WO2025140055A1 - Distributed training method, apparatus and system, and chip module and storage medium - Google Patents

Abstract

An artificial intelligence (AI) distributed training method, apparatus and system, and a chip module and a storage medium. A central node indicates a selection parameter, such that each sub-node participating in training selects an output result of a gating network model on the basis of the selection parameter and trains both the gating network model and expert models on the basis of the selected result. Therefore, the sub-nodes align the selection of output results of the gating network model, thereby improving the performance of an aggregation model at the central node.

Description

Distributed training method, device, system, chip module and storage medium

This application claims priority to the Chinese patent application filed with the State Intellectual Property Office of China on December 26, 2023, with application number 202311819054.6 and invention name “Distributed training method, device, system, chip module and storage medium”, all contents of which are incorporated by reference in this application.

Technical Field

The present application relates to artificial intelligence (AI), and in particular to a distributed training method, device, system, chip module and storage medium.

Background Art

Federated learning (FL) is a machine learning paradigm for distributed training. In the federated learning framework, different sub-nodes train a public model based on local data to obtain a local model that can better fit the characteristics of the local data. However, after the local models of these sub-nodes are aggregated by the central node, the generalization of the entire model is improved, but the public model's ability to fit the local data characteristics of each sub-node decreases, and personalized enhancement is needed.

Therefore, a mixture of experts (MOE) model is proposed. The global model issued by the central node includes multiple expert models. Each child node selects at least a part of the output results of the expert model based on the output results of the gated network model to train the global model, obtains gradient information, and reports the gradient information to the central node for model aggregation.

However, if each child node has a different understanding of how to use the output of the gating network model and how to train the expert model, the performance of the aggregation model of the central node will be degraded.

In view of this, for federated learning scenarios involving hybrid expert models, how to improve the performance of the aggregation model of the central node is an urgent problem to be solved.

Summary of the invention

The present application provides a distributed training method, device, system, chip module and storage medium to improve the performance of the aggregation model of the central node.

In a first aspect, a distributed training method is provided, the method being applied to a distributed training system, the global model of the distributed training system comprising N expert models and one gated network model, wherein N is a positive integer, the method comprising: receiving first information, the first information being used to indicate a selection parameter, the selection parameter being used to select an output result of the gated network model; and sending second information, the second information being used to indicate model parameter information of the gated network model and model parameter information of at least one expert model selected from the N expert models, the selected at least one expert model being obtained based on the output result of the gated network model.

In this aspect, all sub-nodes participating in the training receive the selection parameters sent by the central node, and during the training process of the gated network model and the expert model, the selection parameters are applied to the output results of the gated network model, thereby enabling each sub-node to align the selection of the output results of the gated network model and improving the performance of the aggregation model of the central node.

Exemplarily, the output result of the gated network model is a value between [0, 1] after being processed by a set function.

In a possible implementation, the method further includes: inputting sample data into the N expert models and the gated network model to obtain N first output results of the N expert models and N second output results of the gated network model, wherein the N second output results correspond one-to-one to the N first output results respectively; for each first output result of the N first output results and each second output result of the N second output results, for the second output result that satisfies the selection parameter, selecting the first output result corresponding to the second output result; and based on the selected at least one first output result and the true value information, obtaining model parameter information of the selected at least one expert model and the model parameter information of the gated network model.

In another possible implementation, the selection parameter is a first threshold.

In this implementation, for example, the value range of the first threshold can be (0,1). If the output result of the gated network model is greater than the first threshold, the output result of the gated network model is retained; if the output result of the gated network model is less than the first threshold, the output result of the gated network model is discarded. In particular, if the output result of the gated network model is equal to the first threshold, it can be agreed to retain or discard the output result of the gated network model.

In another possible implementation, the first information is further used to indicate at least one of the following information: identifications of the N expert models, a competition mode of the N expert models, a training task, and types of input and output of the global model.

In this implementation, the competitive mode can also be called a usage mode, a cooperation mode, a collaboration mode, an indication of whether to collaborate, etc. The competitive mode indicates whether the N expert models collaborate or compete. Since the same training node obtains different gate weights and expert weights by training the hybrid expert model based on different competitive modes, it is hoped that all child nodes participating in the training use/train the expert model in at least the same way. Therefore, the central node can indicate the competitive mode of the N expert models to the child nodes.

In another possible implementation, the method further includes: sending third information, wherein the third information is used to indicate at least one of the following information: the memory space size of the child node, the computing power information of the child node, whether model training is supported, and the type of model supported for training.

In yet another possible implementation, the second information is further used to indicate an identifier of the selected at least one expert model.

In another possible implementation, the acquiring, based on the selected at least one first output result and the true value information, model parameter information of the selected at least one expert model and model parameter information of the gated network model comprises any one of the following operations: acquiring model parameter information of the selected at least one expert model and model parameter information of the gated network model based on an average value and true value information of the selected at least one first output result; or weighting and averaging the selected at least one first output result based on at least one second output result respectively corresponding to the selected at least one first output result to obtain a weighted average value of the selected at least one first output result, and acquiring model parameter information of the selected at least one expert model and model parameter information of the gated network model based on the weighted average value and true value information of the selected at least one first output result.

In yet another possible implementation, the method further includes: receiving fourth information, where the fourth information is used to indicate an updated selection parameter, and the updated selection parameter is obtained based on the second information.

In this implementation, if the central node updates the selection parameters, fourth information can be sent to each child node, where the fourth information is used to indicate the updated selection parameters, so that the fourth information selects the output result of the gated network model based on the updated selection parameters during the next round of training.

In a second aspect, a distributed training method is provided, which is applied to a distributed training system, wherein a global model of the distributed training system includes N expert models and a gated network model, wherein the N are positive integers, and the method includes: sending first information to multiple child nodes, wherein the first information is used to indicate a selection parameter, and the selection parameter is used to select an output result of the gated network model; and receiving multiple second information respectively, wherein each of the multiple second information is used to indicate model parameter information of the gated network model and model parameter information of at least one expert model selected from the N expert models, and the selected at least one expert model is obtained based on the output result of the gated network model.

In this aspect, the central node indicates the selection parameters so that all sub-nodes participating in the training will apply the selection parameters to the output results of the gated network model during the training of the gated network model and the expert model, thereby enabling each sub-node to align the selection of the output results of the gated network model and improve the performance of the aggregation model of the central node.

Exemplarily, the output result of the gated network model is a value between [0, 1] after being processed by a set function.

In a possible implementation, the selection parameter is a first threshold.

In another possible implementation, the first information is further used to indicate at least one of the following information: identifications of the N expert models, a competition mode of the N expert models, a training task, and types of input and output of the global model.

In another possible implementation, the method further includes: receiving third information, where the third information is used to indicate at least one of the following information: the memory space size of the child node, the computing power information of the child node, whether model training is supported, and the type of model supported for training.

In yet another possible implementation, the second information is further used to indicate an identifier of the selected at least one expert model.

In yet another possible implementation, the method further includes: updating the global model based on the plurality of second information.

In yet another possible implementation, the method further includes: updating the selection parameter based on the plurality of second information; and sending fourth information, where the fourth information is used to indicate the updated selection parameter.

In a third aspect, a distributed training device is provided for implementing the distributed training method in the above-mentioned first aspect or any one of the implementations of the first aspect. The device may be a sub-node/third-party device, or a module (such as a processor, chip, or chip system, etc.) applied to a sub-node/third-party device, or a logical node, logical module or software that can implement all or part of the functions of a sub-node/third-party device. In one implementation, the distributed training device may include a sending unit, a receiving unit, and may also include a processing unit. The sending unit and the receiving unit may be independent or combined together (which may be referred to as a "transceiver unit").

In a fourth aspect, a distributed training device is provided for implementing the distributed training method in the second aspect or any one of the implementations of the second aspect. The device may be a central node, or a module applied to the central node (such as a processor, a chip, or a chip system, etc.), or a logical node, a logical module, or software that can implement all or part of the functions of the central node. In one implementation, the distributed training device may include a sending unit, a receiving unit, and may also include a processing unit. The sending unit and the receiving unit may be independent or combined together (which may be referred to as a "transceiver unit").

In a possible implementation, the distributed training device in the third to fourth aspects includes a unit for respectively executing the method in any aspect or any implementation of the first to second aspects.

In another possible implementation, the distributed training device in the third to fourth aspects above includes a processing circuit coupled to a memory; the processing circuit is configured to enable the device to perform corresponding functions in the above distributed training method. The memory is used to couple with the processing circuit, which stores the necessary programs (instructions) and/or data for the device. Optionally, the distributed training device may further include a communication interface for enabling communication between the device and other network elements. Optionally, the memory may be located inside the distributed training device or outside the distributed training device. Exemplarily, the processing circuit may be a processor or a circuit in a processor for processing.

When the distributed training device in the third to fourth aspects is a chip, the sending unit may be an output unit, such as an output circuit or a communication interface; the receiving unit may be an input unit, such as an input circuit or a communication interface. When the distributed training device is a terminal device, the sending unit may be a transmitter or a transmitter; the receiving unit may be a receiver or a receiver.

In a fifth aspect, a computer-readable storage medium is provided, in which a computer program or instruction is stored. When the computer program or instruction is executed, the methods described in the above aspects are implemented.

In a sixth aspect, a computer program product comprising instructions is provided, which, when executed on a distributed training device, enables the distributed training device to execute the methods described in the above aspects.

In a seventh aspect, a distributed training system is provided, which includes the distributed training device described in the third aspect and the distributed training device described in the fourth aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the architecture of a distributed training system provided in an embodiment of the present application;

FIG2 is a simplified schematic diagram of a wireless communication system provided by an embodiment of the present application;

FIG3 is a schematic diagram of the architecture of another distributed training system provided by the present application;

4A to 4D are schematic diagrams of a network architecture provided in an embodiment of the present application;

FIG5 is a schematic diagram of a neuron structure;

Fig. 6 is a schematic diagram of a neural network;

FIG7 is a schematic diagram of an AI application framework;

FIG8 is a schematic diagram of the architecture of another communication system provided in an embodiment of the present application;

9A to 9E are schematic diagrams of the structure of a global model provided in an embodiment of the present application;

FIG10 is a schematic diagram of a flow chart of a distributed training method provided in an embodiment of the present application;

FIG11 is a schematic diagram of optimal beam training according to an example of an embodiment of the present application;

FIG12 is a schematic diagram of the structure of a distributed training device provided in an embodiment of the present application;

FIG13 is a schematic diagram of the structure of another distributed training device provided in an embodiment of the present application.

DETAILED DESCRIPTION

The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application.

The embodiments of the present application can be applied to a distributed training system as shown in Figure 1, which includes a central node and multiple child nodes. Exemplarily, the distributed training system can be a federated learning system or a Gossip learning system. Model parameter information can be transmitted between the central node and each child node. The distributed training system can also include a third-party device (not shown in the figure), which can serve as a model training entity for the central node/child node.

The machine learning model trained by the distributed training system can be for non-wireless communication services, such as image recognition, natural language processing, etc., or for wireless communication services, such as beam selection based on environmental information.

The technology provided in this application can be applied to various communication systems, for example, the communication system can be a ^fourth generation (4G) communication system (such as a long term evolution (LTE) system), a ^fifth generation (5G) communication system, a world-wide interoperability for microwave access (WiMAX), a wireless local area network (WLAN) system, a satellite communication system, a fusion system of multiple systems, or a future communication system, such as a ^sixth generation (6G) communication system. Among them, the 5G communication system can also be called a new radio (NR) system.

A network element in a communication system can send a signal to another network element or receive a signal from another network element. The signal may include information, signaling, or data, etc. The network element may also be replaced by an entity, a network entity, a device, a terminal device, a communication module, a node, a communication node, etc. The network element is used as an example for description in this application. For example, a communication system may include at least one terminal device and at least one network device. The network device may send a downlink signal to the terminal device, and/or the terminal device may send an uplink signal to the network device. In addition, it can be understood that if a plurality of terminal devices are included in the communication system, the plurality of terminal devices may also send signals to each other, that is, the signal sending network element and the signal receiving network element may both be terminal devices.

Referring to FIG. 2 , FIG. 2 is a simplified schematic diagram of a wireless communication system provided in an embodiment of the present application. As shown in FIG. 2 , the wireless communication system includes a wireless access network 100. The wireless access network 100 may be a next generation (e.g., 6G or higher) wireless access network, or a traditional (e.g., 5G, 4G) wireless access network. One or more terminal devices (120a-120j, collectively referred to as 120) may be connected to each other, or to one or more network devices (110a, 110b, collectively referred to as 110) in the wireless access network 100. Optionally, FIG. 2 is only a schematic diagram, and other devices may also be included in the wireless communication system, such as core network devices, wireless relay devices, and/or wireless backhaul devices, which are not shown in FIG. 2 .

Optionally, in practical applications, the wireless communication system may include multiple network devices (also referred to as access network devices) at the same time, and may also include multiple terminal devices at the same time. A network device may serve one or more terminal devices at the same time. A terminal device may also access one or more network devices at the same time. The embodiment of the present application does not limit the number of terminal devices and network devices included in the wireless communication system.

Among them, the network device can be an entity on the network side for transmitting or receiving signals. The network device can be an access device for the terminal device to access the wireless communication system by wireless means, such as the network device can be a base station. The base station can broadly cover the following various names, or be replaced with the following names, such as: radio access network (RAN) node, node B (NodeB), evolved NodeB (evolved NodeB, eNB), next generation NodeB (next generation NodeB, gNB), network equipment in open radio access network (open radio access network, O-RAN), relay station, access point, transmission point (transmitting and receiving point, TRP), transmitting point (transmitting point, TP), master eNB (MeNB), secondary eNB (SeNB), multi-standard radio (multi-standard radio, MSR) node, home base station, network controller, access Node, wireless node, access point (AP), transmission node, transceiver node, building baseband unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), centralized unit (CU), distributed unit (DU), radio unit (RU), centralized unit control plane (CU-CP) node, centralized unit user plane (CU-UP) node, positioning node, RAN intelligent controller (RIC), etc. The base station can be a macro base station, a micro base station, a relay node, a donor node or the like, or a combination thereof. The network device can also refer to a communication module, a modem or a chip used to be set in the aforementioned device or apparatus. The network device may also be a mobile switching center and a device that performs base station functions in device-to-device (D2D), vehicle-to-everything (V2X), and machine-to-machine (M2M) communications, a network-side device in a 6G network, and a device that performs base station functions in future communication systems. The network device may support networks with the same or different access technologies. The embodiments of the present application do not limit the specific technology and specific device form used by the network device.

The network equipment can be fixed or mobile. For example, base stations 110a, 110b are stationary and are responsible for wireless transmission and reception in one or more cells from terminal device 120. The helicopter or drone 120i shown in Figure 2 can be configured to act as a mobile base station, and one or more cells can move according to the location of the mobile base station 120i. In other examples, the helicopter or drone (120i) can be configured to act as a terminal device that communicates with base station 110b.

In the present application, the communication device used to implement the above-mentioned access network function can be a network device, or a network device with partial access network functions, or a device capable of supporting the access network function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module, which can be installed in the network device or used in combination with the network device. In the method of the present application, the communication device used to implement the network device function is described as a network device.

The terminal device can be an entity on the user side for receiving or transmitting signals, such as a mobile phone. The terminal device can be used to connect people, objects and machines. The terminal device can communicate with one or more core networks through a network device. The terminal device includes a handheld device with a wireless connection function, other processing devices connected to a wireless modem, or a vehicle-mounted device. The terminal device can be a portable, pocket-sized, handheld, computer-built-in or vehicle-mounted mobile device. The terminal device 120 can be widely used in various scenarios, such as cellular communication, D2D, V2X, point-to-point (P2P), machine-to-machine (M2M), machine type communication (MTC), Internet of Things (IoT), virtual reality (VR), augmented reality (AR), industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, drone, robot, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc. Some examples of terminal devices 120 are: 3GPP standard user equipment (UE), fixed equipment, mobile devices, handheld devices, wearable devices, cellular phones, smart phones, session initiation protocol (SIP) phones, laptops, personal computers, smart books, vehicles, satellites, global positioning system (GPS) equipment, target tracking equipment, drones, helicopters, aircraft, ships, remote control equipment, smart home equipment, industrial equipment, personal communication service (PCS) phones, wireless local loop (WLL) stations, personal digital assistants (PDAs), etc. The terminal device 120 may be a wireless device in the above-mentioned scenarios or a device used to be set in a wireless device, for example, a communication module, a modem or a chip in the above-mentioned device. The terminal device may also be referred to as a terminal, a terminal device, a UE, a mobile station (MS), a mobile terminal (MT), or the like. The terminal device may also be referred to as a terminal, a terminal device, a UE, a mobile station (MS), a mobile terminal (MT), or the like. The terminal device may also be a terminal device in a future wireless communication system. The terminal device can be used in a dedicated network device or a general-purpose device. The embodiments of the present application do not limit the specific technology and specific device form used by the terminal device.

Optionally, the terminal device can be used to act as a base station. For example, the UE can act as a scheduling entity that provides sidelink signals between UEs in V2X, D2D, or P2P, etc. As shown in Figure 2, the cellular phone 120a and the car 120b communicate with each other using sidelink signals. The cellular phone 120a and the smart home device 120e communicate without relaying the communication signal through the base station 110b.

In the present application, the communication device for realizing the functions of the terminal device may be a terminal device, or a terminal device having some functions of the above terminal devices, or a device capable of supporting the functions of the above terminal devices, such as a chip system, which may be installed in the terminal device or used in combination with the terminal device. In the present application, the chip system may be composed of a chip, or may include a chip and other discrete devices. In the technical solution provided in the present application, the communication device is described as a terminal device or UE as an example.

Optionally, a wireless communication system is usually composed of cells, and a base station provides management of the cell. The base station provides communication services to multiple mobile stations (MS) in the cell. The base station includes a baseband unit (BBU) and a remote radio unit (RRU). The BBU and RRU can be placed in different places, for example: the RRU is remote and placed in an area with high traffic volume, and the BBU is placed in a central computer room. The BBU and RRU can also be placed in the same computer room. The BBU and RRU can also be different components under one rack. Optionally, a cell can correspond to one carrier or component carrier.

In some deployments, the network device mentioned in the embodiments of the present application may be a device including a CU, or a DU, or a device including a CU and a DU, or a device including a control plane CU node (centralized unit-control plane (CU-CP)) and a user plane CU node (centralized unit-user plane (CU-UP)) and a DU node. For example, the network device may include a gNB-CU-CP, a gNB-CU-UP, and a gNB-DU.

In some deployments, multiple RAN nodes collaborate to assist the terminal in achieving wireless access, and different RAN nodes implement part of the functions of the base station. For example, the RAN node can be a CU, DU, CU-CP, CU-UP, or RU. The CU and DU can be set separately, or can also be included in the same network element, such as a BBU. The RU can be included in a radio frequency device or a radio frequency unit, such as an RRU, AAU, or RRH.

The RAN node may support one or more types of fronthaul interfaces, and different fronthaul interfaces correspond to DUs and RUs with different functions. If the fronthaul interface between the DU and the RU is a common public radio interface (CPRI), the DU is configured to implement one or more of the baseband functions, and the RU is configured to implement one or more of the radio frequency functions. If the fronthaul interface between the DU and the RU is another interface, relative to the CPRI, part of the baseband functions of the downlink and/or uplink, such as, for the downlink, one or more of precoding, digital beamforming (BF), or inverse fast Fourier transform (IFFT)/adding cyclic prefix (CP), are moved from the DU to the RU for implementation, and for the uplink, one or more of digital beamforming (BF), or fast Fourier transform (FFT)/removing cyclic prefix, are moved from the DU to the RU for implementation. In a possible implementation, the interface may be an enhanced common public radio interface (eCPRI). Under the eCPRI architecture, the division between DU and RU is different, corresponding to different types (category, Cat) of eCPRI, such as eCPRI Cat A, B, C, D, E, and F.

Taking eCPRI Cat A as an example, for downlink transmission, based on layer mapping, the DU is configured to implement one or more functions before layer mapping (i.e., one or more functions of coding, rate matching, scrambling, modulation, and layer mapping), while other functions after layer mapping (e.g., one or more functions of resource element (RE) mapping, digital beamforming (BF), or inverse fast Fourier transform (IFFT)/adding cyclic prefix (CP)) are moved to the RU for implementation. For uplink transmission, based on de-RE mapping, the DU is configured to implement one or more functions before de-mapping (i.e., one or more functions of decoding, de-rate matching, de-scrambling, demodulation, inverse discrete Fourier transform (IDFT), channel equalization, and de-RE mapping), while other functions after de-mapping (e.g., one or more functions of digital BF or FFT/CP removal) are moved to the RU for implementation. It can be understood that for the functional description of DU and RU corresponding to various types of eCPRI, reference can be made to the eCPRI protocol and will not be repeated here.

In one possible design, the processing unit for implementing the baseband function in the BBU is called a baseband high layer (BBH) unit, and the processing unit for implementing the baseband function in the RRU/AAU/RRH is called a baseband low layer (BBL) unit.

In different systems, CU (or CU-CP and CU-UP), DU or RU may also have different names, but those skilled in the art can understand their meanings. For example, in an open radio access network (ORAN) system, CU may also be called O-CU (open CU), DU may also be called O-DU, CU-CP may also be called O-CU-CP, CU-UP may also be called O-CU-UP, and RU may also be called O-RU. Any unit in the CU (or CU-CP, CU-UP), DU and RU in this application may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

In the embodiments of the present application, the device for realizing the function of the network device may be a network device; or it may be a device capable of supporting the network device to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module. The device may be installed in the network device or used in combination with the network device. In the embodiments of the present application, only the device for realizing the function of the network device is a network device as an example for explanation, and the scheme of the embodiments of the present application is not limited.

It can be understood that the present application can be applied between network devices and terminal devices.

Protocol layer structure between network devices and terminal devices:

The communication between the network device and the terminal device follows a certain protocol layer structure. The protocol layer structure may include a control plane protocol layer structure and a user plane protocol layer structure. For example, the control plane protocol layer structure may include the functions of the radio resource control (RRC) layer, the packet data convergence protocol (PDCP) layer, the radio link control (RLC) layer, the medium access control (MAC) layer and the physical layer. For example, the user plane protocol layer structure may include the functions of the PDCP layer, the RLC layer, the MAC layer and the physical layer. In a possible implementation, the service data adaptation protocol (SDAP) layer may also be included above the PDCP layer.

Optionally, the protocol layer structure between the network device and the terminal device may also include an artificial intelligence (AI) layer for transmitting data related to AI functions.

Taking data transmission between network devices and terminal devices as an example, data transmission needs to pass through the user plane protocol layer, such as the SDAP layer, PDCP layer, RLC layer, MAC layer, and physical layer. Among them, the SDAP layer, PDCP layer, RLC layer, MAC layer, and physical layer can also be collectively referred to as the access layer. According to the transmission direction of the data, it is divided into sending or receiving, and each of the above layers is divided into a sending part and a receiving part. Taking downlink data transmission as an example, after the PDCP layer obtains data from the upper layer, it transmits the data to the RLC layer and the MAC layer, and then the MAC layer generates a transmission block, which is then wirelessly transmitted through the physical layer. Data is encapsulated accordingly in each layer. For example, the data received by a layer from the upper layer of the layer is regarded as the service data unit (SDU) of the layer, which becomes a protocol data unit (PDU) after being encapsulated by the layer, and then passed to the next layer.

Exemplarily, the terminal device may also have an application layer and a non-access layer. The application layer may be used to provide services to applications installed in the terminal device. For example, downlink data received by the terminal device may be sequentially transmitted from the physical layer to the application layer, and then provided to the application by the application layer; for another example, the application layer may obtain data generated by the application, and sequentially transmit the data to the physical layer and send it to other communication devices. The non-access layer may be used to forward user data, such as forwarding uplink data received from the application layer to the SDAP layer, or forwarding downlink data received from the SDAP layer to the application layer.

It should be understood that the number and type of each device in the communication system shown in Figure 2 are for illustration only, and the present application is not limited to this. In actual applications, the communication system may also include more terminal devices, more network devices, and other network elements, such as core network devices, and/or network elements for implementing artificial intelligence functions.

It is understandable that all or part of the functions implemented by one or more of the terminal devices, network devices, core network devices, or network elements used to implement artificial intelligence functions can be virtualized, that is, implemented by one or more of the proprietary processors or general-purpose processors and corresponding software modules. Among them, since the terminal devices and network devices involve interfaces for air interface transmission, the transceiver functions of the interfaces can be implemented by hardware. Core network devices, such as operation administration and maintenance (OAM) network elements, can be virtualized. Optionally, one or more functions of the virtualized terminal devices, network devices, core network devices, or network elements used to implement artificial intelligence functions can be implemented by cloud devices, such as cloud devices in over the top (OTT) systems.

In an embodiment of the present application, when the central node is a network device and the child node is a terminal device (such as a UE), the network device and UE1 to UE5 can form a distributed AI training system as shown in Figure 3. In this communication system, UE1 to UE5 can send data to the network device, and the network device needs to receive uplink data sent by UE1 to UE5. The uplink data can be the representation difference or model parameter calculated by the child node, or it can be the feedback amount containing its status information. At the same time, the network device can send configuration information to UE1-UE5. The configuration information can be the model parameter data used by the central node to synchronize each child node, or it can be control data indicating the training method of the child node. The data between the network equipment and the UE can be carried on physical channels, such as the physical downlink control channel (physical downlink control channel, PDCCH), the physical downlink shared channel (physical downlink shared channel, PDSCH), the physical uplink shared channel (physical uplink shared channel, PUSCH) or the physical uplink control channel (physical uplink control channel, PUCCH); for example, the physical sidelink control channel (physical sidelink control channel, PSCCH) and the physical sidelink shared channel (physical sidelink shared channel, PSSCH).

In order to support AI technology in wireless networks, AI nodes may also be introduced into the network.

Optionally, the AI node can be deployed in one or more of the following locations in the communication system: network equipment, terminal equipment, or core network equipment, etc., or the AI node can also be deployed separately, for example, deployed in a location other than any of the above devices, such as a host or cloud server in an over-the-top (OTT) system. The AI node can communicate with other devices in the communication system, and the other devices can be, for example, one or more of the following: network equipment, terminal equipment, or network elements of the core network, etc.

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

It can also be understood that AI nodes can be independent devices, or they can be integrated into the same device to implement different functions, or they can be network elements in hardware devices, or they can be software functions running on dedicated hardware, or they can be virtualized functions instantiated on a platform (for example, a cloud platform). This application does not limit the specific form of the above-mentioned AI nodes.

An AI node can be an AI network element or an AI module.

One or more AI modules are provided in one or more devices of these network element nodes, such as core network equipment, access network nodes (RAN nodes), terminals or OAM. The access network node can be a separate RAN node, or it can include multiple RAN nodes, for example, including CU and DU. The CU and/or DU can also be provided with one or more AI modules. Optionally, the CU can also be split into CU-CP and CU-UP. One or more AI models are provided in the CU-CP and/or CU-UP.

The AI module is used to implement the corresponding AI function. The AI modules deployed in different network elements can be the same or different. The model of the AI module can implement different functions according to different parameter configurations. The model of the AI module can be configured based on one or more of the following parameters: structural parameters (such as the number of neural network layers, the width of the neural network, the connection relationship between layers, the weight of the neuron, the activation function of the neuron, or at least one of the biases in the activation function), input parameters (such as the type of input parameters and/or the dimension of input parameters), or output parameters (such as the type of output parameters and/or the dimension of output parameters). Among them, the bias in the activation function can also be called the bias of the neural network.

An AI module can have one or more models. A model can be inferred to obtain an output, which includes one parameter or multiple parameters. The learning process, training process, or inference process of different models can be deployed in different nodes or devices, or can be deployed in the same node or device.

The communication system includes a RAN intelligent controller. For example, the RIC may be the above-mentioned AI module, which is used to implement AI-related functions. The RIC includes near-real time RIC (near-real time RIC, near-RT RIC) and non-real time RIC (non-real time RIC, Non-RT RIC). Among them, the non-real time RIC mainly processes non-real-time information, such as data that is not sensitive to delay, and the delay of the data can be in the second level. The real-time RIC mainly processes near-real-time information, such as data that is relatively sensitive to delay, and the delay of the data is in the tens of milliseconds.

Near real-time RIC is used for model training and reasoning. For example, it is used to train an AI model and use the AI model for reasoning. Near real-time RIC can obtain information on the network side and/or the terminal side from RAN nodes (such as CU, CU-CP, CU-UP, DU and/or RU) and/or terminals. This information can be used as training data or reasoning data. Optionally, near real-time RIC can submit the reasoning results to the RAN node and/or terminal. Optionally, the reasoning results can be exchanged between the CU and DU, and/or between the DU and RU. For example, the near real-time RIC submits the reasoning results to the DU, and the DU sends it to the RU.

Non-real-time RIC is also used for model training and reasoning. For example, it is used to train an AI model and use the model for reasoning. Non-real-time RIC can obtain information on the network side and/or the terminal side from RAN nodes (such as CU, CU-CP, CU-UP, DU and/or RU) and/or terminals. This information can be used as training data or reasoning data, and the reasoning results can be submitted to the RAN node and/or the terminal. Optionally, the reasoning results can be exchanged between the CU and the DU, and/or between the DU and the RU. For example, the non-real-time RIC submits the reasoning results to the DU, and the DU sends it to the RU.

The near real-time RIC and the non-real-time RIC may also be separately set as a network element. Optionally, the near real-time RIC and the non-real-time RIC may also be part of other devices, for example, the near real-time RIC is set in a RAN node (for example, in a CU or DU), and the non-real-time RIC is set in an OAM, a cloud server, a core network device, or other network devices.

Exemplarily, the configuration of near real-time RIC and non-real-time RIC in the network architecture may be as shown in FIG. 4A to FIG. 4D :

As shown in (a) of FIG. 4A , in a first possible implementation, the network device includes a near real-time RIC module for performing model learning and/or reasoning.

As shown in (b) of FIG. 4A , in a second possible implementation, in a communication system, a non-real-time RIC may be included outside a network device. Optionally, the non-real-time RIC may be located in an OAM or a core network device.

As shown in (c) of Figure 4A, in a third possible implementation, the network device includes a near real-time RIC, and the network device also includes a non-real-time RIC. Optionally, the non-real-time RIC may be located in the OAM or in the core network device.

Compared with (c) in Fig. 4A, the CU is separated into CU-CP and CU-UP in Fig. 4B. The settings of the near real-time RIC and the non-real-time RIC are the same as those in (c) in Fig. 4A.

As shown in Figure 4C, optionally, the network device includes one or more AI entities, and the function of the AI entity is similar to the above-mentioned near real-time RIC. Optionally, the OAM includes one or more AI entities, and the function of the AI entity is similar to the above-mentioned non-real-time RIC. Optionally, the core network device includes one or more AI entities, and the function of the AI entity is similar to the above-mentioned non-real-time RIC. When both the OAM and the core network device include AI entities, the models trained by their respective AI entities are different, and/or the models used for reasoning are different. In the present application, the difference in models may include at least one of the following differences: structural parameters of the model (such as the number of layers of the model, and/or weights, etc.), input parameters of the model, or output parameters of the model.

Relative to Figure 4C, the network device in Figure 4D is separated into CU and DU. Optionally, the CU may include an AI entity, and the function of the AI entity is similar to the above-mentioned near real-time RIC. Optionally, the DU may include an AI entity, and the function of the AI entity is similar to the above-mentioned near real-time RIC. When both the CU and the DU include AI entities, the models trained by their respective AI entities are different, and/or the models used for reasoning are different. Optionally, the CU in Figure 4D can be further split into CU-CP and CU-UP. Optionally, one or more AI models can be deployed in the CU-CP. And/or, one or more AI models can be deployed in the CU-UP. Optionally, in Figure 4C or Figure 4D, the OAM of the network device and the OAM of the core network device can be deployed separately and independently.

For ease of understanding, the AI technology involved in this application is first introduced below. It is understandable that this introduction is not intended to limit this application.

(1) AI Model

AI refers to the intelligence displayed by machines created by humans. Usually, artificial intelligence refers to the technology that presents human intelligence through ordinary computer programs. Artificial intelligence can be defined as machines or computers that imitate humans and have cognitive functions related to human thinking, such as learning and problem solving. Artificial intelligence is able to learn from past experiences, make reasonable decisions, and respond quickly. The goal of artificial intelligence is to understand intelligence by building computer programs with symbolic reasoning or reasoning.

Machine learning (ML) is a way to achieve artificial intelligence, that is, to solve problems in artificial intelligence by means of machine learning. The theory of machine learning mainly designs and analyzes some algorithms that allow computers to "learn" automatically. Machine learning algorithms are a type of algorithm that automatically analyzes data to obtain patterns and uses the patterns to predict unknown data. Because learning algorithms involve a lot of statistical theories, machine learning is particularly closely related to inferential statistics, and is also called statistical learning theory.

Machine learning can be divided into supervised learning, unsupervised learning, and reinforcement learning.

Supervised learning uses machine learning algorithms to learn the mapping relationship from sample values to sample labels based on the collected sample values and sample labels, and uses machine learning models to express the learned mapping relationship. The process of training a machine learning model is the process of learning this mapping relationship. For example, in signal detection, the received signal containing noise is the sample, and the real constellation point corresponding to the signal is the label. Machine learning expects to learn the mapping relationship between samples and labels through training, that is, to make the machine learning model learn a signal detector. During training, the model parameters are optimized by calculating the error between the model's predicted value and the real label. Once the mapping relationship is learned, the learned mapping can be used to predict the label of each new sample. The mapping relationship learned by supervised learning can include linear mapping and nonlinear mapping. According to the type of label, the learning task can be divided into classification task and regression task.

Unsupervised learning is based only on the collected sample values, using algorithms to discover the inherent patterns of samples. There is a type of algorithm in unsupervised learning that uses the sample itself as a supervisory signal, that is, the model learns the mapping relationship from sample to sample, which is called self-supervised learning. During training, the model parameters are optimized by calculating the error between the model's predicted value and the sample itself. Self-supervised learning can be used in applications such as signal compression and decompression recovery. Common algorithms include autoencoders and adversarial generative networks.

Reinforcement learning is different from supervised learning. It is a type of algorithm that learns problem-solving strategies by interacting with the environment. Unlike supervised and unsupervised learning, reinforcement learning problems do not have clear "correct" action label data. The algorithm needs to interact with the environment to obtain reward signals from the environment, and then adjust the decision-making actions to obtain a larger reward signal value. For example, in downlink power control, the reinforcement learning model adjusts the downlink transmission power of each user according to the total system throughput fed back by the wireless network, and then expects to obtain a higher system throughput. The goal of reinforcement learning is also to learn the mapping relationship between the state of the environment and the optimal decision action. However, because the label of the "correct action" cannot be obtained in advance, the network cannot be optimized by calculating the error between the action and the "correct action". Reinforcement learning training is achieved through iterative interaction with the environment.

AI models are algorithms or computer programs that can realize AI functions. They are the specific implementation of AI technology functions. AI models represent the mapping relationship between the input and output of the model. The types of AI models can be neural networks, linear regression models, decision tree models, support vector machines (SVM), Bayesian networks, Q learning models or other machine learning models.

(2) Deep neural network (DNN)

Deep neural network is a specific implementation form of AI or machine learning technology. According to the universal approximation theorem, neural network can theoretically approximate any continuous function, so that neural network has the ability to learn any mapping. Traditional communication systems require rich expert knowledge to design communication modules, while DNN-based deep learning communication systems can automatically discover implicit pattern structures from large data sets, establish mapping relationships between data, and obtain performance that is superior to traditional modeling methods.

The idea of DNN comes from the neuron structure of brain tissue. For example, each neuron performs a weighted sum operation on its input values and outputs the operation result through an activation function. As shown in Figure 5, it is a schematic diagram of the neuron structure. Assume that the input of the neuron is x = [x ₀ , x ₁ ,…, x _n ], and the weights corresponding to each input are w = [w ₀ , w ₁ ,…, w _n ], where w _i is the weight of _xi and is used to weight _xi . The bias for weighted summation of input values according to the weights is, for example, b. The activation function can take many forms. Assuming that the activation function of a neuron is: y = f(z) = max(0,z), the output of the neuron is:

For another example, the activation function of a neuron is: y = f(z) = z, then the output of the neuron is:

b, w _i , _xi can be decimals, integers (such as 0, positive integers or negative integers), or complex numbers. The activation functions of different neurons in a neural network can be the same or different.

A neural network generally includes multiple layers, each of which may include one or more neurons. By increasing the depth and/or width of a neural network, the expressive power of the neural network can be improved, providing a more powerful information extraction and abstract modeling capability for complex systems. Among them, the depth of a neural network may refer to the number of layers included in the neural network, wherein the number of neurons included in each layer may be referred to as the width of the layer. In one implementation, the neural network includes an input layer and an output layer. The input layer of the neural network processes the received input information through neurons, passes the processing results to the output layer, and obtains the output result of the neural network from the output layer. In another implementation, the neural network includes an input layer, a hidden layer, and an output layer, and reference may be made to the schematic diagram of the neural network in FIG6. The input layer of the neural network processes the received input information through neurons, passes the processing results to the middle hidden layer, and the hidden layer calculates the received processing results to obtain the calculation results. The hidden layer passes the calculation results to the output layer or the adjacent hidden layer, and finally obtains the output result of the neural network from the output layer. Among them, a neural network may include one hidden layer, or include multiple hidden layers connected in sequence, without limitation.

According to the way the network is constructed, DNN can include feedforward neural network (FNN), convolutional neural network (CNN) and recurrent neural network (RNN). Figure 6 shows an FNN network, which is characterized by the neurons in adjacent layers being fully connected to each other, which usually requires a large amount of storage space and leads to high computational complexity.

CNN is a neural network that is specifically designed to process data with a grid-like structure. For example, time series data and image data can be considered to be data with a grid-like structure. CNN does not use all the input information for calculations at once, but uses a fixed-size window to intercept part of the information for convolution operations, which greatly reduces the amount of calculation of model parameters. In addition, depending on the type of information intercepted by the window (for example, people and objects in a picture are different types of information), each window can use different convolution kernel operations, which enables CNN to better extract the features of the input data.

RNN is a type of DNN network that uses feedback time series information. Its input includes the new input value at the current moment and its own output value at the previous moment. RNN is suitable for obtaining sequence features that are correlated in time, and is particularly suitable for applications such as speech recognition and channel coding.

The above-mentioned FNN, CNN, and RNN are common neural network structures, which are all constructed based on neurons. As mentioned above, each neuron performs a weighted sum operation on its input values, and the weighted sum result generates an output through a nonlinear function. We call the weights of the weighted sum operation of neurons in the neural network and the nonlinear function the parameters of the neural network. Take the neuron with max{0,x} as the nonlinear function as an example, The parameters of the operated neurons are the weight w = [w ₀ ,…,w _n ], the bias of the weighted sum b, and the nonlinear function max{0,x}. The parameters of all neurons in a neural network constitute the parameters of this neural network.

(3) Training Dataset and Inference Data

The training data set is used for training the AI model. The training data set may include the input of the AI model, or the input and target output of the AI model. Among them, the training data set includes one or more training data, and the training data may be a training sample input to the AI model, or it may be the target output of the AI model. Among them, the target output may also be referred to as a label or a label sample. The training data set is one of the important parts of machine learning. Model training is essentially learning certain features from the training data so that the output of the AI model is as close to the target output as possible, such as the difference between the output of the AI model and the target output is as small as possible. The composition and selection of the training data set can determine the performance of the trained AI model to a certain extent.

In addition, during the training process of an AI model (such as a neural network), a loss function can be defined. The loss function describes the gap or difference between the output value of the AI model and the target output value. This application does not limit the specific form of the loss function. The training process of the AI model is to adjust the model parameters of the AI model so that the value of the loss function is less than the threshold, or the value of the loss function meets the target requirements. For example, the AI model is a neural network, and adjusting the model parameters of the neural network includes adjusting at least one of the following parameters: the number of layers, width, weights of neurons, or parameters in the activation function of neurons.

Inference data can be used as input to the trained AI model for inference. During the model inference process, the inference data is input into the AI model, and the corresponding output is the inference result.

(4) AI model design

The design of the AI model mainly includes a data collection link (for example, collecting training data and/or reasoning data), a model training link, and a model reasoning link. It can also include a reasoning result application link. See Figure 7, which illustrates an AI application framework. In the aforementioned data collection link, the data source is used to provide training data sets and reasoning data. In the model training link, the AI model is obtained by analyzing or training the training data provided by the data source. Among them, the AI model represents the mapping relationship between the input and output of the model. Learning the AI model through the model training node is equivalent to learning the mapping relationship between the input and output of the model using the training data. In the model reasoning link, the AI model trained through the model training link is used to reason based on the reasoning data provided by the data source to obtain the reasoning result. This link can also be understood as: inputting the reasoning data into the AI model, and obtaining the output through the AI model, which is the reasoning result. The reasoning result can indicate: the configuration parameters used (executed) by the execution object, and/or the operation performed by the execution object. The reasoning results are published in the reasoning result application link. For example, the reasoning results can be planned uniformly by the execution entity (actor). For example, the execution entity can send the reasoning results to one or more execution objects (for example, core network equipment, network equipment, or terminal equipment, etc.) for execution. For another example, the execution entity can also feedback the performance of the model to the data source to facilitate the subsequent implementation of the model update training.

It is understandable that a network element with artificial intelligence function may be included in the communication system. The above-mentioned AI model design-related links may be performed by one or more network elements with artificial intelligence function. In one possible design, an AI function (such as an AI module or an AI entity) may be configured in an existing network element in the communication system to implement AI-related operations, such as training and/or reasoning of an AI model. For example, the existing network element may be a network device (such as a gNB), a terminal device, a core network device, or a network management system. Among them, the network management system may divide the network management work into three categories according to the actual needs of the operator's network operation: operation, administration, and maintenance. The network management system may also be referred to as an OAM network element, or OAM for short. Operation mainly completes the analysis, prediction, planning, and configuration of daily networks and services; maintenance mainly involves daily operational activities such as testing and fault management of the network and its services. The network management system may detect the network operation status, optimize network connections and performance, improve network operation stability, and reduce network maintenance costs. Or in another possible design, an independent network element may also be introduced into the communication system to perform AI-related operations, such as training an AI model. The independent network element can be called an AI network element or an AI node, etc., and this application does not limit this name. The AI network element can be directly connected to the network equipment in the communication system, or it can be indirectly connected through a third-party device and the network equipment. Among them, the third-party device can be a core network element such as an authentication management function (AMF) network element, a user plane function (UPF) network element, an OAM, a cloud server or other network elements, without limitation. For example, referring to Figure 8, the communication system includes a network device 810, terminal devices 820, 830, and an AI network element 840 is also introduced in the communication system.

In this application, a model can be inferred to obtain one parameter, or multiple parameters. The training process of different models can be deployed in different devices or nodes, or in the same device or node. The reasoning process of different models can be deployed in different devices or nodes, or in the same device or node.

Among them, the model parameters may include one or more of the following structural parameters of the model (such as the number of layers of the model, and/or weights, etc.), the input parameters of the model (such as input dimension, number of input ports), or the output parameters of the model (such as output dimension, number of output ports). It can be understood that the input dimension may refer to the size of an input data. For example, when the input data is a sequence, the input dimension corresponding to the sequence may indicate the length of the sequence. The number of input ports may refer to the number of input data. Similarly, the output dimension may refer to the size of an output data. For example, when the output data is a sequence, the output dimension corresponding to the sequence may indicate the length of the sequence. The number of output ports may refer to the number of output data.

(5) Centralized training

In the past decade, the number of smart devices such as mobile terminals and wearable devices has continued to increase. It is foreseeable that in the near future, billions of IoT devices will be deployed throughout the communication network to achieve automation and intelligence in social operations. The performance of current smart services based on advanced machine learning may benefit from the explosive growth of data on these devices and the computing power available to these UEs themselves. Most machine learning techniques, such as learning algorithms based on deep neural networks, require all available data to be centralized for training. However, since centralized training requires the collection of enough data, which often comes from UEs, UEs are required to upload these data, and the uploading overhead of these data is large. In addition, a large amount of training data is concentrated in one node, and the training efficiency will be limited by the storage space. In addition, collecting data on the UE side may infringe on the privacy of users. Storing massive data for centralized training may cause users to have significant concerns about privacy data leakage. On the other hand, if users only use the data they own to train machine learning algorithms, the training effect is often limited by the limited amount of local data.

(6) Distributed training

Distributed training is an effective solution to the above challenges. This type of technology allows the machine training process to be divided into multiple sub-nodes on the user side to achieve scalability of the learning algorithm. It allows the cloud or server as a central node to collect machine learning models trained by multiple sub-nodes. The central node improves the effect of the entire machine learning training by integrating the models trained by the sub-nodes. Since the training data is always kept in the sub-nodes, distributed learning technology is expected to achieve the same performance as centralized training while using the UE's data and/or computing power to protect user data privacy.

Federated learning is a machine learning paradigm for distributed training. Its original intention was to effectively help multiple institutions use data and conduct machine learning modeling while meeting the requirements of user privacy protection and data security. In the federated learning framework, what is transmitted between nodes is not the data itself, but the intermediate results obtained during training, such as model parameters or gradients. As a distributed machine learning paradigm, federated learning can effectively solve the problem of data silos, allowing participants to jointly model without sharing data, technically breaking down data silos and achieving AI collaboration. According to the different distribution of data sources among the participating parties, federated learning can be divided into three categories: horizontal federated learning, vertical federated learning, and federated transfer learning.

Horizontal federated learning means that when there is a lot of overlap in user features but little overlap in users of two datasets, we split the dataset horizontally (i.e., user dimension) and take out the data with the same user features but different users for training. Vertical federated learning means that when there is a lot of overlap in users of two datasets but little overlap in user features, we split the dataset vertically (i.e., feature dimension) and take out the data with the same users but different user features for training. Federated transfer learning means that when there is little overlap in users and user features of two datasets, we do not split the data, but use transfer learning to overcome the lack of data or labels.

Taking the high-frequency beam management problem as an example, when the network equipment side performs beam scanning of codebook-based synchronization signals and physical broadcast channel (PBCH) blocks (i.e., synchronization signal blocks (SSB)) or channel state information-reference signals (CSI-RS), the channels between users and network equipment at different locations are different. Users measure the received SSB or CSI-RS beams, such as measuring the physical layer reference signal receiving power (L1-reference signal receiving power, L1-RSRP), and feedback the beam identity (ID) corresponding to the maximum RSRP value. AI/ML can be used to train a model, using, for example, a plurality of received SSB/CSI-RS signals (part or all), or the strength (RSRP) (part or all) of a plurality of received SSB/CSI-RS signals or the estimated channel as input to infer the optimal beam ID and feedback it to the network device. Each user can collect his/her own receiving beam/channel information and the corresponding optimal beam ID as samples (i.e., local samples) for training the above AL/ML model. However, the number of samples that each user can collect is limited. The performance of the model obtained by the user using only local data for training will be limited, that is, due to the position relationship, the optimal beam ID of the user may be only a subset of the SSB/CSI-RS codebook. If the user sends local data to the server, the server aggregates the data of each user for model training. Although the performance of the model can be improved, there is a risk of leaking the user's privacy information, such as the user's current location can be inferred through the channel. In order to solve this problem, federated learning can be used. The central node sends a global model to each user participating in federated learning. Each user uses local data to train the global model to obtain a local model, and sends the parameter information of the local model, such as gradients, weights, etc. (encrypted) to the server. The server performs model fusion (model aggression, MA) to update the global model, and sends the global model to each user again. The user continues to update the local model and sends it to the central node. This is repeated for multiple times until convergence.

In the federated learning framework, different sub-nodes train public models based on local data to obtain local models, which can better fit the characteristics of local data. However, after the local models of these sub-nodes are aggregated by the central node, the generalization of the entire model is improved, but the public model's ability to fit the local data characteristics of each sub-node decreases, and needs to be personalized and strengthened.

Therefore, a hybrid expert model is proposed. The global model sent by the central node includes multiple expert models. Each sub-node selects at least a part of the output results of the expert model based on the output results of the gated network model to train the global model, obtains gradient information, and reports the gradient information to the central node for model aggregation.

The hybrid expert model combines multiple networks or multiple expert models to achieve better model performance. A large model based on the hybrid expert model can effectively improve its training efficiency and the number of model parameters.

As shown in FIG. 9A to FIG. 9E , various global model structures are given.

According to whether the global model includes a general layer, whether the expert layer includes a gated network model, and whether there is a clear definition of the expert layer, it can be classified as follows:

(1) The expert layer includes the gated network model and has a clear definition of the expert layer:

In FIG9A , the global model includes a general layer and an expert layer. The general layer is a network layer that is common to all input data, and may be a feature extraction network such as a convolutional neural network. The expert layer includes a gated network model and N expert models. N is a positive integer. The N expert models may be based on different types of neural networks, such as CNN, RNN, Transformer, MLP, etc.; or based on the same type of neural network, but with different parameter configurations, such as the number of layers, depth, and specific configurations including kernel size.

The gated network model is trained, and the output of the gated network model is used to select the output of the expert model. Selecting an appropriate and matching gated output mechanism can merge and balance the expert's choices. The different expert models here can refer to different network structures, for example, expert model 1 is a convolutional network, expert model 2 is a fully connected network, and so on. However, it is necessary to ensure that the output of each expert network can be merged. Selecting a part of the expert models for prediction based on the output of the gated network model can reduce the amount of calculation and select the most appropriate expert model for different inputs.

In this architecture, the output of the general layer serves as the common input of the expert model and the gating network model.

In FIG9B , the global model includes only the expert layer, but not the general layer. The expert layer includes a gated network model and N expert models. The meanings of the expert model and the gated network model can be referred to in the above description.

In this architecture, the input of the local model serves as the common input of the expert model and the gating network model.

(2) The expert layer does not include the gated network model, and there is a clear definition of the expert layer:

In FIG9C , the expert layer does not include a gated network model, but only includes N expert models. The gated network model is independent of the expert layer. The meanings of the expert model and the gated network model can be referred to the above description. The expert layer has the same input as the gated network model, and the result of the general preprocessing is used as the common input of the expert layer and the gated network model. The result of the general preprocessing can be the output result of the general layer model, or it can be based on the output result of the general layer model.

In FIG. 9D , the expert layer does not include a gated network model, but only includes N expert models. The gated network model is independent of the expert layer. The meanings of the expert model and the gated network model can be referred to the above description. And the input of the expert layer and the gated network model are different. The result of the first preprocessing of the input local data is used as the input of expert layer 1 to expert layer N; the result of the second preprocessing of the input local data is used as the input of the gated network model.

(3) No clear definition of the expert level:

In Figure 9E, there is no clear definition of the expert layer, and the global model includes N networks and a gated network model. The functions and effects of the N networks can refer to the expert model above. The meaning of the gated network model can refer to the above description. Among them, the inputs of network 1 to network N are different from those of the gated network model. The result of the first preprocessing of the input local data is used as the input of network 1; the result of the second preprocessing of the input local data is used as the input of network 2; and so on; the result of the N+1th preprocessing of the input is used as the input of the gated network model.

However, if each child node has a different understanding of how to use the output of the gating network model and how to train the expert model, the performance of the aggregation model of the central node will be degraded.

In response to the above problems, the present application provides a distributed training solution, in which the central node indicates the selection parameters so that all sub-nodes participating in the training will apply the selection parameters to the output results of the gated network model during the training process of the gated network model and the expert model, thereby enabling each sub-node to align the selection of the output results of the gated network model and improve the performance of the aggregation model of the central node.

The distributed training method provided by the embodiment of the present application is described in detail below. It can be understood that the present application uses the central node and the sub-node as an example to illustrate the execution subject of the interactive diagram, but the present application does not limit the execution subject of the interactive diagram. For example, the central node in the method provided by the present application may also be a chip, a chip system, a circuit or a processor applied to the central node, or a logical node, a logical module or software that can implement all or part of the central node; the sub-node in the method provided by the present application may also be a chip, a chip system, a circuit or a processor applied to the sub-node, or a logical node, a logical module or software that can implement all or part of the functions of the sub-node.

As shown in FIG10 , a flow chart of a distributed training method provided in an embodiment of the present application is provided. The method is applied to a distributed training system, which includes a central node and multiple child nodes. The global model of the distributed training system includes N expert models and a gated network model, where N is a positive integer. Exemplarily, the method may include the following steps:

S1001. The subnode sends third information to the central node. Correspondingly, the central node receives the third information.

The child node is any child node participating in the distributed training. This embodiment is described by taking the interaction process between a child node and a central node as an example, and the interaction between other child nodes and the central node refers to the interaction process of this embodiment.

When initially constructing the distributed training system, the child node can report its own capabilities to the central node. Exemplarily, the child node sends third information to the central node, where the third information is used to indicate at least one of the following information: the size of the child node's memory space, the child node's computing power information, whether model training is supported, and the type of model supported for training.

When training starts, the central node will send the initialization model of the central node to the child node, so that the child node can train the initialization model. Therefore, before the training starts, the child node can report the size of the child node's memory space to the central node, so that the central node knows whether the child node's memory space is large enough to store the initialization model of the central node and subsequent training data. The size of the child node's memory space refers to the size of the memory space that the child node can use to store AI/ML models.

Before training begins, the child node can also report the computing power information of the child node to the central node, so that the central node knows whether the child node has strong enough computing power and can timely feedback the model information after training. Among them, the computing power information of the child node refers to the computing power of running the AI/ML model.

Before training begins, the child node can also report to the central node whether it supports model training, so that the central node can determine whether the child node can participate in distributed training and send the global model to it.

Before training begins, the child node can also report the type of model supported for training to the central node so that the central node can determine whether the child node can participate in distributed training. For example, the types of models supported for training by the child node include CNN, RNN, fully connected, random forest models, etc.

In addition, the third information may also include hardware information of the sub-node, including but not limited to the antenna configuration of the sub-node (number of antennas, polarization direction, etc.), number of RF channels, sensor type (position sensor/global positioning system (GPS), motion sensor, etc.) and parameters.

It is understandable that, because it is federated learning, the child nodes use local data for training, so the child nodes do not need to report a series of information related to the actual collected data or involving privacy, such as the amount of data that can be processed.

It is understandable that the central node may also obtain the above information of the child nodes in advance. Therefore, this step is optional and is indicated by a dotted line in the figure.

S1002. The central node sends first information to the child node. Correspondingly, the child node receives the first information.

After receiving the third information reported by each sub-node, the central node selects a sub-node to participate in this round of distributed training according to the third information of each sub-node.

As shown in FIG. 9A to FIG. 9E , the global model includes N expert models (or network 1 to network N as shown in FIG. 9E ), and these N expert models can be based on different types of neural networks, such as CNN, RNN, Transformer, MLP, etc.; or based on the same type of neural network, but corresponding to different parameter configurations, such as the number of layers, depth, and specific configurations including kernel size. Therefore, the output results of the N expert models may be different.

According to experience and simulation results, the more sparsely the output results of the expert model are selected, the more data feature types the expert network can adapt to. The extreme case is that a network only learns one type of feature data.

However, each sub-node participating in the distributed training needs to align the selection of the output results of the N expert models, otherwise the performance of the global model obtained by the central node aggregation will deteriorate. In Figures 9A-9E, the global model includes a gated network model, and the sub-nodes also train the gated network model when training the expert model. The gated network model includes different types of neural networks, or includes the same type of neural network, but corresponds to different parameter configurations, such as the number of layers, depth, and specific configurations including kernel size. Therefore, the gated network model has N outputs, and the N outputs of the gated network model are respectively connected to the outputs of the N expert models. The output results of the gated network model can be used to select the output results of the expert model, and the output results of the gated network model correspond one-to-one to the output results of the expert model connected to it. Therefore, in order to align the selection of the output results of the expert model by each sub-node, the central node sends the first information to the sub-node. Among them, the first information is used to indicate the selection parameter, and the selection parameter is used to select the output result of the gated network model. The sub-node can determine whether to retain the output result of the gated network model or discard the output result of the gated network model according to the selection parameter. If it is determined according to the selection parameters to retain the output result of the gated network model, the output result of the corresponding expert model is also retained; if it is determined according to the selection parameters to discard the output result of the gated network model, the output result of the corresponding expert model is also discarded.

Exemplarily, the output result of the gated network model is a value between [0, 1] after being processed by the set function.

Exemplarily, the selection parameter is a first threshold. For example, the value range of the first threshold may be (0, 1). If the output result of the gated network model is greater than the first threshold, the output result of the gated network model is retained; if the output result of the gated network model is less than the first threshold, the output result of the gated network model is discarded. In particular, if the output result of the gated network model is equal to the first threshold, it can be agreed to retain or discard the output result of the gated network model.

Furthermore, in addition to being used to indicate the selection parameters, the first information can also be used to indicate at least one of the following information: the identification of the N expert models, the competition mode of the N expert models, the training task, and the type of input and output of the global model.

The first information is used to indicate the identifiers of the N expert models, so that the model parameter information of the expert models reported by the subsequent child nodes is based on the output results of which expert models.

For N expert models, the central node can further instruct the child nodes on how to train each expert model. Generally, the central node will instruct the child nodes on the purpose of training the network. For example, channel recovery is to make the recovered channel (as the output of the central node) as close to the true value as possible (that is, to minimize the normalized mean squared error (NMSE)). However, under the architecture of the hybrid expert model, even based on the same task (such as making the recovered channel close to the true value), there may still be at least two of the following methods. The main difference between these two methods is the competitive mode of the N expert models, where the competitive mode can also be called the usage mode, cooperation mode, collaboration mode, and whether to collaborate, etc.:

Mode 1: N expert models collaborate with each other:
loss=||A-∑ _i p _i o _i || ² .

Mode 2: N expert models compete with each other:
loss=∑ _i p _i ||Ao _i || ² .

Among them, in the above two formulas, A represents the true value (which can represent a single sample), _pi represents the weight of the ith expert model (determined based on the output of the gating network model), and _oi represents the output result of the ith expert model. It can be seen that in the collaborative mode, multiple expert models cooperate to approximate the true value; while in the competitive mode, each expert model has a separate loss, which realizes a separate judgment without relying on the results of other expert models. Even if the same training node trains the hybrid expert model based on different competitive methods, the gating weight and expert weight obtained will be different. Therefore, it is hoped that all sub-nodes participating in the training will at least use/train the expert model in the same way.

Therefore, the central node may carry the competition mode of the N experts in the first information, where the competition mode indicates whether the N expert models are cooperating or competing.

The central node also indicates the training task of the distributed training, and the training task can be understood as the function of the global model, that is, what the global model can be used for training. Exemplarily, the training task can be beam prediction, channel recovery, channel prediction, etc.

In the beam management scenario, the input of the global model may be, for example, channel quality, RSRP of the beam, etc.; the output of the global model may be, for example, the optimal beam index, etc.

In addition, the central node can also send model parameter information of the global model to the child nodes.

S1003. The subnode inputs sample data to N expert models and the gated network model to obtain N first output results of the N expert models and N second output results of the gated network model, and the N second output results correspond one-to-one to the N first output results.

When training a child node, it is necessary to input sample data into N expert models and gated network models. Therefore, before executing this step, the child node collects different types of sample data in different application scenarios.

For example, in a beam management scenario, the subnodes in the federated learning architecture can be network devices, while the central node can be an independent federated learning management node; or the subnodes can be terminal devices, while the central node can be a network device that functions as a central node. Assuming that the global model to be trained is an AI/ML model that takes the estimated channel measurement value or the received signal itself as input and the optimal beam index as output, then the subnode is responsible for collecting the channel measurement value or received signal as the model input and the label used for training the model, that is, the optimal beam index, during the data collection stage. All possible beams (codebook-based synchronization signal blocks (SSB) or channel state information-reference signal (CSI-RS) beams) can be sent to the terminal device one by one through the network device, and the terminal device selects the beam direction index with the best performance (the best can refer to the beam with the largest physical layer-reference signal receiving power (layer1-reference signal receiving power, L1-RSRP) or signal-to-noise ratio (SNR) measurement value among all SSB/CSI-RS beams) as a label.

Before model training, the central node can also configure downlink resources for the subnodes to send the initial global model information of the central node. The downlink resources can be control channel resources, such as PDCCH resources; or data channel resources, such as PDSCH resources. Exemplarily, the downlink resources include frequency domain resource block number, starting position, subband number, subband bandwidth, frequency hopping parameters, modulation and coding scheme (MCS) and other parameters.

The global model can be sent down by the central node in a broadcast or multicast manner. For example, in a single-cell federated learning architecture where the central node is a network device and the sub-nodes are terminal devices, the global model can be sent down in a broadcast manner. Due to the characteristics of broadcast, sub-nodes that are not involved in the federated learning can also receive the broadcast information; in a multi-cell federated learning architecture where a network device with a federated learning management function is used as the central node and other network devices are used as sub-nodes, the central node can also send the global model to each sub-node in a broadcast manner. Similarly, other sub-nodes that are not involved in the federated learning can also receive the broadcast information; multicast can also be used for sub-nodes participating in the federated learning. Sub-nodes associated with the same central node are grouped together, have the same group number, and are configured with the same downlink resources. In multicast mode, sub-nodes that do not participate in the federated learning will not receive the multicast information.

The central node can also configure uplink resources for the subnodes to report local models for the subnodes to report models/gradients/weights. Another federated learning management node can also configure uplink resources for the central node and subnodes for the subnodes to report local model information and necessary signaling. Similar to the downlink resource configuration, the uplink resources can be control channel resources, such as PUCCH resources, or data channel resources, such as PUSCH resources.

After the child node collects a certain amount of sample data, it inputs the sample data into the N expert models and the gated network model to obtain the N first output results of the N expert models and the N second output results of the gated network model. Since the N outputs of the gated network model are connected to the outputs of the N expert models, the N second output results correspond to the N first output results one by one.

S1004. The child node selects, for each first output result among the N first output results and each second output result among the N second output results, a first output result corresponding to the second output result that satisfies the selection parameter.

For the second output result of each gated network model, the child node determines whether to retain or discard the second output result according to the selection parameter. For example, the selection parameter is a first threshold, and the child node determines whether the second output result is greater than or equal to the first threshold. If so, the second output result is retained; otherwise, the second output result is discarded.

After the at least one second output result to be retained is determined, at least one first output result corresponding to the at least one second output result to be retained is also retained or selected.

S1005. The subnode obtains model parameter information of at least one selected expert model and model parameter information of the gated network model based on the selected at least one first output result and the true value information.

After the subnode selects at least one first output result, the subnode obtains model parameter information of the selected at least one expert model and model parameter information of the gated network model based on the selected at least one first output result and the true value information.

The model parameter information of the expert model includes the weight, gradient, gradient change, etc. of the expert model. The model parameter information of the gated network model includes the weight, gradient, gradient change, etc. of the gated network model.

Exemplarily, the child node obtains the model parameter information of the selected at least one expert model and the model parameter information of the gated network model based on the selected at least one first output result and the true value information, and there are two possible implementation methods:

One possible implementation is that the child node obtains the model parameter information of the selected at least one expert model and the model parameter information of the gated network model based on the average value and true value information of the selected at least one first output result. For example, if the first threshold is 0.5, the child node can retain the second output results of the gated network model whose values are greater than 0.5, and select at least one first output result of the expert model corresponding to the retained at least one second output result. The child node multiplies each of the selected at least one first output results by 0.5, sums and averages their products, and obtains the average value of the selected at least one first output result.

Another possible implementation is that the child node weights and averages the at least one selected first output result based on the at least one second output result corresponding to the at least one selected first output result, obtains the weighted average value of the at least one selected first output result, and obtains the model parameter information of the at least one selected expert model and the model parameter information of the gated network model based on the weighted average value and true value information of the at least one selected first output result. For example, the first threshold is 0.5, the child node can retain the second output result of the gated network model whose value is greater than 0.5, and select at least one first output result of the expert model corresponding to the at least one retained second output result. The child node multiplies each first output result of the at least one selected first output result by its corresponding second output result (for example, some second output results are 0.6, and some second output results are 0.8), and sums and averages their products to obtain the weighted average value of the at least one selected first output result.

S1006. The subnode sends the second information to the central node. Correspondingly, the central node receives the second information.

After the child node obtains the model parameter information of at least one selected expert model and the model parameter information of the gated network model, it sends second information to the central node, wherein the second information is used to indicate the model parameter information of the gated network model and the model parameter information of at least one selected expert model from the N expert models.

Furthermore, the second information is also used to indicate the identifier of the selected at least one expert model. The subnode carries the indication information of the identifier of the selected at least one expert model in the second information, so that the central node performs model aggregation based on the indication information.

S1007. The central node updates the global model and the selection parameters based on the plurality of second information received from the plurality of child nodes.

The central node receives multiple pieces of second information from multiple child nodes respectively, and can update the global model based on the multiple pieces of second information.

In some scenarios, the central node can further update the selection parameters based on the updated global model. For example, during initial training, the first threshold configured by the central node is relatively large, and the first threshold is updated based on information such as training feedback from the child nodes. The first threshold can also be obtained based on neural network learning.

S1008. The central node sends the fourth information. Correspondingly, the subnode receives the fourth information.

If the central node updates the selection parameters, the fourth information may be sent to each child node, wherein the fourth information is used to indicate the updated selection parameters, so that the fourth information selects the output result of the gated network model based on the updated selection parameters during the next round of training.

The distributed training system may execute the above steps S1002 to S1008 multiple times until the global model converges.

In one example, the global model includes a general layer Net_com, three expert models Net1/2/3, and a fully connected network Net_FC. The central node (e.g., a base station) sends model parameter information of the global model to each child node, including Net_com, Net1/2/3, and Net_FC. Their corresponding neural network gradient/weight data are represented as W_com, W_1/W_2/W_3, and W_FC.

Assume that only two child nodes participate in this round of federated learning: UE1 and UE2. Their respective training processes are the same. Input local sample data into the model and perform the following training:

ⅰ) The outputs after the general layer and Net_1/2/3 are W_1_out, W_2_out, W_3_out, and the output of gated Net_FC is W_FC (a vector of three [0,1] elements). After fusion, the final output is:
Out＝W_FC(1)*W_1_out+W_FC(2)*W_2_out+W_FC(3)*W_3_out;

ii) Calculate loss based on, for example, the loss function calculation formula loss = ||Out–A||^2, where A is the true value/label;

ⅲ) After completing the training, the following weights are obtained:
UE1: W_com(UE1), W_1(UE1)/W_2(UE1)/W_3(UE1) and W_FC(UE1);
UE2: W_com(UE2), W_1(UE2)/W_2(UE2)/W_3(UE2) and W_FC(UE2).

The sub-nodes (UE1, UE2) feed back the above weights (which may also be gradients or gradient changes) to the central node (base station).

After the central node obtains the model parameter information from the two child nodes, it corresponds to the same global model and starts to fuse the two sets of parameters. Here, we take the averaging method as an example to obtain:

ⅰ) Fusion results of general layer

ii) Fusion results of the three expert models

ⅲ) Fusion results of the gated network model

After completing this round of federated learning, the central node sends new model parameter information to the child nodes for the next round of training.

As shown in Figure 11, it is a schematic diagram of the optimal beam training of an example embodiment of the present application. Taking the high-frequency beam management problem as an example, when the network device side performs beam scanning of the synchronization signal block or CSI-RS based on the codebook, the channels between users (UE1~UEN) at different locations (may also include training nodes with similar data characteristics to these UEs) and the network device are different. The network device sends a unified global model to users at different locations. Users at different locations measure the received SSB or CSI-RS beam, for example, measure L1-RSRP, and feedback the beam identifier corresponding to the maximum RSRP value. AI/ML can be used to train a model, using, for example, a plurality of received SSB/CSI-RS signals (part or all of them), or the strength (RSRP) of a plurality of received SSB/CSI-RS signals (part or all of them) or the estimated channel as input to infer the optimal beam ID and feed it back to the network device.

The global model includes a gated network model and N expert models, and may also include other layers (i.e., the above-mentioned general layers, which may be trained or not), and the other layers are connected to the gated network model and the N expert models. The gated network model has N outputs, and each output is connected to the output of an expert model. During the training process of each user, each user inputs the measured RSRP into other layers, and the training output results of other layers are used as the common input of the gated network model and the N expert models. Each user trains the gated network model and the N expert models based on the input. Before training, the network device sends selection parameters to each user, and each user determines whether to retain the N second output results of the gated network model according to the selection parameters, and obtains the first output result of at least one expert model corresponding to the at least one second output result retained according to the retained at least one second output result. Then, each user obtains the model parameter information of the selected at least one expert model and the model parameter information of the gated network model based on the selected at least one first output result and the true value information, and sends the model parameter information of the selected at least one expert model and the model parameter information of the gated network model to the network device. Furthermore, each user obtains an optimal beam identifier based on the selected at least one first output result and the true value information.

Optionally, the central node is a third-party device that performs the aforementioned actions related to the central node. For example, the above steps S1001-S1002, S1006-S1008 are all performed by a third-party device.

Optionally, the subnode is a third-party device that performs the aforementioned subnode-related actions. For example, the above steps S1001-S1008 are all performed by a third-party device.

Optionally, the central node is a network device. In this case, the network device can complete the training of the model. For example, the above steps S1001-S1002 and S1006-S1008 are all performed by the network device.

Optionally, the child node is a terminal device. In this case, the terminal device can complete the training of the model. For example, the above steps S1001-S1008 are all performed by the terminal device.

Optionally, the central node includes a network device and a third-party device. In one example, the above step S1007 can be performed by a third-party device, such as an OTT, or a cloud server, and one or more of the above steps S1001-S1002, S1006, and S1008 can be performed by the network device. In addition, the network device and the third-party device can also communicate with each other to transmit the content transmitted by one or more of the above steps S1001-S1002, S1006-S1008.

Optionally, the subnode includes a terminal device and a third-party device. In one example, one or more of the above steps S1003-S1005 may also be performed by a third-party device, such as an OTT, or a cloud server, and one or more of the above steps S1001-S1002, S1006, and S1008 may be performed by the terminal device. In addition, the terminal device and the third-party device may also communicate with each other to transmit the content transmitted by one or more of the above steps S1001-S1008.

According to a distributed training method provided in an embodiment of the present application, the central node indicates a selection parameter so that all sub-nodes participating in the training will apply the selection parameter to the output result of the gated network model during the training process of the gated network model and the expert model, thereby enabling each sub-node to align the selection of the output result of the gated network model and improve the performance of the aggregation model of the central node.

In this application, "sending information to... (e.g., a child node)" or the related illustrations in the accompanying drawings can be understood as the destination end of the information is a child node. It can include sending information to a child node directly or indirectly. "Receiving information from... (e.g., a child node)" or "receiving information from... (e.g., a child node)", or the related illustrations in the accompanying drawings can be understood as the source end of the information is a child node, which can include receiving information from a child node directly or indirectly. The information may be processed as necessary between the source end and the destination end of the information transmission, such as format changes, etc., but the destination end can understand the valid information from the source end. Similar expressions in this application can be understood similarly and will not be repeated here.

The above mainly introduces the scheme provided by the embodiment of the present application from the perspective of interaction between various nodes. Accordingly, the embodiment of the present application also provides a distributed training device, which is used to implement the above various methods. The distributed training device can be a central node in the above method embodiment, or a component that can be used for the central node; or, the distributed training device can be a sub-node in the above method embodiment, or a component that can be used for the sub-node. It can be understood that in order to achieve the above functions, the distributed training device includes a hardware structure and/or software module corresponding to each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiment disclosed in this article, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

The embodiment of the present application can divide the functional modules of the distributed training device according to the above method embodiment. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing unit. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical functional division. There may be other division methods in actual implementation.

Based on the same concept of the above distributed training method, the present application also provides the following distributed training device:

As shown in FIG. 12 , it is a schematic diagram of the structure of a distributed training device provided in an embodiment of the present application. The distributed training device 1200 includes a transceiver unit 1201 and a processing unit 1202; wherein:

When the distributed training device is used to implement the functions of the subnode in the above method embodiment, the transceiver unit 1201 is used to perform one or more of the operations of the subnode in steps S1001, S1002, S1006 and S1008 of the embodiment shown in Figure 10, and the processing unit 1202 is used to perform one or more of steps S1003-S1005 of the embodiment shown in Figure 10. Optionally, the distributed training device can be a terminal device, or a third-party device such as an OTT or cloud server, or a system composed of a terminal device and a third-party device.

When the distributed training device is used to implement the function of the central node in the above method embodiment, the transceiver unit 1201 is used to perform one or more of the operations of the central node in steps S1001, S1002, S1006 and S1008 of the embodiment shown in Figure 10, and the processing unit 1202 is used to perform step S1007 of the embodiment shown in Figure 10. Optionally, the distributed training device can be a network device, or a third-party device such as an OTT or cloud server, or a system consisting of a network device and a third-party device.

For the specific implementation of the above-mentioned transceiver unit 1201 and the processing unit 1202, reference may be made to the description in the above-mentioned method embodiment.

In addition, it should be noted that the aforementioned transceiver unit and/or processing unit can be implemented through a virtual module, for example, the processing unit can be implemented through a software function unit or a virtual device, and the transceiver unit can be implemented through a software function or a virtual device. Alternatively, the processing unit or the transceiver unit can also be implemented through a physical circuit, for example, if the device is implemented using a chip/chip circuit, the transceiver unit can be an input-output circuit and/or a communication interface, performing input operations (corresponding to the aforementioned receiving operations) and output operations (corresponding to the aforementioned sending operations); the processing unit is a processing circuit, such as an integrated processor or microprocessor or integrated circuit.

The division of modules in this application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each example of this application may be integrated into one processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.

As shown in Figure 13, it is a schematic diagram of the structure of another distributed training device provided in an embodiment of the present application. The distributed training device 1300 includes one or more processing circuits 1301 (one processing circuit is illustrated in the figure). Optionally, the distributed training device 1300 may also include a memory 1303 (indicated by a dotted line in the figure). The memory 1303 is used to store instructions executed by the processing circuit 1301, or to store input data required for the processing circuit 1301 to run instructions, or to store data generated after the processing circuit 1301 runs instructions. Optionally, the distributed training device 1300 may also include an interface circuit 1302 (indicated by a dotted line in the figure), and the processing circuit 1301 and the interface circuit 1302 are coupled to each other. It can be understood that the interface circuit 1302 can be a transceiver or an input-output interface.

The processing circuit may be a processor or a circuit in a processor used for processing.

When the distributed training device is used to implement the functions of the sub-node in the above method embodiment, the interface circuit 1302 is used to execute one or more of the operations of the sub-node in steps S1001, S1002, S1006 and S1008 of the embodiment shown in Figure 10, and the processing circuit 1301 is used to execute one or more of steps S1003-S1005 of the embodiment shown in Figure 10.

When the distributed training device is used to implement the function of the central node in the above method embodiment, the interface circuit 1302 is used to execute one or more of the operations of the central node in steps S1001, S1002, S1006 and S1008 of the embodiment shown in Figure 10, and the processing circuit 1301 is used to execute step S1007 of the embodiment shown in Figure 10.

When the above-mentioned distributed training device is a chip applied to the central node, the chip implements the function of the central node in the above-mentioned method embodiment. The chip receives information from other modules in the central node, and the information is sent by the subnode to the central node; or, the chip sends information to other modules in the central node, and the information is sent by the central node to the subnode. When the central node is a network device, the module of the central node here can be the baseband chip of the central node, or it can be a CU, DU or other module, or it can be a device under the open radio access network (open radio access network, O-RAN) architecture, such as an open CU, open DU and other devices. When the central node is a third-party device, the module of the subnode here can be a processing chip of a third-party device. Among them, the processing chip can be used to implement AI training.

When the above-mentioned distributed training device is a chip applied to a subnode, the chip implements the functions of the subnode in the above-mentioned method embodiment. The chip receives information from other modules in the subnode, and the information is sent from the central node to the subnode; or, the chip sends information to other modules in the subnode, and the information is sent from the subnode to the central node. When the subnode is a terminal device, the module of the subnode here can be the baseband chip of the subnode, or, the baseband chip and the processing chip. Among them, the processing chip can be used to implement AI training. When the subnode is a third-party device, the module of the subnode here can be the processing chip of the third-party device. Among them, the processing chip can be used to implement AI training.

An embodiment of the present application further provides a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed, the method in the above embodiment is implemented.

The embodiments of the present application also provide a computer program product including instructions, which, when executed on a computer, enables the computer to execute the method in the above embodiments.

An embodiment of the present application also provides a distributed training system, including the above-mentioned distributed training device.

The embodiment of the present application also provides a circuit, which is coupled to a memory and is used to execute the method shown in the above embodiment. The circuit may include a chip circuit.

Optionally, the embodiment of the present application further provides a chip system, including: at least one processor and an interface, the at least one processor is coupled to a memory via the interface, and when the at least one processor runs a computer program or instruction in the memory, the chip system executes a method in any of the above method embodiments. Optionally, the chip system may be composed of a chip, or may include a chip and other discrete devices, which is not specifically limited in the embodiment of the present application.

The memory in the present application may also be a circuit or any other device capable of realizing a storage function, for storing program instructions and/or data. The memory is any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer, but is not limited thereto. For example, the memory may be a non-volatile memory, such as a digital versatile disc (DVD), a hard disk drive (HDD), or a solid-state drive (SSD), etc., or a volatile memory (volatile memory), such as a random-access memory (RAM).

The terms "including" and "having" and any variations thereof mentioned in the following description of this application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

It should be understood that in the description of the present application, unless otherwise specified, "/" indicates that the objects associated before and after are in an "or" relationship, for example, A/B can represent A or B; wherein A and B can be singular or plural. Also, in the description of the present application, unless otherwise specified, "multiple" refers to two or more than two. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a-b, a-c, b-c, or a-b-c, wherein a, b, c can be single or multiple. In addition, in order to facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, the words "first", "second", etc. are used to distinguish the same items or similar items with substantially the same functions and effects. Those skilled in the art can understand that the words "first", "second", etc. do not limit the quantity and execution order, and the words "first", "second", etc. do not limit them to be necessarily different. Meanwhile, in the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete manner for ease of understanding.

In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using a software program, it can 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 program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.

Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by viewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in a claim. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

It is understood that the various numbers involved in the embodiments of the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application. The size of the sequence number of the above-mentioned processes does not mean the order of execution, and the execution order of each process should be determined by its function and internal logic.

In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

The components in the device of the embodiment of the present application can be merged, divided and deleted according to actual needs. Those skilled in the art can combine or combine the different embodiments and features of the different embodiments described in this specification.

In the present application, under the premise of no logical contradiction, the examples may reference each other, for example, the methods and/or terms between method embodiments may reference each other, for example, the functions and/or terms between device embodiments may reference each other, for example, the functions and/or terms between device examples and method examples may reference each other.

Claims (21)

A distributed training method, characterized in that the method is applied to a distributed training system, the global model of the distributed training system includes N expert models and a gated network model, N is a positive integer, and the method includes: receiving first information, where the first information is used to indicate a selection parameter, and the selection parameter is used to select an output result of the gated network model; Sending second information, where the second information is used to indicate model parameter information of the gated network model and model parameter information of at least one expert model selected from the N expert models, where the at least one selected expert model is obtained based on an output result of the gated network model. The method according to claim 1, characterized in that the method further comprises: Inputting sample data into the N expert models and the gated network model to obtain N first output results of the N expert models and N second output results of the gated network model, wherein the N second output results correspond one-to-one to the N first output results respectively; For each first output result of the N first output results and each second output result of the N second output results, for the second output result satisfying the selection parameter, selecting a first output result corresponding to the second output result; Based on the selected at least one first output result and the true value information, model parameter information of the selected at least one expert model and model parameter information of the gated network model are obtained. The method according to claim 1 or 2, characterized in that the selection parameter is a first threshold. The method according to any one of claims 1 to 3 is characterized in that the first information is also used to indicate at least one of the following information: the identification of the N expert models, the competition mode of the N expert models, the training task, and the type of input and output of the global model. The method according to any one of claims 1 to 4, characterized in that the method further comprises: Send third information, where the third information is used to indicate at least one of the following information: the memory space size of the child node, the computing power information of the child node, whether model training is supported, and the type of model supported for training. The method according to any one of claims 1 to 5, characterized in that the second information is also used to indicate an identifier of the at least one selected expert model. The method according to any one of claims 1 to 4, characterized in that the acquiring, based on the selected at least one first output result and the true value information, the model parameter information of the selected at least one expert model and the model parameter information of the gated network model comprises any one of the following operations: Based on the average value and true value information of the at least one selected first output result, obtaining model parameter information of the at least one selected expert model and model parameter information of the gated network model; or Based on at least one second output result corresponding to the at least one selected first output result, the at least one selected first output result is weighted and averaged to obtain a weighted average value of the at least one selected first output result, and based on the weighted average value and true value information of the at least one selected first output result, model parameter information of the at least one selected expert model and model parameter information of the gated network model are obtained. The method according to any one of claims 1 to 7, characterized in that the method further comprises: Fourth information is received, where the fourth information is used to indicate updated selection parameters, where the updated selection parameters are obtained based on the second information. A distributed training method, characterized in that the method is applied to a distributed training system, the global model of the distributed training system includes N expert models and a gated network model, the N is a positive integer, and the method includes: Sending first information to multiple child nodes, where the first information is used to indicate a selection parameter, and the selection parameter is used to select an output result of the gated network model; A plurality of second information are received respectively, each of the plurality of second information is used to indicate model parameter information of the gated network model and model parameter information of at least one expert model selected from the N expert models, wherein the at least one selected expert model is obtained based on an output result of the gated network model. The method according to claim 9, characterized in that the selection parameter is a first threshold. The method according to claim 9 or 10 is characterized in that the first information is also used to indicate at least one of the following information: the identification of the N expert models, the competition mode of the N expert models, the training task, and the type of input and output of the global model. The method according to any one of claims 9 to 11, characterized in that the method further comprises: Receive third information, where the third information is used to indicate at least one of the following information: the memory space size of the child node, the computing power information of the child node, whether model training is supported, and the type of model supported for training. The method according to any one of claims 9 to 12, characterized in that the second information is also used to indicate an identification of the at least one selected expert model. The method according to any one of claims 9 to 13, characterized in that the method further comprises: The global model is updated based on the plurality of second information. The method according to any one of claims 9 to 14, characterized in that the method further comprises: updating the selection parameter based on the plurality of second information; Send fourth information, where the fourth information is used to indicate the updated selection parameter. A distributed training device, characterized in that it comprises a unit for executing the method as described in any one of claims 1-8, or comprises a unit for executing the method as described in any one of claims 9-15. A distributed training system, characterized in that the system comprises a first distributed training device and a second distributed training device, the first distributed training device is used to implement the method as described in any one of claims 1-8, and the second distributed training device is used to implement the method as described in any one of claims 9-15. A distributed training device, characterized in that it includes a processing circuit and an interface circuit, wherein the interface circuit is used to receive signals from other distributed training devices outside the distributed training device and transmit them to the processing circuit, or to send signals from the processing circuit to other distributed training devices outside the distributed training device, and the processing circuit is used to implement the method described in any one of claims 1 to 8, or to implement the method described in any one of claims 9 to 15 through a logic circuit or execution code instructions. The distributed training device according to claim 18 is characterized in that the distributed training device is a chip. A chip module, characterized in that it includes a transceiver component and a chip, wherein the chip is used to execute the method as described in any one of claims 1 to 8, or execute the method as described in any one of claims 9 to 15. A computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented, or the method according to any one of claims 9 to 15 is implemented.