、需接收的参数对应的时频资源位置，和/或，需发送的参数、需发送的参数对应的时频资源位置，从而有利于参与训练的任一第二通信装置在进行训练任务的训练过程中，进行相应输出的接收和/或发送，以保障其他各第二通信装置的协作训练。
一种可选的实施方式中，发送第四指示信息的资源是协作控制资源，协作控制资源是在侧行链路配置信息中配置的。可见，第一通信装置是采用侧行链路配置信息中的协作控制资源发送第四指示信息的。
一种可选的实施方式中，第一通信装置还可在协作同步资源上发送同步信号，以使得第二通信装置根据该同步信号与第一通信装置进行同步。另外，该协作同步资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源可是预先配置的，或是动态指示的，或是非授权频谱资源。
第四方面，本申请还提供了一种模型训练方法，该方面的模型训练方法与第三方面所述的模型训练方法相对应，该方面的模型训练方法是从第二通信装置侧进行阐述的。该方法中，第二通信装置接收协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的。第二通信装置确定参与第一训练任务的训练时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
可见，本申请实施例中，第二通信装置在确定参与第一通信装置请求的多个训练任务中的第一训练任务的训练时，发送指示参与该第一训练任务的指示信息，以告知第一通信装置自身可协助第一通信装置参与第一训练任务的训练，从而有利于降低对第一通信装置能力的需求。
一种可选的实施方式中，协作请求信息在协作发现资源上接收的，协作发现资源是在侧行链路配置信息中配置的。可见，第二通信装置采用侧行链路配置信息中的协作发现资源接收协作请求信息。
一种可选的实施方式中，第二通信装置还可接收第三指示信息，第三指示信息用于指示 第一训练任务中的其中一个训练任务，从而第二通信装置获知参与训练的训练任务。
一种可选的实施方式中，第二通信装置还可接收第四指示信息，第四指示信息用于指示第二通信装置接收的第一输出、第一输出对应的时频资源位置，和/或发送的第二输出、第二输出对应的时频资源位置。第一输出是第一通信装置训练的神经网络模型的输出，或者是除第二通信装置外的其他第二通信装置训练的神经网络模型的输出；第二输出是第二通信装置训练的神经网络模型的输出。从而，第二通信装置在进行训练任务的训练过程中，进行相应输出的接收和/或发送，以保障其他各第二通信装置的协作训练。
一种可选的实施方式中，接收第四指示信息的资源是协作控制资源，协作控制资源是在侧行链路配置信息中配置的。可见，第二通信装置是采用侧行链路配置信息中的协作控制资源接收第四指示信息的。
一种可选的实施方式中，第二通信装置还可在协作同步资源上接收同步信号，并根据同步信号，与第一通信装置进行同步。从而第二通信装置与第一通信装置同步后，可与第一通信装置通信。另外，协作同步资源是在上述侧行链路配置信息中配置的。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源可是预先配置的，或是动态指示的，或是非授权频谱资源。
第五方面，本申请还提供一种通信装置。该通信装置具有实现上述第一方面所述的第二通信装置的部分或全部功能，或者具有实现上述第二方面所述的第一通信装置的部分或全部功能，或者具有实现上述第三方面所述的第一通信装置的部分或全部功能，或者具有实现上述第四方面所述的第二通信装置的部分或全部功能。比如，该通信装置的功能可具备本申请中第一方面所述的第二通信装置的部分或全部实施例中的功能，也可以具备单独实施本申请中的任一个实施例的功能。所述功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的设计中，该通信装置的结构中可包括处理单元和通信单元，所述处理单元被配置为支持通信装置执行上述方法中相应的功能。所述通信单元用于支持该通信装置与其他通信装置之间的通信。所述通信装置还可以包括存储单元，所述存储单元用于与处理单元和通信单元耦合，其保存通信装置必要的程序指令和数据。
一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信单元进行数据/信令收发；通信单元，用于接收第一通信装置的第一神经网络参数；通信单元，还用于在第一神经网络参数与通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息；第一指示信息用于指示该通信装置参与第一通信装置的第一神经网络模型的训练。
另外，该方面中，通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处不再详述。
另一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信单元进行数据/信令收发；该通信单元，用于发送该通信装置的第一神经网络参数；该通信单元，还用于接收来自第二通信装置的第一指示信息；第一指示信息是第二通信装置在第一神经网络参数与第二装置的第二神经网络参数之间的相关系数小于第一阈值时发送的；第一指示信息用于指示第二通信装置参与该通信装置的第一神经网络模型的训练。
另外，该方面中，通信装置其他可选的实施方式可参见上述第二方面的相关内容，此处不再详述。
又一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信 单元进行数据/信令收发；该通信单元，用于发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；该通信单元，还用于接收来自第二通信装置的第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
另外，该方面中，通信装置其他可选的实施方式可参见上述第三方面的相关内容，此处不再详述。
又一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信单元进行数据/信令收发；该通信单元，用于接收协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；该通信单元，还用于确定参与第一训练任务的训练时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
另外，该方面中，通信装置其他可选的实施方式可参见上述第四方面的相关内容，此处不再详述。
作为示例，通信单元可以为收发器或通信接口，存储单元可以为存储器，处理单元可以为处理器。
一种实施方式中，所述通信装置包括：处理器和收发器，处理器用于控制收发器进行数据/信令收发；收发器，用于接收第一通信装置的第一神经网络参数；收发器，还用于在第一神经网络参数与通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息；第一指示信息用于指示该通信装置参与第一通信装置的第一神经网络模型的训练。
另外，该方面中，上行通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处不再详述。
另一种实施方式中，所述通信装置包括：处理器和收发器，处理器用于控制收发器进行数据/信令收发；该收发器，用于发送该通信装置的第一神经网络参数；收发器，还用于接收来自第二通信装置的第一指示信息；第一指示信息是第二通信装置在第一神经网络参数与第二装置的第二神经网络参数之间的相关系数小于第一阈值时发送的；第一指示信息用于指示第二通信装置参与该通信装置的第一神经网络模型的训练。
另外，该方面中，通信装置其他可选的实施方式可参见上述第二方面的相关内容，此处不再详述。
又一种实施方式中，所述通信装置包括：处理器和收发器，处理器用于控制收发器进行数据/信令收发；收发器，用于发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；收发器，还用于接收来自第二通信装置的第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
另外，该方面中，通信装置其他可选的实施方式可参见上述第三方面的相关内容，此处不再详述。
又一种实施方式中，所述通信装置包括：处理器和收发器，处理器用于控制收发器进行数据/信令收发；该收发器，用于接收协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；收发器，还用于确定参与第一训练任务的训练时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
另外，该方面中，通信装置其他可选的实施方式可参见上述第四方面的相关内容，此处不再详述。
另一种实施方式中，该通信装置为芯片或芯片系统。所述处理单元也可以体现为处理电路或逻辑电路；所述收发单元可以是该芯片或芯片系统上的输入/输出接口、接口电路、输出电路、输入电路、管脚或相关电路等。
在实现过程中，处理器可用于进行，例如但不限于，基带相关处理，收发器可用于进行，例如但不限于，射频收发。上述器件可以分别设置在彼此独立的芯片上，也可以至少部分的或者全部的设置在同一块芯片上。例如，处理器可以进一步划分为模拟基带处理器和数字基带处理器。其中，模拟基带处理器可以与收发器集成在同一块芯片上，数字基带处理器可以设置在独立的芯片上。随着集成电路技术的不断发展，可以在同一块芯片上集成的器件越来越多。例如，数字基带处理器可以与多种应用处理器(例如但不限于图形处理器，多媒体处理器等)集成在同一块芯片之上。这样的芯片可以称为系统芯片(System on a Chip,SoC)。将各个器件独立设置在不同的芯片上，还是整合设置在一个或者多个芯片上，往往取决于产品设计的需要。本申请实施例对上述器件的实现形式不做限定。
第六方面，本申请还提供一种处理器，用于执行上述各种方法。在执行这些方法的过程中，上述方法中有关发送上述信息和接收上述信息的过程，可以理解为由处理器输出上述信息的过程，以及处理器接收输入的上述信息的过程。在输出上述信息时，处理器将该上述信息输出给收发器，以便由收发器进行发射。该上述信息在由处理器输出之后，还可能需要进行其他的处理，然后才到达收发器。类似的，处理器接收输入的上述信息时，收发器接收该上述信息，并将其输入处理器。更进一步的，在收发器收到该上述信息之后，该上述信息可能需要进行其他的处理，然后才输入处理器。
基于上述原理，举例来说，前述方法中提及的接收第一通信装置的第一神经网络参数可以理解为处理器输入第一通信装置的第一神经网络参数。
对于处理器所涉及的发送和接收等操作，如果没有特殊说明，或者，如果未与其在相关描述中的实际作用或者内在逻辑相抵触，则均可以更加一般性的理解为处理器输出和接收、输入等操作，而不是直接由射频电路和天线所进行的发送和接收操作。
在实现过程中，上述处理器可以是专门用于执行这些方法的处理器，也可以是执行存储器中的计算机指令来执行这些方法的处理器，例如通用处理器。上述存储器可以为非瞬时性(non-transitory)存储器，例如只读存储器(Read Only Memory，ROM)，其可以与处理器集成在同一块芯片上，也可以分别设置在不同的芯片上，本申请实施例对存储器的类型以及存储器与处理器的设置方式不做限定。
第七方面，本申请还提供了一种通信系统，该系统包括上述方面的至少一个第一通信装置、至少两个第二通信装置。在另一种可能的设计中，该系统还可以包括本申请提供的方案中与第一通信装置、第二通信装置进行交互的其他设备。
第八方面，本申请提供了一种计算机可读存储介质，用于储存指令，当所述指令被计算机运行时，实现上述第一方面至第四方面任一项所述的方法。
第九方面，本申请还提供了一种包括指令的计算机程序产品，当其在计算机上运行时，实现上述第一方面至第四方面任一项所述的方法。
第十方面，本申请提供了一种芯片系统，该芯片系统包括处理器和接口，所述接口用于获取程序或指令，所述处理器用于调用所述程序或指令以实现或者支持第二通信装置实现第一方面所涉及的功能，或者用于调用所述程序或指令以实现或者支持第一通信装置备实现第 二方面所涉及的功能，用于调用所述程序或指令以实现或者支持第一通信装置实现第三方面所涉及的功能，用于调用所述程序或指令以实现或者支持第二通信装置实现第四方面所涉及的功能。例如，确定或处理上述方法中所涉及的数据和信息中的至少一种。在一种可能的设计中，所述芯片系统还包括存储器，所述存储器，用于保存终端必要的程序指令和数据。该芯片系统，可以由芯片构成，也可以包括芯片和其他分立器件。
附图说明
图1是本申请实施例提供的一种通信系统的结构示意图；
图2是本申请实施例提供的另一种通信系统的结构示意图；
图3是本申请实施例提供的一种联邦学习的系统示意图；
图4是本申请实施例提供的一种分割学习的示意图；
图5是本申请实施例提供的一种模型训练方法的交互流程示意图；
图6是本申请实施例提供的另一种模型训练方法的交互流程示意图；
图7是本申请实施例提供的又一种模型训练方法的交互流程示意图；
图8是本申请实施例提供的一种待训练的神经网络模型拆分示意图；
图9是本申请实施例提供的又一种模型训练方法的交互流程示意图；
图10是本申请实施例提供的一种通信装置的结构示意图；
图11是本申请实施例提供的另一种通信装置的结构示意图；
图12是本申请实施例提供的一种芯片的结构示意图。
具体实施方式
下面结合本申请实施例中的附图对本申请实施例中的技术方案进行清楚、完整的描述。
为了更好的理解本申请实施例公开的模型训练方法，对本申请实施例适用的通信系统进行描述。
本申请实施例可应用于第五代移动通信(5th generation mobile communication，5G)系统、卫星通信及短距等无线通信系统中，系统架构如图1所示。无线通信系统可以包括一个或多个网络设备以及一个或多个终端设备。无线通信系统也可以进行点对点通信，如多个终端设备之间互相通信。
可理解的，本申请实施例提及的无线通信系统包括但不限于：窄带物联网系统(narrow band-internet of things，NB-IoT)、长期演进系统(long term evolution，LTE)，5G移动通信系统的三大应用场景：增强移动宽带(enhanced mobile broadband，eMBB)、超可靠低时延通信(ultra reliable low latency communication，URLLC)和海量机器类通信(massive machine type of communication，mMTC)，无线保真(wireless fidelity，WiFi)系统，或者5G之后的移动通信系统等。
请参见图2，图2为本申请实施例提供的另一种通信系统的结构示意图。该通信系统可包括但不限于一个第一通信装置201、两个第二通信装置202。图2所示的设备数量和形态用于举例并不构成对本申请实施例的限定，实际应用中可以包括两个或两个以上的第二通信装置202，三个以上的第二通信装置202。其中，第一通信装置可以为网络设备，也可为终端设备，第二通信装置为终端设备。本申请实施例中以第一通信装置和第二通信装置均为终端设备为例进行阐述。
本申请实施例中，第二通信装置为第一通信装置的周边终端设备，即各第二通信装置和第一通信装置位于同一小区内。第一通信装置和第二通信装置中均设置有神经网络模型，第二通信装置可协作第一通信装置，参与第一通信装置的第一神经网络模型的训练。
本申请实施例中，网络设备是具有无线收发功能的设备，用于与终端设备进行通信，可以是LTE中的演进型基站(evolved Node B，eNB或eNodeB)；或者5G网络中的基站或者未来演进的公共陆地移动网络(public land mobile network，PLMN)中的基站，宽带网络业务网关(broadband network gateway，BNG)，汇聚交换机或者非第三代合作伙伴项目(3rd generation partnership project，3GPP)接入设备等。可选的，本申请实施例中的网络设备可以包括各种形式的基站，例如：宏基站、微基站(也称为小站)、中继站、接入点、未来实现基站功能的设备、WiFi系统中的接入节点，传输接收点(transmitting and receiving point，TRP)、发射点(transmitting point，TP)、移动交换中心以及设备到设备(Device-to-Device，D2D)、车辆外联(vehicle-to-everything，V2X)、机器到机器(machine-to-machine，M2M)通信中承担基站功能的设备等，本申请实施例对此不作具体限定。
网络设备可以和核心网设备进行通信交互，向终端设备提供通信服务。核心网设备例如为5G网络核心网(core network，CN)中的设备。核心网作为承载网络提供到数据网络的接口，为终端提供通信连接、认证、管理、策略控制以及对数据业务完成承载等。
本申请实施例所涉及到的终端设备可以包括各种具有无线通信功能的手持设备、车载设备、可穿戴设备、计算设备或连接到无线调制解调器的其它处理设备。终端设备也可以指用户设备(user equipment，UE)、接入终端、用户单元(subscriber unit)、用户代理、蜂窝电话(cellular phone)、智能手机(smart phone)、无线数据卡、个人数字助理(personal digital assistant，PDA)电脑、平板型电脑、无线调制解调器(modem)、手持设备(handset)、膝上型电脑(laptop computer)、机器类型通信(machine type communication，MTC)终端、高空飞机上搭载的通信设备、可穿戴设备、无人机、机器人、设备到设备通信(device-to-device，D2D)中的终端、车到一切(vehicle to everything，V2X)中的终端、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端、远程医疗(remote medical)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端、智慧家庭(smart home)中的无线终端或者未来通信网络中的终端设备等，本申请不作限制。
为了便于理解本申请公开的实施例，作以下两点说明。
(1)本申请公开的实施例中场景以无线通信网络中5G新空口(new radio，NR)网络的场景为例进行说明，应当指出的是，本申请公开的实施例中的方案还可以应用于其他无线通信网络中，相应的名称也可以用其他无线通信网络中的对应功能的名称进行替代。
(2)本申请公开的实施例将围绕包括多个设备、组件、模块等的系统来呈现本申请的各个方面、实施例或特征。应当理解和明白的是，各个系统可以包括另外的设备、组件、模块等，并且/或者可以并不包括结合附图讨论的所有设备、组件、模块等。此外，还可以使用这些方案的组合。
为了更好的理解本申请实施例公开的模型训练方法，对本申请实施例涉及的相关概念进行简单的介绍。
1.联邦学习FL。
联邦学习是在充分保障用户数据隐私和安全的前提下，通过促使各个周边设备和中心端服务器协同合作来高效地完成模型的一种学习方式。FL算法如下：
(1)在中心端服务器的第i∈[1,T]轮训练中，终端设备m采用局部数据集训练本地的神经网络模型，并将本地的梯度
Figure PCTCN2022133214-appb-000001
通过空口传输至中心端服务器。其中，γ表示回传的梯度参数的数量，T为阈值次数，且T大于或等于2，M为终端设备的总个数，
Figure PCTCN2022133214-appb-000002
代表第i轮训练中，终端设备m的第γ个梯度参数对应的梯度；
(2)中心端服务器汇总收集来自全部(部分)终端设备的梯度，并对其进行加权求均得到新的全局梯度：
Figure PCTCN2022133214-appb-000003
(3)中心端根据新的全局梯度更新本地神经网络模型，获得更新后的神经网络模型，若更新后的神经网络模型不收敛，且训练次数未达到阈值，则将新的全局梯度广播给各终端设备。终端设备接收到该新的全局梯度后，根据该新的全局梯度更新自身本地神经网络模型，直至中心端的神经网络模型收敛或训练轮数达到阈值次数。
示例性的，联邦学习的系统示意图如图3所示。图3中，中心端为网络设备，周边设备为各种终端设备。联邦学习中，各终端设备将本地计算的梯度通过无线信道上传至网络设备。网络设备进行多个本地梯度的汇总，即将接收的多个梯度进行加权求均处理，获得全局梯度，并根据该全局梯度更新本地的神经网络模型。若更新后的神经网络模型仍不收敛，且训练次数未达到阈值次数，则将该全局梯度广播给各终端设备。任一终端设备接收到该全局梯度后，采用该全局梯度更新本终端设备的神经网络模型，并将更新的神经网络模型的梯度上传至网络设备，进行下一轮神经网络模型的训练，直至网络设备的神经网络模型收敛，或训练次数达到阈值次数。
2.分割学习(split learning)。
分割学习如图4所示。在分割学习中，完整的神经网络模型被分割为两部分(即两个子网络)，神经网络的一部分子网络部署在分布式节点上，另一部分子网络部署在中心节点上。完整的神经网络被分割的地方被称为“分割层”。神经网络模型的前向推理过程中，分布式节点将本地数据输入本地的子网络，推理到分割层，将分割层的结果F1通过通信链路发送到中心节点，中心节点将收到的F1输入自身部署的另一个子网络，并继续进行前向推理，得到最终的推理结果。神经网络模型训练的梯度反向传递中，梯度通过中心节点的子网络反向传递到分割层，得到反向传递结果G1，然后中心节点将G1发送给分布式节点，使得G1继续在分布式节点的子网络上进行梯度反向传递。
分割学习的前向推理和梯度反向传递过程中，只涉及一个分布式节点和一个中心节点。训练好的分布式节点上的子网络可以保存在分布式节点本地或特定的模型存储服务器上。当有新的分布式节点加入学习系统时，该新的分布式节点可以先下载已训练好的分布式节点子网络，再使用本地数据进行进一步的训练。
从上述的联邦学习可知，目前的分布式学习中，中心节点汇总各分布节点上报的本地模型，并将各分布节点的神经网络模型进行融合处理，再下发给各分布节点，进行下一轮训练，直至中心节点的神经网络模型收敛。然而在经过若干轮的训练后，部分分布节点对中心节点的神经网络模型的收敛的贡献会逐渐降低，此时若继续参与中心节点的神经网络模型的训练，带来的增益可能不足以弥补中心节点的信令开销。
本申请实施例提供了一种模型训练方法100。模型训练方法100中，第一通信装置发送该第一通信装置的第一神经网络参数。第二通信装置接收第一通信装置的第一神经网络参数， 且第二通信装置在第一神经网络参数与第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息，该第一指示信息用于指示第二通信装置参与第一通信装置的第一神经网络模型的训练。从而第二通信装置接收该第一指示信息。第二通信装置在第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置反馈参与第一神经网络模型的训练，可避免第二通信装置在第一神经网络参数与第二神经网络参数之间的相关系数等于或大于第一阈值时，仍参与第一神经网络模型的训练，从而可减少第二通信装置的信令开销。
本申请实施例还提供了一种模型训练方法200。模型训练方法200中，第一通信装置发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的。第二通信装置接收协作请求信息。第二通信装置确定参与第一训练任务时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。第一通信装置接收来自第二通信装置的第二指示信息。可见，第一通信装置将待训练的神经网络模型拆分为了多个训练任务，并通过协作请求信息将多个训练任务广播给周边的各第二通信装置，周边的第二通信装置反馈第二指示信息，以通过第二指示信息告知第一通信装置自身可参与训练的训练任务。也就是说，周边的各第二通信装置协助参与第一通信装置的待训练的神经网络模型的训练，从而可降低对第一通信装置能力的需求。
本申请实施例提出一种模型训练方法100，图5是该模型训练方法100的交互流程示意图。该模型训练方法100从第一通信装置与第二通信装置之间交互的角度进行阐述。该模型训练方法100包括但不限于以下步骤：
S101.第一通信装置发送第一通信装置的第一神经网络参数。
可理解的，第一神经网络参数是神经网络模型，或者是神经网络的梯度，或者还可以是训练神经网络模型的训练数据。也就是说，第一神经网络参数是第一通信装置的第一神经网络模型，或者是第一神经网络的梯度，或者是训练第一神经网络模型的训练数据。其中，神经网络模型包括神经网络中包括的神经元，以及每层神经元之间的权重。
一种可选的实施方式中，第一通信装置基于自身的本地数据集，训练第一神经网络模型。第一通信装置在经过阈值轮数的训练后，第一神经网络模型仍未满足预设收敛条件，则可触发协作机制。第一通信装置触发协作机制可以包括：第一通信装置向网络设备发送请求消息，以请求网络设备给该第一通信装置配置用于进行协作训练的相关资源。可选地，该请求消息可以是按需系统消息块(on demand system information block，on demand SIB)。网络设备接收到来自第一通信装置的请求消息后，为第一通信装置和第一通信装置的周边设备(各第二通信装置)发送侧行链路配置信息。
其中，侧行链路配置信息可以是SIB_AI_sidelink，且侧行链路配置信息用于配置协作同步资源、协作发现资源或协作控制资源。其中，协作同步资源可以是人工智能协作同步(AI-cooperation-sync)资源，该协作同步资源用于各第二通信装置与第一通信装置进行同步；协作发现资源可以是人工智能协作发现(AI-cooperation-discover)资源，该协作发现资源用于第一通信装置发送第一神经网络参数，还用于第二通信装置监听第一通信装置的第一神经网络参数；协作控制资源可以是人工智能控制发现(AI-cooperation-control)资源，该协作控制资源用于第一通信装置为各第二通信装置指示发送第二神经网络参数的资源。
从而，该方式中，第一神经网络参数是在协作发现资源上发送的，例如，第一神经网络参数是第一通信装置在AI-cooperation-discover资源上发送的。协作发现资源是在上述侧行链路配置信息中配置的。
另一种可选的实施方式中，当第一通信装置是网络设备时，可自行预配置上述侧行链路配置信息，并下发给各第二通信装置。
一种可选的实施方式中，第一通信装置在发送第一神经网络参数之前，还可在上述协作同步资源上发送同步信号，以使得第二通信装置根据该同步信号与第一通信装置进行同步。
S102.第二通信装置接收第一通信装置的第一神经网络参数。
一种可选的实施方式中，网络设备在接收到来自第一通信装置的请求消息后，为各第二通信装置发送侧行链路配置信息，从而第二通信装置在上述侧行链路配置信息所配置的协作发现资源上接收第一神经网络参数。
一种可选的实施方式中，第二通信装置接收第一神经网络参数之前，还可在上述侧行链路配置信息所配置的协作同步资源上监听同步信号，并根据该同步信号与第一通信装置进行同步。从而第二通信装置完成与第一通信装置的同步后，可与第一通信装置进行通信，比如接收来自第一通信装置的第一神经网络参数。
S103.第二通信装置在第一神经网络参数与第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置发送第一指示信息，第一指示信息用于指示第二通信装置参与第一通信装置第一神经网络模型的训练。
可理解的，当上述第一神经网络参数是第一神经网络的模型参数时，第二神经网络参数是第二神经网络的模型参数，第一神经网络是第一通信装置的神经网络，第二神经网络是第二通信装置的神经网络；当第一神经网络参数是第一神经网络模型的梯度时，第二神经网络参数是第二神经网络模型的梯度；当第一神经网络参数是训练第一神经网络模型的训练数据时，第二神经网络参数是训练第二神经网络模型的训练数据。其中，模型参数包括神经网络结构、神经网络结构中神经元之间的权重等。
其中，第一阈值是第二通信装置预先设定的。第二通信装置的第二神经网络参数是第二通信装置接收到第一神经网络参数后，根据第一神经网络参数更新第二通信装置的本地神经网络模型后的神经网络模型的神经网络参数。从而第二通信装置将更新后的神经网络模型的神经网络参数与第一神经网络参数进行对比，以确定第一神经网络参数与第二神经网络参数之间的相关系数。
示例性的，第二通信装置接收第一神经网络参数之前的本地神经网络模型为神经网络模型X，第二通信装置接收到第一神经网络参数之后，根据该第一神经网络参数对神经网络模型X进行更新，获得神经网络模型Y。该神经网络模型Y为第二神经网络模型，进而第二通信装置将神经网络模型Y的神经网络参数与第一神经网络参数进行对比，以确定第一神经网络参数与第二神经网络参数之间的相关系数。
另外，当第一神经网络参数和第二神经网络参数所属的类型不同时，其两者之间的相关系数的确定方式也不相同。以下，结合第一神经网络参数和第二神经网络参数所属的类型，阐述第一神经网络参数与第二神经网络参数之间的相关系数的确定方式：
1.第一神经网络参数是第一神经网络的模型参数，第二神经网络参数是第二神经网络的模型参数。
当第一神经网络参数是第一神经网络的模型参数，第二神经网络参数是第二神经网络的模型参数时，第一神经网络参数与第二神经网络参数之间的相关系数是根据第一参数和第二 参数确定的。其中，第一参数是第二通信装置对第一神经网络模型输入训练数据时，第一神经网络模型输出的参数；第一神经网络模型是根据第一神经网络的模型参数确定的；第二参数是第二通信装置对第二通信装置的第二神经网络模型也输入该训练数据时，第二神经网络模型输出的参数。
也就是说，第二通信装置接收到第一神经网络参数后，根据第一神经网络的模型参数确定第一神经网络模型，再根据第一神经网络模型和第二神经网络模型确定第一参数和第二参数，并根据第一参数和第二参数确定第一神经网络参数与第二神经网络参数之间的相关系数。其中，第一参数和第二参数是第二通信装置对第一神经网络模型和第二神经网络模型输入相同的训练数据时，第一神经网络模型和第二神经网络模型分别输出的参数。
一种可能的实现中，第二通信装置将第一参数和第二参数的协方差除以第一参数的标准差与第二参数的标准差之积的结果作为第一神经网络参数与第二神经网络参数之间相关性的评价准则。例如，第二通信装置对第一神经网络模型和第二神经网络模型输入相同的训练数据时，第一神经网络模型和第二神经网络模型分别输出X、Y，即X为第一参数，Y为第二参数，那么第一神经网络参数与第二神经网络参数之间的相关系数为：
Figure PCTCN2022133214-appb-000004
其中，Cov(X,Y)表示X和Y的协方差，σ X、σ Y分别表示X和Y的标准差。
2.第一神经网络参数是第一神经网络模型的梯度，第二神经网络参数是第二神经网络模型的梯度；或者，第一神经网络参数是训练第一神经网络模型的训练数据，第二神经网络参数是训练第二神经网络模型的训练数据。
当第一神经网络参数是第一神经网络模型的梯度，第二神经网络参数是第二神经网络模型的梯度时，第一神经网络参数与第二神经网络参数之间的相关系数是根据第一神经网络模型的梯度的概率分布与第二神经网络模型的梯度的概率分布确定的。当第一神经网络参数是训练第一神经网络模型的训练数据，第二神经网络参数是训练第二神经网络模型的训练数据时，第一神经网络参数与第二神经网络参数之间的相关系数是根据训练第一神经网络模型的训练数据的概率分布与训练第二神经网络模型的训练数据的概率分布确定的。
一种可能的实现中，第一神经网络参数与第二神经网络参数之间的相关系数是根据第一神经网络参数的概率分布、第二神经网络参数的概率分布，以及海林格距离(hellinger distance)的定义确定的。即第一神经网络参数与第二神经网络参数之间的相关系数为：
Figure PCTCN2022133214-appb-000005
其中，Z a、Z b分别表示第一神经网络参数和第二神经网络参数，S(Z a)、S(Z b)分别表示第一神经网络参数的概率分布、第二神经网络参数的概率分布。
可理解的，第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值，意味着第二神经网络参数与第一神经网络参数的相关性较大，从而第二神经网络参数对第一神经网络模型的收敛贡献较大。也就是说，第二通信装置是根据第二神经网络参数对第一神经网络模型收敛的贡献大小，确定是否参与第一神经网络模型的训练。第二通信装置在第二神经网络参数对第一神经网络模型的收敛贡献较大时，确定参数第一神经网络模型的训练，且通过第一指示信息告知给第一通信装置。从而可避免第二通信装置在第二神经网络参数对第一神经网络模型收敛的贡献较低时，仍参与第一神经网络模型的训练，进而可减少第二通信装置的信令开销，即减少第一通信装置不必要的传输开销。
一种可选的实施方式中，上述侧行链路配置信息中还配置了第二通信装置发送第一指示信息的协作响应资源，从而第二通信装置在该协作响应资源上向第一通信装置发送第一指示信息。
另一种可选的实施方式中，第二通信装置还可在上述协作控制资源上向第一通信装置发送第一指示信息。
另一种可选的实施方式中，第二通信装置在第一神经网络参数与第二神经网络参数的相关系数等于或大于第一阈值时，第二通信装置向第一通信装置发送第五指示信息，第五指示信息用于指示第二通信装置不参与第一神经网络模型的训练。也就是说，第二通信装置在确定不参与第一神经网络模型的训练时，通过第五指示信息向第一通信装置指示不参与第一神经网络模型的训练，从而第一通信装置获知到该第二通信装置不参与此轮神经网络参数的训练。
又一种可选的实施方式中，第一通信装置和第二通信装置预先约定，第二通信装置在第一神经网络参数与第二神经网络参数的相关系数等于或大于第一阈值时，不向第一通信装置反馈任何信息，即第二通信装置在本轮确定不参与第一神经网络模型的训练时，不做任何处理。从而第一通信装置在预设时间内未接收到来自第二通信装置的反馈信息时，确定该第二通信装置不参与本轮第一神经网络模型的训练，从而可节省系统的信令开销。
第二通信装置在确定不参与本轮第一神经网络模型的训练时，等待下一轮第一通信装置发送神经网络参数，并再次和第二通信装置更新的本地神经网络模型的神经网络参数进行对比，以确定是否参与下一轮的训练。
S104.第一通信装置接收来自第二通信装置的第一指示信息。
第一通信装置可在上述侧行链路配置信息所配置的协作响应资源上接收第一指示信息，也可以是在上述协作发现资源接收第一指示信息。本申请实施例不做限定。
第一通信装置通过接收第一指示信息，获知到愿意参与第一神经网络模型训练的第二通信装置，从而第二通信装置根据反馈了第一指示信息的第二通信装置的第二神经网络参数更新第一神经网络模型，有利于节省第二通信装置的信令开销，即节省第二通信装置的传输开销。
一种可选的实施方式中，第一通信装置通过第一指示信息获知到愿意参与第一神经网络模型训练的第二通信装置后，向该部分的第二通信装置发送控制信号，控制信号用于指示时频资源，指示的时频资源用于第二通信装置发送第二神经网络参数。
其中，第一通信装置发送控制信号的资源是协作控制资源，该协作控制资源是在上述侧行链路配置信息中配置的。
也就是说，第一通信装置在确定愿意参与第一神经网络模型训练的第二通信装置后，通过向该部分的第二通信装置发送控制信号，向该部分的第二通信装置指示发送第二神经网络参数的时频资源，以使该部分的第二通信装置在各自对应的时频资源上向第一通信装置发送第二神经网络参数。
一种可选的实施方式中，第二通信装置发送第二神经网络参数的时频资源是网络设备动态调度给第一通信装置的。第一通信装置通过控制信号将时频资源调度给反馈了第一指示信息的第二通信装置。第一通信装置给不同第二通信装置调度的时频资源是不相同的。该方式中，第一通信装置每次给反馈了第一指示信息的第二通信装置动态调度时频资源，可使得资源的利用率较高。
另一种可选的实施方式中，第二通信装置发送第二神经网络参数的时频资源是网络设备半静态配置给第一通信装置的。该半静态资源是周期性出现的，从而无需第一通信装置给第二通信装置调度时频资源。但第一通信装置仍需通过控制信号向反馈了第一指示信息的第二通信装置指示第二通信装置中的半静态资源，以激活该半静态资源。进而第二通信装置可采用该半静态资源向第一通信装置发送第二神经网络参数。该方式中，第一通信装置无需给第二通信装置调度时频资源，可减少信令开销。
一种可选的实施方式中，第二通信装置接收到上述控制信号后，采用控制信号指示的时频资源，向第一通信装置发送第二神经网络参数。从而第一通信装置接收来自第二通信装置的第二神经网络参数，并根据第二神经网络参数更新第一神经网络模型。可理解的，第一通信装置接收多个第二通信装置的第二神经网络参数，该多个第二通信装置均是反馈了第一指示信息的第二通信装置，从而第一通信装置根据多个第二神经网络参数更新第一神经网络模型。
可理解的，第一通信装置根据多个第二神经网络参数更新第一神经网络模型是指：第一通信装置将各个第二神经网络模型参数进行平均求和处理，获得处理后的第二神经网络参数，再根据处理后的第二神经网络参数，更新第一神经网络模型。若更新后的第一神经网络模型仍不收敛，且训练次数未达到阈值次数，则第一通信装置将更新后的第一神经网络模型的第一神经网络参数广播给周边的各第二通信装置，以使得各第二通信装置再次根据自身本地的神经网络参数与接收的神经网络参数，决策是否参与下一轮更新后的第一神经网络模型的训练。
如上所述，侧行链路配置信息所配置的协作同步资源、协作发现资源、协作控制资源可是网络设备在接收到请求消息后动态指示的。可选的，侧行链路配置信息所配置的协作同步资源、协作发现资源、协作控制资源也可以是网络设备预先配置的，还可以是非授权频谱资源。本申请实施例对此不做限定。
请参见图6，图6是本申请实施例以第一通信装置为终端设备A、第二通信装置包括终端设备B和终端设备C为例的模型训练方法的交互流程示意图。如图6所示：
601.若终端设备A在采用本地的训练数据进行N轮训练后，终端设备A的第一神经网络模型还未收敛，则终端设备A向周边终端设备(终端设备B和终端设备C)发送同步信号。可选的，N的数值小于阈值次数。
602.终端设备B和终端设备C监听到同步信号后，分别根据该同步信号与终端设备A进行同步，然后监听终端设备A的第一神经网络参数。终端设备B、终端设备C与终端设备A进行同步，是为了保证后续可与终端设备A进行通信。终端设备A发送同步信号的资源，以及终端设备B和终端设备C监听同步信号的资源均可以是上述侧行链路配置信息配置的协作同步资源，不再赘述。
603.终端设备A在上述协作发现资源上广播终端设备A的第一神经网络参数。
604.终端设备B和终端设备C在协作发现资源上监听到第一神经网络参数后，将第一神经网络参数和自身的第二神经网络参数进行对比，判断第一神经网络参数与第二神经网络参数之间的相关系数是否小于第一阈值。
605.终端设备B和终端设备C在两者的相关系数小于第一阈值时，向终端设备A发送用于指示参与第一神经网络模型的训练的第一指示信息。
606.终端设备A接收到来自终端设备B和终端设备C的第一指示信息后，向终端设备B和终端设备C发送控制信号，通过控制信号向终端设备B和终端设备C分别指示发送自身第二神经网络参数的时频资源。
607.终端设备B和终端设备C分别根据指示的时频资源，向终端设备A发送第二各自的第二神经网络参数。
例如，终端设备A向终端设备B发送的控制信号#b用于指示时频资源#b，向终端设备C发送的控制信号#c用于指示时频资源#c。从而终端设备B在时频资源#b上向终端设备A发送终端设备B的第二神经网络参数，终端设备C在时频资源#c上向终端设备A发送终端设备C的第二神经网络参数。
608.终端设备A接收到来自终端设备B的第二神经网络参数和终端设备C的第二神经网络参数后，对两个第二神经网络参数进行融合，即对两个神经网络参数加权求均，获得全局神经网络参数。
609.终端设备A再采用该全局神经网络参数开始第N+1轮训练，即根据该全局神经网络参数更新第一神经网络模型，获得更新后的第一神经网络模型。
若更新后的第一神经网络模型不收敛，且训练次数还未达到阈值次数时，终端设备A将更新后的第一神经网络模型的第一神经网络参数广播，终端设备B和终端设备C再次根据接收的神经网络参数和此时自身本地的神经网络参数，确定是否参与终端设备A的下一轮训练，直至终端设备A的神经网络模型收敛，或者训练次数达到阈值次数。
可见，本申请实施例中，第二通信装置将自身的第二神经网络参数与接收的第一神经网络参数进行比较，在第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值时，向第一通信装置反馈用于指示参与第一神经网络模型训练的第一指示信息，从而第一通信装置可根据反馈了第一指示信息的第二通信装置的第二神经网络参数，更新第一神经网络模型。第一神经网络参数与第二神经网络参数之间的相关系数小于第一阈值，表明第二神经网络参数对第一神经网络模型收敛的贡献较大。从而第二通信装置是根据第二神经网络参数对第一神经网络模型收敛的贡献大小，确定是否参与第一神经网络模型的训练，可避免第二通信装置在第二神经网络参数对第一神经网络模型收敛贡献较小时，仍参与第一神经网络模型的训练，进而可减少第二通信装置的信令开销。另外，第一通信装置不再是根据周边所有第二通信装置的第二神经网络参数更新第一神经网络模型，而是根据反馈了第一指示信息的第二通信装置的第二神经网络参数更新第一神经网络模型，从而可减少第一通信装置的信令开销。
本申请实施例还提出一种模型训练方法200，图7是该模型训练方法200的交互流程示意图。该模型训练方法200也从第一通信装置与第二通信装置之间交互的角度进行阐述。该模型训练方法200包括但不限于以下步骤：
S201.第一通信装置发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的。
可理解的，第一通信装置在自身资源或能力有限，无法独立完成待训练的神经网络模型的训练时，将待训练的神经网络模型进行拆分，拆分为多个训练任务，并以协作请求信息的方式将该多个训练任务广播给周边的第二通信装置，以请求各第二通信装置协助第一通信装置参与多个训练任务的训练。
其中，每个训练任务包括一个简单神经网络模型，每个简单神经网络模型是该待训练的神经网络模型的一部分子网络。
一种可能的实现中，第一通信装置基于切分后子网络的延迟，对待训练的神经网络模型进行拆分。可选的，第一通信装置基于各第二通信装置的算力，对待训练的神经网络模型进行拆分。可选的，第一通信装置基于切分位置产生的数据量，对待训练的神经网络模型进行拆分。
一种可选的实施方式中，第一通信装置根据待训练的神经网络模型的结构，将待训练的神经网络模型平均分成多个训练任务，从而每个训练任务包括的待训练的神经网络模型的神经网络层数相同。本申请对每个训练任务包括的神经网络层数不做限定。
示例性的，如图8所示，第一通信装置将待训练的神经网络模型拆分为3个训练任务，分别为训练任务A、训练任务B、训练任务C，训练任务A、训练任务B、训练任务C均包括两层神经网络。第一通信装置将训练任务C作为自身的训练部分，将训练任务A和训练任务B作为两个训练任务，并以协作请求的方式广播给周边的第二通信装置。
另一种可选的实施方式中，第一通信装置将待训练的神经网络模型拆分为多个训练程度不均的训练任务，即每个训练任务中包括的神经网络层数不相同。从而有利于剩余资源或自身算力较小的第二通信装置参与神经网络层数较少的训练任务的训练，剩余资源或自身算力较大的第二通信装置参与神经网络层数较多的训练任务的训练。
一种可选的实施方式中，第一通信装置在确定无法完成待训练的神经网络模型时，向网络设备发送请求消息(例如，on demand SIB)，以请求网络设备给该第一通信装置配置用于进行协作训练的相关资源。网络设备接收到来自第一通信装置的请求消息后，为第一通信装置和第一通信装置的周边设备(各第二通信装置)发送侧行链路配置信息。
其中，侧行链路配置信息的实施方式可参见上述S101中所述，不再赘述。
从而，协作请求信息可是在该侧行链路配置信息所配置的协作发现资源上发送的。
一种可选的实施方式中，上述协作请求信息还可包括每个训练任务对应的预估开销等，从而有利于第二通信装置接收到该协作请求信息后，根据自身的剩余资源情况，确定是否可参与其中某些训练任务的训练。
S202.第二通信装置接收协作请求信息。
一种可选的实施方式中，网络设备为第二通信装置发送侧行链路配置信息，侧行链路配置信息配置有些用于接收请求信息的协作发现资源。因此，第二通信装置可在上述侧行链路配置信息所配置的协作发现资源上接收协作请求信息。
S203.第二通信装置确定参与第一训练任务时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
可理解的，第二通信装置根据自身的剩余资源量、自身的算力等，从协作请求信息中的多个训练任务中确定自身能参与的一个或多个训练任务。然后，第二通信装置通过第二指示信息，将自身能参与的第一训练任务告知给第一通信装置。第一训练任务包括第二通信装置能训练的一个或多个训练任务。
一种可选的实施方式中，第一通信装置和各第二通信装置预先约定：第二通信装置接收到协作请求信息后，若向第一通信装置反馈能参与的训练任务，反馈自身能参与训练的一个训练任务。也就是说，当第二通信装置可参与一个或多个训练任务的训练时，通过第二指示信息，向第一通信装置反馈上述一个或多个训练任务中的一个，即第一训练任务包括协作请求信息携带的多个训练任务的其中一个训练任务。
另一种可选的实施方式中，第一通信装置和各第二通信装置预先约定：第二通信装置接收到协作请求信息后，若向第一通信装置反馈能参与的训练任务，可反馈自身所有能参与训练的训练任务。也就是说，第二通信转置可通过第二指示信息向第一通信装置反馈自身所有能参与训练的训练任务，即第一训练任务包括协作请求信息中多个训练任务携带的一个或多个训练任务。
S204.第一通信装置接收来自第二通信装置的第二指示信息。
可理解的，第一通信装置通过接收来自第二通信装置的第二指示信息，获知到愿意参与训练的第二通信装置，以及每个第二通信装置可参与的训练任务。
由上述S203可知，第二通信装置向第一通信装置反馈可参与的一个训练任务，或者反馈所有能参与的训练任务。
当每个第二通信装置均反馈可参与的一个训练任务时，部分第二通信装置反馈的训练任务可能是同一训练任务。第一通信装置需从反馈相同训练任务的多个第二通信装置中，确定出训练该相同训练任务的第二通信装置。可理解的，第一通信装置通过协商的方式，或者第一通信装置根据多个第二通信装置的剩余资源，或算力大小，或每个第二通信装置与第一通信装置的信道质量等，确定出参与相同训练任务的一个第二通信装置，并向确定的第二通信装置发送第三指示信息，以告知该第二通信装置可进行第三指示信息指示的训练任务的训练。
其中，第一通信装置通过协商的方式可指第一通信装置与多个第二通信装置协商，确定距离第一通信装置最近的第二通信装置参与该相同训练任务的训练。本申请实施例并不限定具体的协商方式。
示例性的，第一通信装置根据多个第二通信装置中每个第二通信装置的算力大小，确定参与相同训练任务训练的第二通信装置，可指第一通信装置将多个第二通信装置中算力大小最大的第二通信装置参与该相同训练任务的训练，以保障该相同训练任务被完整训练。
示例性的，第一通信装置发送的协作请求信息中携带训练任务A、训练任务B，周边的第二通信装置A、第二通信装置B、第二通信装置C均接收到了该协作请求信息。第二通信装置A发送的第二指示信息指示了训练任务A，第二通信装置B和第二通信装置C发送的第二指示信息均指示了训练任务B。可见，第二通信装置B和第二通信装置C反馈了相同的训练任务，此时第一通信装置需从第二通信装置B和第二通信装置C中确定一个进行训练任务B训练的第二通信装置。若第一通信装置通过与第二通信装置B、第二通信装置C进行协商的方式，确定出让第二通信装置B参与训练任务B的训练，则第一通信装置向第二通信装置B发送第三指示信息，该第三指示信息用于指示训练任务B。从而第二通信装置B接收到该第三指示信息后，可进行训练任务B的训练。
示例性的，第一通信装置发送的协作请求信息中携带训练任务A、训练任务B，周边的第二通信装置A、第二通信装置B、第二通信装置C均接收到了该协作请求信息。第二通信装置A发送的第二指示信息指示了训练任务A和训练任务C，第二通信装置B和第二通信装置C发送的第二指示信息均指示了训练任务B。那么第一通信装置确定第二通信装置A参与训练任务A和训练任务C的训练，以及通过协商的方式确定第二通信装置C参与训练任务B的训练。从而第一通信装置向第二通信装置A发送的第三指示信息用于指示训练任务A和训练任务C，向第二通信装置C发送的第三指示信息用于指示训练任务B。
可选的，当每个第二通信装置均反馈可参与的一个训练任务，且每个第二通信装置反馈训练任务不相同时，第一通信装置也向每个反馈了第二指示信息的第二通信装置发送第三指 示信息，该第三指示信息用于指示该第二通信装置反馈的训练任务，以告知每个反馈了第二指示信息的第二通信装置，可进行训练任务的训练。
另一种可选的实施方式中，第二通信装置向第一通信装置反馈自身能参与训练的所有训练任务，那么第一通信装置根据每个第二通信装置可参与训练的训练任务，确定每个通信装置需参与的训练任务，并以第三指示信息的方式告知每个第二通信装置需参与的训练任务。也就是说，第一通信装置向每个第二通信装置发送第三指示信息，第三指示信息用于指示第一训练任务中的其中一个训练任务。
示例性的，第一通信装置发送的协作请求信息中携带训练任务A、训练任务B、训练任务C，周边的第二通信装置A、第二通信装置B、第二通信装置C均接收到了该协作请求信息。第二通信装置A发送的第二指示信息中指示了训练任务C，第二通信装置B发送的第二指示信息中指示了训练任务B，第二通信装置C发送的第二指示信息中指示了训练任务A和训练任务B。从而第一通信装置确定第二通信装置可参与训练任务A的训练。进而，第一通信装置向第二通信装置A发送的第三指示信息用于指示训练任务C，向第二通信装置B发送的第三指示信息用于指示训练任务B，向第二通信装置C发送的第三指示信息用于训练任务A。
示例性的，第一通信装置发送的协作请求信息中携带训练任务A、训练任务B、训练任务C，周边的第二通信装置A、第二通信装置B、第二通信装置C均接收到了该协作请求信息。第二通信装置A发送的第二指示信息中指示了训练任务A和训练任务C，第二通信装置B发送的第二指示信息中指示了训练任务B，第二通信装置C发送的第二指示信息中指示了训练任务A和训练任务B。第二通信装置根据第二通信装置A、第二通信装置B、第二通信装置C分别可参与的训练任务，确定第二通信装置A可参与训练任务C的训练，第二通信装置B可参与训练任务B的训练，第二通信装置C可参与训练任务A的训练。从而第一通信装置向第二通信装置A发送的第三指示信息用于指示训练任务C，向第二通信装置B发送的第三指示信息用于指示训练任务B，向第二通信装置C发送的第三指示信息用于指示训练任务A。
可见，各愿意参与训练任务的第二通信装置通过第三指示信息获知到自身需训练的训练任务。从而该部分的第二通信装置根据本地的训练数据，训练对应的神经网络模型，直至第一通信装置确定神经网络模型收敛时，停止对神经网络模型的训练。可理解的，第一通信装置通过待训练的神经网络模型的输入和输出，确定神经网络模型是否收敛。
示例性的，第一通信装置将待训练的神经网络模型拆分为上述图8所示的三个训练任务，且第一通信装置进行训练任务C的训练，即第一通信装置进行包含有神经网络输出的子网络的训练。从而，第一通信装置与各第二通信装置协商神经网络的输入(X)，并通过自身训练的训练任务C的输出(神经网络的输出)Y，确定自身和各第二通信装置训练的神经网络模型是否收敛。
一种可选的实施方式中，第一通信装置确定出反馈了第二指示信息的第二通信装置中每个第二通信装置参与的训练任务后，向每个第二通信装置发送第四指示信息。该第四指示信息用于指示第二通信装置需接收的第一输出、第一输出对应的时频位置，和/或需发送的第二输出、第二输出对应的时频资源位置。
其中，第一输出是第一通信装置训练的神经网络模型的输出，或者是除该第二通信装置外的其他第二通信装置训练的神经网络模型的输出；第二输出是该第二通信装置训练的神经网络模型的输出。
可理解的，第一通信装置通过第四指示信息为每个参与训练任务的第二通信装置指示需接收和/或需接收的参数，以及为每个第二通信装置调度需接收和/或需发送的参数的时频资源，以使得参与训练的第二通信装置获知自身需在哪些时频资源上接收和/或发送哪些参数。
示例性的，第一通信装置将待训练的神经网络模型拆分为上述如图8所示的训练任务，且第一通信装置进行训练任务A的训练，第二通信装置A进行训练任务B的训练，第二通信装置B进行训练任务C的训练。那么对于第二通信装置A而言，第一输出是第一通信装置训练的神经网络模型的输出(输出A)，第二输出是第二通信装置A训练的训练任务B中神经网络模型的输出(输出B)。对于第二通信装置B而言，第一输出是第二通信装置A训练的训练任务B中神经网络模型的输出(输出B)，第二输出是第二通信装置B训练的训练任务C中神经网络模型的输出(输出C)。
第一通信装置确定自身需在时频资源#a上发送训练任务A中的神经网络模型的输出。另外，第一通信装置通过第四指示信息向第二通信装置A指示第二通信装置需在时频资源#b中的部分时频资源上接收输出A，并在时频资源#b中的另一部分时频资源上发送输出B，或者，第一通信装置通过第四指示信息向第二通信装置A指示第二通信装置需在时频资源#b对应的不同频域资源上接收输出A和发送输出B；以及第一通信装置通过第四指示信息向第二通信装置B指示第二通信装置B需在时频资源#c中的部分时频资源上接收输出B，以及在时频资源#c中的另一部分时频资源上发送输出C，或者，第一通信装置通过第四指示信息向第二通信装置B指示第二通信装置B需在时频资源#c对应的不同频域资源上接收输出B和发送输出C。可见，第二通信装置A和第二通信装置B按照时分的方式，传输自身的训练结果。
一种可选的实施方式中，发送第四指示信息的资源可是上述侧行配置信息所配置的协作控制资源。
另一种可选的实施方式中，第一通信装置发送协作请求信息之前，在协作同步资源上发送同步信号，从而第二通信装置在协作同步资源上接收同步信号，并根据该同步信号与第一通信装置进行同步，以使得后续可与第一通信装置进行通信。其中，协作同步资源是在上述侧行链路配置信息中配置的。
请参见图9，图9是本申请实施例以第一通信装置为终端设备A、第二通信装置包括终端设备B和终端设备C为例的另一种模型训练方法的交互流程示意图。如图9所示：
901.若终端设备确定自身无法独自完成待训练的神经网络模型，则终端设备A向周边终端设备(终端设备B和终端设备C)发送同步信号。
902.终端设备B和终端设备C监听到同步信号后，分别根据该同步信号与终端设备A进行同步，且监听终端设备A的协作请求信息。终端设备B、终端设备C与终端设备A进行同步是为了保证后续可与终端设备A进行通信。终端设备A发送同步信号的资源，以及终端设备B和终端设备C监听同步信号的资源均可以是上述侧行链路配置信息配置的协作同步资源，不再赘述。
903.终端设备A在上述协作发现资源上广播协作请求信息，以请求周边终端设备协助多个训练任务的训练。
904.终端设备B和终端设备C在协作发现资源上监听到协作请求信息后，根据自身的剩余资源情况，从协作请求信息携带的多个训练任务中确定自身可参与的一个或多个训练任务。
905.终端设备B和终端设备C向终端设备A发送第二指示信息，以通过第二指示信息向终端设备A反馈自身可参与的一个或多个训练任务。
906.终端设备A再根据终端设备B和终端设备C可参与的训练任务，确定终端设备B和终端设备C分别需参与的训练任务，并发送第三指示信息，以第三指示信息的方式告知终端设备B和终端设备C分别需参与的训练任务。另外，终端设备A向终端设备B和终端设备C发送第四指示信息，以告知终端设备B和终端设备C各自需接收和/或发送的参数，以及每个参数对应的时频资源位置。
907.终端设备B和终端设备C对第三指示信息指示的训练任务进行训练，并根据每个参数的时频资源位置，发送或接收对应的参数。
908.终端设备A确定神经网络模型收敛时，确定神经网络模型训练完成。
本申请实施例中，第一通信装置将待训练的神经网络模型拆分为多个简单的神经网络模型，即多个训练任务，并通过广播协作请求信息的方式，将多个训练任务广播给周边的第二通信装置。第二通信装置接收到该多个训练任务后，根据自身的资源剩余量，确定可参与的训练任务。第二通信装置通过第二指示信息的方式将自身可参与的训练任务指示给第一通信装置。从而第一通信装置根据各第二通信装置可参与的训练任务，确定每个第二通信装置的训练任务，并以第三指示信息告知给各第二通信装置。进而参与协作训练的各第二通信装置训练第三指示信息指示的训练任务。一个或多个第二通信装置通过训练协作请求信息携带的多个训练任务中的一个或多个训练任务，协作第一通信装置完成待训练的神经网络模型的训练，从而可降低对第一通信装置的能力需求。
本申请实施例中，待训练的神经网络模型的训练完成，没有网络设备的参与，只有各终端设备的协作参与。另外，第一通信装置根据周边各第二通信装置的本地数据，完成待训练的神经网络模型的训练，使得训练的神经网络模型更加准确。
为了实现上述本申请实施例提供的方法中的各功能，第一通信装置或第二通信装置可以包括硬件结构和/或软件模块，以硬件结构、软件模块、或硬件结构加软件模块的形式来实现上述各功能。上述各功能中的某个功能以硬件结构、软件模块、还是硬件结构加软件模块的方式来执行，取决于技术方案的特定应用和设计约束条件。
如图10所示，本申请实施例提供了一种通信装置1000。该通信装置1000可以是第一通信装置的部件(例如，集成电路，芯片等等)，也可以是第二通信装置的部件(例如，集成电路，芯片等等)。该通信装置1000也可以是其他通信单元，用于实现本申请方法实施例中的方法。该通信装置1000可以包括：通信单元1001和处理单元1002。可选的，还可以包括存储单元1003。
在一种可能的设计中，如图10中的一个或者多个单元可能由一个或者多个处理器来实现，或者由一个或者多个处理器和存储器来实现；或者由一个或多个处理器和收发器实现；或者由一个或者多个处理器、存储器和收发器实现，本申请实施例对此不作限定。所述处理器、存储器、收发器可以单独设置，也可以集成。
所述通信装置1000具备实现本申请实施例描述的第二通信装置的功能，可选的，通信装置1000具备实现本申请实施例描述的第一通信装置的功能。比如，所述通信装置1000包括第二通信装置执行本申请实施例描述的第二通信装置涉及步骤所对应的模块或单元或手段(means)，所述功能或单元或手段(means)可以通过软件实现，或者通过硬件实现，也可以通过硬件执行相应的软件实现，还可以通过软件和硬件结合的方式实现。详细可进一步参考前述对应方法实施例中的相应描述。
在一种可能的设计中，一种通信装置1000可包括：处理单元1002和通信单元1001，处理单元1002用于控制通信单元1001进行数据/信令收发；
通信单元1001，用于接收第一通信装置的第一神经网络参数；
通信单元1001，还用于在所述第一神经网络参数与所述装置的第二神经网络参数之间的相关系数小于第一阈值时，向所述第一通信装置发送第一指示信息；
所述第一指示信息用于指示所述装置参与所述第一通信装置的第一神经网络模型的训练。
一种可选的实现方式中，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
一种可选的实现方式中，所述第一神经网络参数是在协作发现资源上接收的；所述协作发现资源是在侧行链路配置信息中配置的。
一种可选的实现方式中，在所述相关系数小于所述第一阈值时，所述通信单元1001还用于：向所述第一通信装置发送所述第二神经网络参数。
另一种可选的实现方式中，所述通信单元1001还用于：接收来自所述第一通信装置的控制信号；所述控制信号用于指示时频资源；指示的时频资源用于所述装置发送所述第二神经网络参数。
一种可选的实现方式中，接收所述控制信号的资源是协作控制资源；所述协作控制资源是在所述侧行链路配置信息中配置的。
一种可选的实现方式中，所述通信单元1001，还用于在协作同步资源上接收同步信号；所述处理单元1002，用于根据所述同步信号，与所述第一通信装置进行同步；所述协作同步资源是在所述侧行链路配置信息中配置的。
一种可选的实现方式中，所述侧行链路配置信息所配置的所述协作发现资源、所述协作控制资源、所述协作同步资源中的一种或多种资源是预先配置的，或是动态指示的，或是非授权频谱资源。
一种可选的实现方式中，所述第一神经网络参数是所述第一神经网络的模型参数，所述第二神经网络参数是所述第二神经网络的模型参数；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据第一参数和第二参数确定的；所述第一参数是所述第二通信装置对所述第一神经网络模型输入训练数据时，所述第一神经网络模型输出的参数；所述第一神经网络模型是根据所述第一神经网络的模型参数确定的；所述第二参数是所述第二通信装置对所述第二通信装置的第二神经网络模型输入所述训练数据时，所述第二神经网络模型输出的参数。
另一种可选的实现方式中，所述第一神经网络参数是所述第一神经网络的梯度，所述第二神经网络参数是所述第二神经网络的梯度；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据所述第一神经网络参数的概率密度分布和所述第二神经网络参数的概率密度分布确定的。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
在另一种可能的设计中，一种通信装置1000可包括：处理单元1002和通信单元1001，处理单元1002用于控制通信单元1001进行数据/信令收发；
通信单元1001，用于发送所述装置的第一神经网络参数；
通信单元1001，还用于接收来自第二通信装置的第一指示信息；
所述第一指示信息是所述第二通信装置在所述第一神经网络参数与所述第二装置的第二神经网络参数之间的相关系数小于第一阈值时发送的；
所述第一指示信息用于指示所述第二通信装置参与所述装置的第一神经网络模型的训练。
一种可选的实现方式中，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
一种可选的实现方式中，所述第一神经网络参数是在协作发现资源上发送的；所述协作发现资源是在侧行链路配置信息中配置的。
又一种可选的实现方式中，所述通信单元1001，还用于接收来自所述第二通信装置的第二神经网络参数；所述处理单元1002，用于根据所述第二神经网络参数，更新所述第一神经网络模型。
又一种可选的实施方式中，所述通信单元1001，还用于向所述第二通信装置发送控制信号；所述控制信号用于指示时频资源；指示的时频资源用于所述第二通信装置发送所述第二神经网络参数。
一种可选的实现方式中，发送所述控制信号的资源是协作控制资源；所述协作控制资源是在所述侧行链路配置信息中配置的。
一种可选的实施方式中，所述通信单元1001，还用于在协作同步资源上发送同步信号；所述协作同步资源是在所述侧行链路配置信息中配置的。
一种可选的实施方式中，所述侧行链路配置信息所配置的所述协作发现资源、所述协作控制资源、所述协作同步资源中的一种或多种资源是预先配置的，或是动态指示的，或是非授权频谱资源。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
在又一种可能的设计中，一种通信装置1000可包括：处理单元1002和通信单元1001，处理单元1002用于控制通信单元1001进行数据/信令收发；
通信单元1001，用于发送协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
通信单元1001，还用于接收来自第二通信装置的第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
一种可选的实施方式中，上述协作请求信息在协作发现资源上发送的，协作发现资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，第二通信装置发送的第二指示信息指示的第一训练任务包括多个训练任务，此时第一通信装置还可发送第三指示信息，第三指示信息用于指示第一训练任务中的其中一个训练任务。
另一种可选的实施方式中，多个第二通信装置发送的第二指示信息指示的第一训练任务是多个训练任务中的相同训练任务，此时第一通信装置也可通过第三指示信息，向其中的一个第二通信装置指示参与训练的训练任务。
一种可选的实施方式中，第一通信装置还可向第二通信装置发送第四指示信息，第四指示信息用于指示第二通信装置需接收的第一输出、第一输出对应的时频资源位置，和/或需发送的第二输出、第二输出对应的时频资源位置。第一输出是第一通信装置训练的神经网络模 型的输出，或者是除第二通信装置外的其他第二通信装置训练的神经网络模型的输出；第二输出是第二通信装置训练的神经网络模型的输出。
一种可选的实施方式中，发送第四指示信息的资源是协作控制资源，协作控制资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，第一通信装置还可在协作同步资源上发送同步信号，以使得第二通信装置根据该同步信号与第一通信装置进行同步。另外，该协作同步资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
在又一种可能的设计中，一种通信装置1000可包括：处理单元1002和通信单元1001，处理单元1002用于控制通信单元1001进行数据/信令收发；
通信单元1001，用于接收协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
通信单元1001，还用于确定参与第一训练任务的训练时，发送第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
一种可选的实施方式中，协作请求信息在协作发现资源上接收的，协作发现资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，第二通信装置还可接收第三指示信息，第三指示信息用于指示第一训练任务中的其中一个训练任务。
一种可选的实施方式中，第二通信装置还可接收第四指示信息，第四指示信息用于指示第二通信装置接收的第一输出、第一输出对应的时频资源位置，和/或发送的第二输出、第二输出对应的时频资源位置。第一输出是第一通信装置训练的神经网络模型的输出，或者是除第二通信装置外的其他第二通信装置训练的神经网络模型的输出；第二输出是第二通信装置训练的神经网络模型的输出。
一种可选的实施方式中，接收第四指示信息的资源是协作控制资源，协作控制资源是在侧行链路配置信息中配置的。
一种可选的实施方式中，第二通信装置还可在协作同步资源上接收同步信号，并根据同步信号，与第一通信装置进行同步。
一种可选的实施方式中，上述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
本申请实施例还提供一种通信装置1100，图11为通信装置1100的结构示意图。所述通信装置1100可以是第一通信装置或第二通信装置，也可以是支持第一通信装置实现上述方法的芯片、芯片系统、或处理器等，还可以是支持第二通信装置实现上述方法的芯片、芯片系统、或处理器等。该装置可用于实现上述方法实施例中描述的方法，具体可以参见上述方法实施例中的说明。
所述通信装置1100可以包括一个或多个处理器1101。所述处理器1101可以是通用处理器或者专用处理器等。例如可以是基带处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件或中央处理器(Central Processing Unit，CPU)。基带处理器可以用于对通信协议以及通信数据进行处理，中央处理器可以用于对通信装置(如，基站、基带芯片，终端、终端芯片，分布单元(distributed unit，DU)或集中单元(centralized unit，CU)等)进行控制，执行软件程序，处理软件程序的数据。
可选的，所述通信装置1100中可以包括一个或多个存储器1102，其上可以存有指令1104，所述指令可在所述处理器1101上被运行，使得所述通信装置1100执行上述方法实施例中描述的方法。可选的，所述存储器1102中还可以存储有数据。所述处理器1101和存储器1102可以单独设置，也可以集成在一起。
存储器1102可包括但不限于硬盘(hard disk drive，HDD)或固态硬盘(solid-state drive，SSD)等非易失性存储器，随机存储记忆体(Random Access Memory，RAM)、可擦除可编程只读存储器(Erasable Programmable ROM，EPROM)、ROM或便携式只读存储器(Compact Disc Read-Only Memory，CD-ROM)等等。
可选的，所述通信装置1100还可以包括收发器1105、天线1106。所述收发器1105可以称为收发单元、收发机、或收发电路等，用于实现收发功能。收发器1105可以包括接收器和发送器，接收器可以称为接收机或接收电路等，用于实现接收功能；发送器可以称为发送机或发送电路等，用于实现发送功能。
所述通信装置1100为第二通信装置：收发器1105用于执行上述模型训练方法100中的S102、S103，以及用于执行模型训练方法200中的S202、S203。
所述通信装置1100为第二通信装置：收发器1105用于模型训练方法100中的S101、S104，以及用于执行模型训练方法200中的S201、S204。
另一种可能的设计中，处理器1101中可以包括用于实现接收和发送功能的收发器。例如该收发器可以是收发电路，或者是接口，或者是接口电路。用于实现接收和发送功能的收发电路、接口或接口电路可以是分开的，也可以集成在一起。上述收发电路、接口或接口电路可以用于代码/数据的读写，或者，上述收发电路、接口或接口电路可以用于信号的传输或传递。
又一种可能的设计中，可选的，处理器1101可以存有指令1103，指令1103在处理器1101上运行，可使得所述通信装置1100执行上述方法实施例中描述的方法。指令1103可能固化在处理器1101中，该种情况下，处理器1101可能由硬件实现。
又一种可能的设计中，通信装置1100可以包括电路，所述电路可以实现前述方法实施例中发送或接收或者通信的功能。本申请实施例中描述的处理器和收发器可实现在集成电路(integrated circuit，IC)、模拟IC、射频集成电路(radio frequency integrated circuit，RFIC)、混合信号IC、专用集成电路(application specific integrated circuit，ASIC)、印刷电路板(printed circuit board，PCB)、电子设备等上。该处理器和收发器也可以用各种IC工艺技术来制造，例如互补金属氧化物半导体(complementary metal oxide semiconductor，CMOS)、N型金属氧化物半导体(nMetal-oxide-semiconductor，NMOS)、P型金属氧化物半导体(positive channel metal oxide semiconductor，PMOS)、双极结型晶体管(Bipolar Junction Transistor，BJT)、双极CMOS(BiCMOS)、硅锗(SiGe)、砷化镓(GaAs)等。
以上实施例描述中的通信装置可以是第一通信装置或第二通信装置，但本申请实施例中描述的通信装置的范围并不限于此，而且通信装置的结构可以不受图11的限制。通信装置可以是独立的设备或者可以是较大设备的一部分。例如所述通信装置可以是：
(1)独立的集成电路IC，或芯片，或，芯片系统或子系统；
(2)具有一个或多个IC的集合，可选的，该IC集合也可以包括用于存储数据，指令的存储部件；
(3)ASIC，例如调制解调器(modulator)；
(4)可嵌入在其他设备内的模块；
(5)接收机、终端、智能终端、蜂窝电话、无线设备、手持机、移动单元、车载设备、网络设备、云设备、人工智能设备等等；
(6)其他等等。
对于通信装置可以是芯片或芯片系统的情况，可参见图12所示的芯片的结构示意图。图12所示的芯片1200包括处理器1201和接口1202。其中，处理器1201的数量可以是一个或多个，接口1202的数量可以是多个。该处理器1201可以是逻辑电路，该接口1202可以是输入输出接口、输入接口或输出接口。所述芯片1200还可包括存储器1203。
一种设计中，对于芯片用于实现本申请实施例中第二通信装置的功能的情况：处理器1201，用于控制接口1202进行输出或接收。
所述接口1202，用于接收第一通信装置的第一神经网络参数；
所述接口1202，还用于在所述第一神经网络参数与所述装置的第二神经网络参数之间的相关系数小于第一阈值时，输出第一指示信息；所述第一指示信息用于指示所述装置参与所述第一通信装置的第一神经网络模型的训练。
另一种设计中，对于芯片用于实现本申请实施例中第一通信装置的功能的情况：
所述接口1202，用于输出所述装置的第一神经网络参数；
所述接口1202，还用于接收来自第二通信装置的第一指示信息；
所述第一指示信息是所述第二通信装置在所述第一神经网络参数与所述第二装置的第二神经网络参数之间的相关系数小于第一阈值时输出的；所述第一指示信息用于指示所述第二通信装置参与所述装置的第一神经网络模型的训练。
又一种设计中，对于芯片用于实现本申请实施例中第一通信装置的功能的情况：
所述接口1202，用于输出协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
所述接口1202，还用于接收来自第二通信装置的第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
又一种设计中，对于芯片用于实现本申请实施例中第二通信装置的功能的情况：
所述接口1202，用于接收协作请求信息，协作请求信息包括多个训练任务，多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
所述接口1202，还用于确定参与第一训练任务的训练时，输出第二指示信息，第二指示信息用于指示第二通信装置参与第一训练任务的训练，第一训练任务是多个训练任务中的其中一个或多个。
本申请实施例中通信装置1100、芯片1200还可执行上述通信装置1000所述的实现方式。本领域技术人员还可以了解到本申请实施例列出的各种说明性逻辑块(illustrative logical block)和步骤(step)可以通过电子硬件、电脑软件，或两者的结合进行实现。这样的功能 是通过硬件还是软件来实现取决于特定的应用和整个系统的设计要求。本领域技术人员可以对于每种特定的应用，可以使用各种方法实现所述的功能，但这种实现不应被理解为超出本申请实施例保护的范围。
本申请实施例和上述模型训练方法100和模型训练方法200所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述模型训练方法100和模型训练方法200所示实施例的描述，不再赘述。
本申请还提供了一种计算机可读存储介质，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序产品，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序，当其在计算机上运行时，实现上述任一方法实施例的功能。
本申请还提供了一种通信系统，该系统包括上述方面的至少一个第一通信装置、至少两个第二通信装置。在另一种可能的设计中，该系统还可以包括本申请提供的方案中与第一通信装置、第二通信装置进行交互的其他设备。
上述实施例中，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指令。在计算机上加载和执行所述计算机指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机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(例如同轴电缆、光纤、数字用户线(digital subscriber line，DSL))或无线(例如红外、无线、微波等)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包含一个或多个可用介质集成的服务器、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)、光介质(例如，高密度数字视频光盘(digital video disc，DVD))、或者半导体介质(例如，SSD)等。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (67)

  1. 一种模型训练方法，其特征在于，所述方法包括：
    第二通信装置接收第一通信装置的第一神经网络参数；
    所述第二通信装置在所述第一神经网络参数与所述第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时，向所述第一通信装置发送第一指示信息；
    所述第一指示信息用于指示所述第二通信装置参与所述第一通信装置的第一神经网络模型的训练。
  2. 根据权利要求1所述的方法，其特征在于，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
  3. 根据权利要求1或2所述的方法，其特征在于，所述第一神经网络参数是在协作发现资源上接收的，所述协作发现资源是在侧行链路配置信息中配置的。
  4. 根据权利要求1至3任一项所述的方法，其特征在于，在所述相关系数小于所述第一阈值时，所述方法还包括：
    所述第二通信装置向所述第一通信装置发送所述第二神经网络参数。
  5. 根据权利要求1至4任一项所述的方法，其特征在于，所述方法还包括：
    所述第二通信装置接收来自所述第一通信装置的控制信号；
    所述控制信号用于指示时频资源；指示的时频资源用于所述第二通信装置发送所述第二神经网络参数。
  6. 根据权利要求5所述的方法，其特征在于，所述控制信号是在协作控制资源上接收的，所述协作控制资源是侧行链路配置信息中配置的。
  7. 根据权利要求1至6任一项所述的方法，其特征在于，所述方法还包括：
    所述第二通信装置在协作同步资源上接收同步信号；
    所述第二通信装置根据所述同步信号，与所述第一通信装置进行同步；
    所述协作同步资源是在侧行链路配置信息中配置的。
  8. 根据权利要求3，6，7任一项所述的方法，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  9. 根据权利要求1至8任一项所述的方法，其特征在于，所述第一神经网络参数是所述第一神经网络的模型参数，所述第二神经网络参数是所述第二神经网络的模型参数；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据第一参数和第二参数确定的；
    所述第一参数是所述第二通信装置对所述第一神经网络模型输入训练数据时，所述第一神经网络模型输出的参数；所述第一神经网络模型是根据所述第一神经网络的模型参数确定的；所述第二参数是所述第二通信装置对所述第二通信装置的第二神经网络模型输入所述训练数据时，所述第二神经网络模型输出的参数。
  10. 根据权利要求1至8任一项所述的方法，其特征在于，所述第一神经网络参数是所述第一神经网络的梯度，所述第二神经网络参数是所述第二神经网络的梯度；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据所述第一神经网络参数的概率密度分布和所述第二神经网络参数的概率密度分布确定的。
  11. 一种模型训练方法，其特征在于，所述方法包括：
    第一通信装置发送所述第一通信装置的第一神经网络参数；
    所述第一通信装置接收来自第二通信装置的第一指示信息；
    所述第一指示信息是所述第二通信装置在所述第一神经网络参数与所述第二通信装置的第二神经网络参数之间的相关系数小于第一阈值时发送的；
    所述第一指示信息用于指示所述第二通信装置参与所述第一通信装置的第一神经网络模型的训练。
  12. 根据权利要求11所述的方法，其特征在于，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
  13. 根据权利要求11或12所述的方法，其特征在于，所述第一神经网络参数是在协作发现资源上发送的，所述协作发现资源是在侧行链路配置信息中配置的。
  14. 根据权利要求11至13任一项所述的方法，其特征在于，所述方法还包括：
    所述第一通信装置接收来自所述第二通信装置的所述第二神经网络参数；
    所述第一通信装置根据所述第二神经网络参数，更新所述第一神经网络模型。
  15. 根据权利要求11至14任一项所述的方法，其特征在于，所述方法还包括：
    所述第一通信装置向所述第二通信装置发送控制信号；
    所述控制信号用于指示时频资源；指示的时频资源用于所述第二通信装置发送所述第二神经网络参数。
  16. 根据权利要求15所述的方法，其特征在于，所述控制信号是在协作控制资源上发送的，所述协作控制资源是在侧行链路配置信息中配置的。
  17. 根据权利要求11至15任一项所述的方法，其特征在于，所述方法还包括：
    所述第一通信装置在协作同步资源上发送同步信号；
    所述协作同步资源是在侧行链路配置信息中配置的。
  18. 根据权利要求13，16，17任一项所述的方法，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  19. 一种模型训练方法，其特征在于，所述方法包括：
    第一通信装置发送协作请求信息，所述协作请求信息包括多个训练任务，所述多个训练任务是所述第一通信装置将待训练的神经网络模型进行拆分获得的；
    所述第一通信装置接收来自第二通信装置的第二指示信息，所述第二指示信息用于指示所述第二通信装置参与第一训练任务的训练，所述第一训练任务是所述多个训练任务中的其中一个或多个。
  20. 根据权利要求19所述的方法，其特征在于，所述协作请求信息在协作发现资源上发送的，所述协作发现资源是在侧行链路配置信息中配置的。
  21. 根据权利要求19或20所述的方法，其特征在于，所述第一训练任务是所述多个训练任务中的其中多个；所述方法还包括：
    所述第一通信装置发送第三指示信息，所述第三指示信息用于指示第一训练任务中的其中一个训练任务。
  22. 根据权利要求19至21任一项所述的方法，其特征在于，所述方法还包括：
    所述第一通信装置向第二通信装置发送第四指示信息，所述第四指示信息用于指示所述第二通信装置需接收的第一输出、所述第一输出对应的时频资源位置，和/或需发送的第二输出、所述第二输出对应的时频资源位置；
    第一输出是所述第一通信装置训练的神经网络模型的输出，或者是除所述第二通信装置外的其他第二通信装置训练的神经网络模型的输出；所述第二输出是所述第二通信装置训练的神经网络模型的输出。
  23. 根据权利要求22所述的方法，其特征在于，所述第四指示信息是在协作控制资源上发送的，所述协作控制资源是在侧行链路配置信息中配置的。
  24. 根据权利要求19至23任一项所述的方法，其特征在于，所述方法还包括：
    所述第一通信装置在协作同步资源上发送同步信号，以使所述第二通信装置根据所述同步信号与所述第一通信装置进行同步；
    所述协作同步资源是在侧行链路配置信息中配置的。
  25. 根据权利要求20，23，24任一项所述的方法，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  26. 一种模型训练方法，其特征在于，所述方法包括：
    第二通信装置接收协作请求信息，所述协作请求信息包括多个训练任务，所述多个训练 任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
    所述第二通信装置确定参与第一训练任务的训练时，发送第二指示信息；所述第二指示信息用于指示所述第二通信装置参与第一训练任务的训练，所述第一训练任务是所述多个训练任务中的其中一个或多个。
  27. 根据权利要求26所述的方法，其特征在于，所述协作请求信息在协作发现资源上接收的，所述协作发现资源是在侧行链路配置信息中配置的。
  28. 根据权利要求26或27所述的方法，其特征在于，所述方法还包括：
    所述第二通信装置接收第三指示信息；所述第三指示信息用于指示所述第一训练任务中的其中一个训练任务。
  29. 根据权利要求26至28任一项所述的方法，其特征在于，所述方法还包括：
    所述第二通信装置接收第四指示信息，所述第四指示信息用于指示所述第二通信装置接收的第一输出、所述第一输出对应的时频资源位置，和/或发送的第二输出、所述第二输出对应的时频资源位置；
    所述第一输出是所述第一通信装置训练的神经网络模型的输出，或者是除所述第二通信装置外的其他第二通信装置训练的神经网络模型的输出；所述第二输出是所述第二通信装置训练的神经网络模型的输出。
  30. 根据权利要求29所述的方法，其特征在于，所述第四指示信息是协作控制资源上接收的，所述协作控制资源是在侧行链路配置信息中配置的。
  31. 根据权利要求26至30任一项所述的方法，其特征在于，所述方法还包括：
    所述第二通信装置在协作同步资源上接收同步信号；
    所述第二通信装置根据所述同步信号，与所述第一通信装置进行同步。
  32. 根据权利要求27，30，31任一项所述的方法，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  33. 一种通信装置，其特征在于，所述装置包括：
    通信单元，用于接收第一通信装置的第一神经网络参数；
    所述通信单元，还用于在所述第一神经网络参数与所述装置的第二神经网络参数之间的相关系数小于第一阈值时，向所述第一通信装置发送第一指示信息；
    所述第一指示信息用于指示所述装置参与所述第一通信装置的第一神经网络模型的训练。
  34. 根据权利要求33所述的装置，其特征在于，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
  35. 根据权利要求33或34所述的装置，其特征在于，所述第一神经网络参数是在协作发现资源上接收的，所述协作发现资源是在侧行链路配置信息中配置的。
  36. 根据权利要求33至35任一项所述的装置，其特征在于，在所述相关系数小于所述第一阈值时，所述通信单元还用于：
    向所述第一通信装置发送所述第二神经网络参数。
  37. 根据权利要求33至36任一项所述的装置，其特征在于，所述通信单元还用于：
    接收来自所述第一通信装置的控制信号；所述控制信号用于指示时频资源；指示的时频资源用于所述装置发送所述第二神经网络参数。
  38. 根据权利要求37所述的装置，其特征在于，所述控制信号是在协作控制资源上接收的，所述协作控制资源是侧行链路配置信息中配置的。
  39. 根据权利要求33至38任一项所述的装置，其特征在于，所述装置还包括处理单元；
    所述通信单元，还用于在协作同步资源上接收同步信号；
    所述处理单元，用于根据所述同步信号，与所述第一通信装置进行同步；
    所述协作同步资源是在侧行链路配置信息中配置的。
  40. 根据权利要求35，38，39所述的装置，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  41. 根据权利要求33至40任一项所述的装置，其特征在于，所述第一神经网络参数是第一神经网络的模型参数，所述第二神经网络参数是第二神经网络的模型参数；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据第一参数和第二参数确定的；
    所述第一参数是所述装置对所述第一神经网络模型输入训练数据时，所述第一神经网络模型输出的参数；所述第一神经网络模型是根据所述第一神经网络的模型参数确定的；所述第二参数是所述第二通信装置对所述装置的第二神经网络模型输入所述训练数据时，所述第二神经网络模型输出的参数。
  42. 根据权利要求33至40任一项所述的装置，其特征在于，所述第一神经网络参数是所述第一神经网络的梯度，所述第二神经网络参数是所述第二神经网络的梯度；所述第一神经网络参数与所述第二神经网络参数之间的相关系数是根据所述第一神经网络参数的概率密度分布和所述第二神经网络参数的概率密度分布确定的。
  43. 一种通信装置，其特征在于，所述装置包括：
    通信单元，用于发送所述装置的第一神经网络参数；
    所述通信单元，还用于接收来自第二通信装置的第一指示信息；
    所述第一指示信息是所述第二通信装置在所述第一神经网络参数与所述第二装置的第二神经网络参数之间的相关系数小于第一阈值时发送的；
    所述第一指示信息用于指示所述第二通信装置参与所述装置的第一神经网络模型的训练。
  44. 根据权利要求43所述的装置，其特征在于，所述第一神经网络参数是第一神经网络的模型参数或所述第一神经网络的梯度；所述第二神经网络参数是第二神经网络的模型参数或所述第二神经网络的梯度。
  45. 根据权利要求43或44所述的装置，其特征在于，所述第一神经网络参数是在协作发现资源上发送的，所述协作发现资源是在侧行链路配置信息中配置的。
  46. 根据权利要求43至45任一项所述的装置，其特征在于，所述装置还包括处理单元；
    所述通信单元，还用于接收来自所述第二通信装置的第二神经网络参数；
    所述处理单元，用于根据所述第二神经网络参数，更新所述第一神经网络模型。
  47. 根据权利要求43至46任一项所述的装置，其特征在于，
    所述通信单元，还用于向所述第二通信装置发送控制信号；
    所述控制信号用于指示时频资源；指示的时频资源用于所述第二通信装置发送所述第二神经网络参数。
  48. 根据权利要求47所述的装置，其特征在于，所述控制信号是在协作控制资源上发送的，所述协作控制资源是在侧行链路配置信息中配置的。
  49. 根据权利要求43至48任一项所述的装置，其特征在于，
    所述通信单元，还用于在协作同步资源上发送同步信号；所述协作同步资源是在侧行链路配置信息中配置的。
  50. 根据权利要求45，48，49任一项所述的装置，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  51. 一种通信装置，其特征在于，所述装置包括：
    通信单元，用于发送协作请求信息，所述协作请求信息包括多个训练任务，所述多个训练任务是所述第一通信装置将待训练的神经网络模型进行拆分获得的；
    所述通信单元，还用于接收来自第二通信装置的第二指示信息，所述第二指示信息用于指示所述第二通信装置参与第一训练任务的训练，所述第一训练任务是所述多个训练任务中的其中一个或多个。
  52. 根据权利要求51所述的装置，其特征在于，所述协作请求信息在协作发现资源上发送的，所述协作发现资源是在侧行链路配置信息中配置的。
  53. 根据权利要求51或52所述的装置，其特征在于，所述第一训练任务是所述多个训练 任务中的其中多个；
    所述通信单元，还用于发送第三指示信息，所述第三指示信息用于指示第一训练任务中的其中一个训练任务。
  54. 根据权利要求51至53任一项所述的装置，其特征在于，所述通信单元，还用于：
    向第二通信装置发送第四指示信息，所述第四指示信息用于指示所述第二通信装置需接收的第一输出、所述第一输出对应的时频资源位置，和/或需发送的第二输出、所述第二输出对应的时频资源位置；
    第一输出是所述装置训练的神经网络模型的输出，或者是除所述第二通信装置外的其他第二通信装置训练的神经网络模型的输出；所述第二输出是所述第二通信装置训练的神经网络模型的输出。
  55. 根据权利要求54所述的装置，其特征在于，所述第四指示信息是在协作控制资源上发送的，所述协作控制资源是在侧行链路配置信息中配置的。
  56. 根据权利要求51至55任一项所述的装置，其特征在于，所述通信单元，还用于：
    在协作同步资源上发送同步信号，以使所述第二通信装置根据所述同步信号与所述装置进行同步；
    所述协作同步资源是在侧行链路配置信息中配置的。
  57. 根据权利要求52，55，56任一项所述的装置，其特征在于，所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  58. 一种通信装置，其特征在于，所述装置包括：
    通信单元，用于接收协作请求信息，所述协作请求信息包括多个训练任务，所述多个训练任务是第一通信装置将待训练的神经网络模型进行拆分获得的；
    处理单元，用于确定参与第一训练任务的训练时，发送第二指示信息；所述第二指示信息用于指示所述第二通信装置参与第一训练任务的训练，所述第一训练任务是所述多个训练任务中的其中一个或多个。
  59. 根据权利要求58所述的装置，其特征在于，所述协作请求信息在协作发现资源上接收的，所述协作发现资源是在侧行链路配置信息中配置的。
  60. 根据权利要求58或59所述的装置，其特征在于，所述通信单元，还用于：
    接收第三指示信息；所述第三指示信息用于指示所述第一训练任务中的其中一个训练任务。
  61. 根据权利要求58至60任一项所述的装置，其特征在于，所述通信单元，还用于：
    接收第四指示信息，所述第四指示信息用于指示所述装置接收的第一输出、所述第一输出对应的时频资源位置，和/或发送的第二输出、所述第二输出对应的时频资源位置；
    所述第一输出是所述第一通信装置训练的神经网络模型的输出，或者是除所述装置外的其他装置训练的神经网络模型的输出；所述第二输出是所述装置训练的神经网络模型的输出。
  62. 根据权利要求61所述的装置，其特征在于，所述第四指示信息是协作控制资源上接收的，所述协作控制资源是在侧行链路配置信息中配置的。
  63. 根据权利要求58至62任一项所述的装置，其特征在于，
    所述通信单元，还用于在协作同步资源上接收同步信号；
    所述处理单元，还用于根据所述同步信号，与所述第一通信装置进行同步。
  64. 根据权利要求59，62，63任一项所述的装置，其特征在于，
    所述侧行链路配置信息所配置的协作发现资源、协作控制资源、协作同步资源是预先配置的，或是动态指示的，或是非授权频谱资源。
  65. 一种通信装置，其特征在于，包括处理器和收发器，所述收发器用于与其它通信装置进行通信；所述处理器用于运行程序，以使得所述通信装置实现权利要求1至10任一项所述的方法，或者，以使得所述通信装置实现权利要求11至18任一项所述的方法，或者，以使得所述通信装置实现权利要求19至25任一项所述的方法，或者，以使得所述通信装置实现权利要求26至32任一项所述的方法。
  66. 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用于存储有指令，当其在计算机上运行时，使得权利要求1至10任一项所述的方法被执行；或者权利要求11至18任一项所述的方法被执行，或者权利要求19至25任一项所述的方法被执行，或者权利要求26至32任一项所述的方法被执行。
  67. 一种包含指令的计算机程序产品，其特征在于，当其在计算机上运行时，使得权利要求1至10任一项所述的方法被执行；或者权利要求11至18任一项所述的方法被执行；或者权利要求19至25任一项所述的方法被执行；或者权利要求26至32任一项所述的方法被执行。
PCT/CN2022/133214 2021-11-22 2022-11-21 一种模型训练方法及相关装置 WO2023088465A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111386640.7A CN116151348A (zh) 2021-11-22 2021-11-22 一种模型训练方法及相关装置
CN202111386640.7 2021-11-22

Publications (1)

Publication Number Publication Date
WO2023088465A1 true WO2023088465A1 (zh) 2023-05-25

Family

ID=86352991

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/133214 WO2023088465A1 (zh) 2021-11-22 2022-11-21 一种模型训练方法及相关装置

Country Status (2)

Country Link
CN (1) CN116151348A (zh)
WO (1) WO2023088465A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116436802B (zh) * 2023-06-13 2023-08-18 山西凯瑞科技有限公司 基于机器学习的智能控制方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110490335A (zh) * 2019-08-07 2019-11-22 深圳前海微众银行股份有限公司 一种计算参与者贡献率的方法及装置
CN111340242A (zh) * 2020-05-15 2020-06-26 支付宝(杭州)信息技术有限公司 保护隐私的模型联合训练方法及装置
US20200285980A1 (en) * 2019-03-08 2020-09-10 NEC Laboratories Europe GmbH System for secure federated learning
CN113505520A (zh) * 2021-05-17 2021-10-15 京东科技控股股份有限公司 用于支持异构联邦学习的方法、装置和系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200285980A1 (en) * 2019-03-08 2020-09-10 NEC Laboratories Europe GmbH System for secure federated learning
CN110490335A (zh) * 2019-08-07 2019-11-22 深圳前海微众银行股份有限公司 一种计算参与者贡献率的方法及装置
CN111340242A (zh) * 2020-05-15 2020-06-26 支付宝(杭州)信息技术有限公司 保护隐私的模型联合训练方法及装置
CN113505520A (zh) * 2021-05-17 2021-10-15 京东科技控股股份有限公司 用于支持异构联邦学习的方法、装置和系统

Also Published As

Publication number Publication date
CN116151348A (zh) 2023-05-23

Similar Documents

Publication Publication Date Title
EP4266756A1 (en) Network resource selection method, and terminal device and network device
CN113179539A (zh) 蜂窝系统中的网络切片选择
WO2023088465A1 (zh) 一种模型训练方法及相关装置
CN114026902A (zh) 小区配置的确定方法及装置
CN116746085A (zh) 传输数据的方法和装置
WO2022056764A1 (en) Multicast service configuration
CN113453356A (zh) 通信优先级确定方法及装置
CN116848885A (zh) 一种通信方法及装置
CN116602010A (zh) 小数据传输控制
WO2021043403A1 (en) Data transmission
WO2024087116A1 (zh) 一种通信方法及装置
WO2023246267A1 (zh) 通信方法、通信装置和系统
WO2023015573A1 (zh) 通信方法、设备及存储介质
WO2023137660A1 (zh) 无线通信的方法、终端设备和网络设备
CN117580183B (zh) 网络优先级的协商方法和相关设备
WO2023001060A1 (zh) 一种通信方法及相关装置
WO2023070684A1 (zh) 无线通信的方法和设备
WO2024092791A1 (zh) 一种资源指示方法及通信装置
US20240196385A1 (en) Improving sidelink communication
WO2023056852A1 (zh) 一种通信方法、装置及系统
WO2022135288A1 (zh) 信息处理的方法和装置
WO2023236774A1 (zh) 意图管理的方法与装置
WO2023115279A1 (zh) 数据传输方法及装置
WO2023283942A1 (zh) 无线通信的方法、终端设备和网络设备
WO2024011581A1 (zh) 一种通信方法及装置

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: 22894994

Country of ref document: EP

Kind code of ref document: A1