Communication method and related device

Abstract

A communication method and a related device are used for improving model performance. In the method, a first communication device may determine a first distillation loss of a first output and a second output from received first information, where the first output may be obtained by processing based on a first input by a first AI model deployed in the first communication device. Thereafter, the first communication device may update the first AI model based on the first distillation loss. Through the mode, the computing power of the communication node can be applied to processing of the AI model, and meanwhile, the first communication device can obtain knowledge of the second AI model in a mode of performing model updating through distillation loss, so that the model performance is improved.

Description

Technical Field

[0001] The present application relates to the field of communications, and in particular to a communication method and related devices. Background Art

[0002] Wireless communication can be the transmission communication between two or more communication nodes without propagating through conductors or cables. The communication nodes generally include network devices and terminal devices.

[0003] Currently, in wireless communication systems, communication nodes generally possess both signal transceiver capabilities and computing capabilities. For example, network devices with computing capabilities primarily provide computing power to support signal transceiver capabilities (e.g., processing both sending and receiving signals), enabling communication between the network device and other communication nodes.

[0004] However, in communication networks, communication nodes may have excess computing power beyond just supporting the aforementioned communication tasks. Therefore, how to utilize this computing power is a pressing technical issue. Summary of the Invention

[0005] The present application provides a communication method and related devices for improving model performance.

[0006] In a first aspect, the present application provides a communication method, which is applied to a first communication device. The first communication device may be a terminal device, or the first communication device may be a component of the terminal device (such as a processor, chip, or chip system), or the first communication device may be a logic module or software that can implement all or part of the terminal device's functions. In this method, the first communication device receives first information, and the first information is used to determine a first distillation loss for a first output and a second output; wherein the first output is obtained by processing a first artificial intelligence (AI) model based on a first input, and the second output is obtained by processing a second AI model based on a second input; and the first communication device updates the first AI model based on the first distillation loss.

[0007] Based on the above scheme, the first communication device can determine the first distillation loss of the first output and the second output through the received first information, wherein the first output can be obtained by processing the first AI model deployed on the first communication device based on the first input, and the second output can be obtained by processing the second AI model deployed on other communication devices (such as the second communication device) based on the second input. Thereafter, the first communication device can update the first AI model based on the first distillation loss. In this way, while the computing power of the communication node can be applied to the processing of the AI ​​model, the first communication device can also obtain the knowledge of the second AI model by updating the model through the distillation loss to improve the model performance.

[0008] In addition, data privacy is relatively strong in wireless communication scenarios. In the above process, different communication devices can update the AI ​​model through knowledge distillation without transmitting their respective input data. While improving model performance, it can also protect the privacy of the data of each communication device to enhance user experience.

[0009] In this application, AI model can be replaced by other terms, such as neural network, neural network model, AI neural network model, machine learning model, or AI processing model.

[0010] It should be noted that the first AI model can implement a model update process through knowledge distillation based on the second output provided by the second AI model. In this process, the first AI model can obtain the knowledge of the second AI model. Accordingly, the second AI model can be a larger or more complex model (such as a large model), and the first AI model can be a smaller or simpler model. Thus, in the case that the first communication device may not have strong AI capabilities, the model knowledge of the larger or more complex model can be deployed in the first AI model in the first communication device through knowledge distillation, which can improve the efficiency of model deployment.

[0011] Optionally, the first AI model and the second AI model can be described in a variety of ways. For example, the first AI model can be called a student model, and the second AI model can be called a teacher model; or, the first AI model can be called a knowledge learning model, and the second AI model can be called a knowledge transfer model.

[0012] In a possible implementation manner of the first aspect, the first information is used to indicate the second output.

[0013] Based on the above scheme, the first communication device can obtain a second output through the received first information. Accordingly, the first communication device can determine the first distillation loss based on the second output and the first output of the first AI model, so that the first communication device updates the first AI model based on the first distillation loss.

[0014] In a possible implementation manner of the first aspect, the method further includes: the first communication device receiving indication information indicating that the characteristic dimension of the second output is the same as the characteristic dimension of the first output.

[0015] Based on the above solution, the first AI model and the second AI model can be deployed in different AI models, respectively. Accordingly, the feature dimensions of the first output obtained based on the first AI model and the second output obtained based on the second AI model may be different or the same. In the above process, the first communication device can determine that the feature dimension of the second output is the same as the feature dimension of the first output through the above indication information, so that the first communication device can determine the first distillation loss based on the second output and the first output obtained by the first information without performing feature dimension processing, which can reduce implementation complexity.

[0016] In a possible implementation of the first aspect, the first distillation loss is determined by a first processing result obtained by performing feature dimension processing on the first output and the second output, wherein the first processing result and the feature dimension of the first output are the same.

[0017] Based on the above solution, the first AI model and the second AI model can be deployed in different AI models respectively. Accordingly, the feature dimensions of the first output obtained based on the first AI model and the second output obtained based on the second AI model may be different or the same. In the above process, the first communication device can obtain the second output by receiving the first information. Accordingly, the first communication device can perform feature dimension processing based on the second output to obtain a first processing result, so that the first communication device can determine the first distillation loss based on the first processing result and the first output of the same feature dimension, thereby avoiding the situation where the distillation loss cannot be determined due to different feature dimensions.

[0018] Optionally, the feature dimension processing can be performed by an internal processing module of the first AI model, i.e., the first AI model has the feature dimension processing function. Alternatively, the feature dimension processing can be performed independently of the processing of an external processing module of the first AI model.

[0019] In a possible implementation manner of the first aspect, the first information is used to indicate the first distillation loss; the method further includes: the first communication device sending second information, where the second information is used to indicate the first output.

[0020] Based on the above solution, the first communication device may also transmit second information indicating the first output, enabling a recipient of the second information (e.g., the second communication device) to determine the first distillation loss based on the second information. Accordingly, the first information transmitted by the recipient to the first communication device may indicate the first distillation loss, allowing the first communication device to update the first AI model based on the first distillation loss, thereby reducing the implementation complexity of the first communication device.

[0021] In a possible implementation manner of the first aspect, the feature dimension of the second output is the same as the feature dimension of the first output.

[0022] Based on the above scheme, the characteristic dimension of the first output indicated by the second information sent by the first communication device can be the same as the characteristic dimension of the second output. In this way, the implementation complexity of the recipient of the second information (such as the second communication device) can be reduced.

[0023] In a possible implementation manner of the first aspect, the method further includes: the first communication device sending third information, where the third information is used to determine the second input; wherein the third information is associated with the first input.

[0024] Based on the above solution, the first communication device may also transmit third information for determining the second input, and the third information is associated with the first input of the first AI model. In this way, the first and second AI models can determine corresponding first and second outputs based on the associated first and second inputs, thereby improving the performance of knowledge distillation based on the first and second outputs.

[0025] In a possible implementation manner of the first aspect, the first input is different from the second input.

[0026] Based on the above solution, the first input to the first AI model in the first communication device and the second input to the second AI model in the second communication device can be different. For example, the second input can contain more information than the first input. In this way, the second output obtained by the second communication device based on the second input can obtain more knowledge, thereby improving the performance of knowledge distillation based on the second output.

[0027] In a possible implementation of the first aspect, before the first communication device receives the first information, the method further includes: the first communication device receives configuration information of the first AI model; wherein the configuration information of the first AI model includes at least one of the following: a task identifier corresponding to the first AI model, an identifier of the configuration information of the first AI model, a network structure parameter of the first AI model, a training parameter of the first AI model, and a distillation parameter of the first AI model.

[0028] Based on the above solution, the first communication device and the second communication device can implement the deployment of the first AI model in the first communication device through the interaction of the configuration information of the first AI model. In addition, the configuration information of the first AI model can include at least one of the above items to enhance the flexibility of the solution implementation.

[0029] In addition, when the configuration information of the first AI model includes the distillation parameters of the first AI model, the first communication device is enabled to implement a subsequent knowledge distillation process based on the distillation parameters of the first AI model.

[0030] Optionally, the configuration information of the first AI model may be pre-configured.

[0031] In a possible implementation of the first aspect, the method further includes: the first communication device receives fourth information, the fourth information is used to indicate one or more distillation loss configuration information corresponding to the second AI model; the first communication device sends fifth information, the fifth information is used to indicate the first distillation loss configuration information among the one or more distillation loss parameters, and the first distillation loss configuration information is used to determine the first distillation loss.

[0032] Based on the above scheme, the first communication device and the second communication device can realize the process of selecting or indicating the distillation loss configuration information through the interaction process of the fourth information and the fifth information, so that the first communication device can realize the subsequent knowledge distillation process based on the distillation loss configuration adapted to the first communication device (and / or the first AI model).

[0033] Optionally, in the one or more distillation loss configuration information, each distillation loss configuration information includes at least one of the following: an identifier of each distillation loss configuration information, an indication information that the parameter type of the distillation loss is a hard label, a soft label, or intermediate features of some or all layers of the AI ​​model, or a layer index in the second AI model for outputting the second output.

[0034] In a possible implementation of the first aspect, the method further includes: the first communication device sends sixth information, where the sixth information is used to indicate AI capabilities and / or AI tasks; wherein the sixth information is used to determine configuration information of the first AI model and / or the sixth information is used to determine the second AI model.

[0035] Based on the above solution, the first communication device can send sixth information indicating AI capabilities and / or AI tasks, so that the recipient of the sixth information (e.g., the second communication device) can determine the first AI model and / or second AI model based on the sixth information. In this way, the first and second AI models used in the subsequent knowledge distillation process can be adapted to the AI ​​capabilities and / or AI tasks of the first communication device.

[0036] In a possible implementation of the first aspect, the first communication device updates the first AI model based on the first distillation loss, including: the first communication device updates the first AI model based on the first distillation loss and the task loss, wherein the task loss is determined based on the labels corresponding to the first output and the first input.

[0037] Based on the above scheme, when the first communication device performs the knowledge distillation process, the basis for updating the first AI model may include task loss in addition to the first distillation loss, which can further improve the performance of the updated first AI model obtained based on the knowledge distillation process.

[0038] The second aspect of the present application provides a communication method, which is applied to a second communication device, which can be a communication device (such as a terminal device or a network device), or the second communication device can be a partial component in the communication device (such as a processor, a chip or a chip system, etc.), or the second communication device can also be a logic module or software that can realize all or part of the functions of the communication device. In this method, the second communication device determines first information, and the first information is used to determine a first distillation loss of a first output and a second output; wherein the first output is obtained by processing the first AI model based on the first input, and the second output is obtained by processing the second AI model based on the second input; the first AI model is a student model, and the second AI model is a teacher model; the first distillation loss is used to update the first AI model; and the second communication device sends the first information to the terminal device.

[0039] Based on the above scheme, after the second communication device sends the first information to the first communication device (the first communication device is a terminal device or a partial component, logic module or software in the terminal device), the first communication device can determine the first distillation loss of the first output and the second output through the received first information, wherein the first output can be obtained by processing based on the first input by the first AI model deployed on the first communication device, and the second output can be obtained by processing based on the second input by the second AI model deployed on other communication devices (such as the second communication device). Thereafter, the first communication device can update the first AI model based on the first distillation loss. In this way, the computing power of the communication node can be applied to the processing of the AI ​​model, and the first communication device can also obtain the knowledge of the second AI model by updating the model through the distillation loss to improve the model performance.

[0040] In addition, data privacy is relatively strong in wireless communication scenarios. In the above process, different communication devices can update the AI ​​model through knowledge distillation without transmitting their respective input data. While improving model performance, it can also protect the privacy of the data of each communication device to enhance user experience.

[0041] In a possible implementation manner of the second aspect, the first information is used to indicate the second output.

[0042] Based on the above scheme, the first communication device can obtain a second output through the received first information. Accordingly, the first communication device can determine the first distillation loss based on the second output and the first output of the first AI model, so that the first communication device updates the first AI model based on the first distillation loss.

[0043] In a possible implementation manner of the second aspect, the method further includes: the second communication device sending indication information indicating that the characteristic dimension of the second output is the same as the characteristic dimension of the first output.

[0044] Based on the above solution, the first AI model and the second AI model can be deployed in different AI models, respectively. Accordingly, the feature dimensions of the first output obtained based on the first AI model and the second output obtained based on the second AI model may be different or the same. In the above process, the second communication device can indicate through the above indication information that the feature dimensions of the second output are the same as the feature dimensions of the first output, allowing the first communication device to determine the first distillation loss based on the second output and the first output obtained by the first information without performing feature dimension processing, thereby reducing implementation complexity.

[0045] In a possible implementation of the second aspect, the first distillation loss is determined by performing feature dimension processing on the first output and a first processing result obtained by performing feature dimension processing on the second output, wherein the first processing result and the feature dimension of the first output are the same.

[0046] Based on the above solution, the first AI model and the second AI model can be deployed in different AI models respectively. Accordingly, the feature dimensions of the first output obtained based on the first AI model and the second output obtained based on the second AI model may be different or the same. In the above process, the first communication device can obtain the second output by receiving the first information. Accordingly, the first communication device can perform feature dimension processing based on the second output to obtain a first processing result, so that the first communication device can determine the first distillation loss based on the first processing result and the first output of the same feature dimension, thereby avoiding the situation where the distillation loss cannot be determined due to different feature dimensions.

[0047] In a possible implementation manner of the second aspect, the first information is used to indicate the first distillation loss; the method further includes: the second communication device receiving second information, where the second information is used to indicate the first output.

[0048] Based on the above solution, the second communication device may also receive second information indicating the first output, enabling the second communication device to determine the first distillation loss based on the second information. Accordingly, the first information sent by the second communication device to the first communication device may indicate the first distillation loss, allowing the first communication device to update the first AI model based on the first distillation loss, thereby reducing the implementation complexity of the first communication device.

[0049] In a possible implementation manner of the second aspect, the feature dimension of the second output is the same as the feature dimension of the first output.

[0050] Based on the above solution, the characteristic dimension of the first output indicated by the second information received by the second communication device may be the same as the characteristic dimension of the second output. In this way, the implementation complexity of the second communication device may be reduced.

[0051] In a possible implementation manner of the second aspect, the method further includes: the second communication device receiving third information, where the third information is used to determine the second input; wherein the third information is associated with the first input.

[0052] Based on the above solution, the second communication device can also receive third information used to determine the second input, and this third information is associated with the first input of the first AI model. In this way, the first and second AI models can determine corresponding first and second outputs based on the associated first and second inputs, thereby improving the performance of knowledge distillation based on the first and second outputs.

[0053] In a possible implementation manner of the second aspect, the first input is different from the second input.

[0054] Based on the above solution, the first input to the first AI model in the first communication device and the second input to the second AI model in the second communication device can be different. For example, the second input can contain more information than the first input. In this way, the second output obtained by the second communication device based on the second input can obtain more knowledge, thereby improving the performance of knowledge distillation based on the second output.

[0055] In a possible implementation of the second aspect, before receiving the first information, the method also includes: the second communication device sends configuration information of the first AI model; wherein the configuration information of the first AI model includes at least one of the following: a task identifier corresponding to the first AI model, an identifier of the configuration information of the first AI model, a network structure parameter of the first AI model, a training parameter of the first AI model, and a distillation parameter of the first AI model.

[0056] Based on the above solution, the first communication device and the second communication device can implement the deployment of the first AI model in the first communication device through the interaction of the configuration information of the first AI model. In addition, the configuration information of the first AI model can include at least one of the above items to enhance the flexibility of the solution implementation.

[0057] In addition, when the configuration information of the first AI model includes the distillation parameters of the first AI model, the first communication device is enabled to implement a subsequent knowledge distillation process based on the distillation parameters of the first AI model.

[0058] Optionally, the configuration information of the first AI model may be pre-configured.

[0059] In a possible implementation of the second aspect, the method further includes: the second communication device sends fourth information, which is used to indicate one or more distillation loss configuration information corresponding to the second AI model; the second communication device receives fifth information, which is used to indicate the first distillation loss configuration information among the one or more distillation loss parameters, and the first distillation loss configuration information is used to determine the first distillation loss.

[0060] Based on the above scheme, the first communication device and the second communication device can realize the process of selecting or indicating the distillation loss configuration information through the interaction process of the fourth information and the fifth information, so that the first communication device can realize the subsequent knowledge distillation process based on the distillation loss configuration adapted to the first communication device (and / or the first AI model).

[0061] Optionally, in the one or more distillation loss configuration information, each distillation loss configuration information includes at least one of the following: an identifier of each distillation loss configuration information, an indication information that the parameter type of the distillation loss is a hard label, a soft label, or intermediate features of some or all layers of the AI ​​model, or a layer index in the second AI model for outputting the second output.

[0062] In a possible implementation of the second aspect, the method further includes: the second communication device receives sixth information, where the sixth information is used to indicate AI capabilities and / or AI tasks; wherein the sixth information is used to determine the configuration information of the first AI model and / or the sixth information is used to determine the second AI model.

[0063] Based on the above solution, the second communication device can receive sixth information indicating AI capabilities and / or AI tasks, enabling the second communication device to determine the first AI model and / or the second AI model based on the sixth information. In this way, the first and second AI models used in the subsequent knowledge distillation process can be adapted to the AI ​​capabilities and / or AI tasks of the first communication device.

[0064] According to a third aspect of the present application, a communication device is provided, which is a first communication device and includes a transceiver unit and a processing unit; the transceiver unit is used to receive first information, and the first information is used to determine a first distillation loss of a first output and a second output; wherein the first output is obtained by processing a first AI model based on a first input, and the second output is obtained by processing a second AI model based on a second input; the processing unit updates the first AI model based on the first distillation loss.

[0065] In the third aspect of the present application, the constituent modules of the communication device can also be used to execute the steps performed in each possible implementation method of the first aspect and achieve corresponding technical effects. For details, please refer to the first aspect and will not be repeated here.

[0066] According to a fourth aspect of the present application, a communication device is provided, which is a second communication device. The device includes a transceiver unit and a processing unit. The processing unit is used to determine first information, and the first information is used to determine a first distillation loss of a first output and a second output; wherein the first output is obtained by processing the first AI model based on the first input, and the second output is obtained by processing the second AI model based on the second input; the first AI model is a student model, and the second AI model is a teacher model; the first distillation loss is used to update the first AI model; the transceiver unit is used to send the first information to the terminal device.

[0067] In the fourth aspect of the present application, the constituent modules of the communication device can also be used to execute the steps performed in each possible implementation method of the second aspect and achieve corresponding technical effects. For details, please refer to the second aspect and will not be repeated here.

[0068] In a fifth aspect, the present application provides a communication device, comprising at least one processor coupled to a memory; the memory is configured to store programs or instructions; the at least one processor is configured to execute the programs or instructions, so that the device implements the method described in any possible implementation of any one of the first to second aspects. Optionally, the communication device may include the memory.

[0069] In a sixth aspect, the present application provides a communication device comprising at least one logic circuit and an input / output interface; the logic circuit is used to execute the method described in any possible implementation of any one of the first to second aspects.

[0070] In a seventh aspect, the present application provides a communication system, which includes the above-mentioned first communication device and second communication device.

[0071] In an eighth aspect, the present application provides a computer-readable storage medium for storing one or more computer-executable programs or instructions. When the computer-executable programs or instructions are executed by a processor, the processor executes the method described in any possible implementation of any one of the first to second aspects above.

[0072] The ninth aspect of the present application provides a computer program product (or computer program). When the computer program or instructions in the computer program product are executed by the processor, the processor executes the method described in any possible implementation of any one of the first to second aspects above.

[0073] In a tenth aspect, the present application provides a chip system comprising at least one processor for supporting a communication device to implement the method described in any possible implementation of any one of the first to second aspects.

[0074] In one possible design, the chip system may further include a memory for storing programs / instructions and data necessary for the communication device. The chip system may be composed of a chip or may include a chip and other discrete components. Optionally, the chip system may further include an interface circuit for providing program instructions and / or data to the at least one processor.

[0075] Among them, the technical effects brought about by any design method in the third to tenth aspects can refer to the technical effects brought about by the different design methods in the above-mentioned first to second aspects, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figures 1a to 1c A schematic diagram of the communication system provided for this application;

[0077] Figures 2a to 2e A schematic diagram of the AI ​​processing process involved in this application;

[0078] Figure 3 An interactive schematic diagram of the communication method provided by this application;

[0079] Figure 4 Schematic diagram of the model processing provided for this application;

[0080] Figure 5a to Figure 5b Schematic diagram of the model processing provided for this application;

[0081] Figures 6 to 10 A schematic diagram of a communication device provided in this application. DETAILED DESCRIPTION

[0082] First, some of the terms used in the embodiments of the present application are explained to facilitate understanding by those skilled in the art.

[0083] (1) Terminal device: It can be a wireless terminal device that can receive network device scheduling and instruction information. The wireless terminal device can be a device that provides voice and / or data connectivity to the user, or a handheld device with wireless connection function, or other processing device connected to a wireless modem.

[0084] Terminal devices can communicate with one or more core networks or the Internet via a radio access network (RAN). Terminal devices can be mobile terminal devices, such as mobile phones (also known as "cellular" phones, mobile phones), computers, and data cards. For example, they can be portable, pocket-sized, handheld, computer-built-in, or vehicle-mounted mobile devices that exchange voice and / or data with the radio access network. Examples include personal communication service (PCS) phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, personal digital assistants (PDAs), tablets (tablets or pads), and computers with wireless transceiver capabilities. A wireless terminal device may also be referred to as a system, a subscriber unit, a subscriber station, a mobile station or mobile station (MS), a remote station (remotestation), an access point (AP), a remote terminal, an access terminal (accessterminal), a user terminal (user terminal), a user agent (user agent), a subscriber station (SS), a customer premises equipment (CPE), a terminal, a user equipment (UE), a mobile terminal (MT), etc.

[0085] As an example and not a limitation, in the embodiments of the present application, the terminal device may also be a wearable device. Wearable devices may also be referred to as wearable smart devices or smart wearable devices, etc., which are a general term for wearable devices that are intelligently designed and developed using wearable technology for daily wear, such as glasses, gloves, watches, clothing, and shoes. A wearable device is a portable device that is worn directly on the body or integrated into the user's clothes or accessories. Wearable devices are not only hardware devices, but also achieve powerful functions through software support, data interaction, and cloud interaction. Broadly speaking, wearable smart devices include those that are fully functional, large in size, and can achieve complete or partial functions without relying on smartphones, such as smart watches or smart glasses, etc., as well as those that only focus on a certain type of application function and need to be used in conjunction with other devices such as smartphones, such as various smart bracelets, smart helmets, and smart jewelry for vital sign monitoring.

[0086] The terminal may also be a drone, a robot, a terminal in device-to-device (D2D) communication, a terminal in vehicle-to-everything (V2X), a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in telemedicine (telehealth services), a wireless terminal in a smart grid, a wireless terminal in transportation safety, a wireless terminal in a smart city, a wireless terminal in a smart home, etc.

[0087] In addition, the terminal device may also be a terminal device in a communication system evolved after the fifth generation (5G) communication system (such as 5G Advanced or sixth generation (6G) communication system) or a terminal device in a future-evolved public land mobile network (PLMN). For example, 5G Advanced or 6G networks can further expand the form and functions of 5G communication terminals. 6G terminals include but are not limited to vehicles, cellular network terminals (with integrated satellite terminal functions), drones, and Internet of Things (IoT) devices.

[0088] In an embodiment of the present application, the terminal device may also obtain artificial intelligence (AI) services provided by the network device. Optionally, the terminal device may also have AI processing capabilities.

[0089] (2) Network equipment: It can be a device in a wireless network. For example, the network equipment can be a RAN node (or device) that connects a terminal device to a wireless network, which can also be called a base station. At present, some examples of RAN equipment are: base station, evolved NodeB (eNodeB), gNB (gNodeB) in a 5G communication system, transmission reception point (TRP), evolved Node B (eNB), radio network controller (RNC), Node B (NB), home base station (e.g., home evolved Node B, or home Node B, HNB), base band unit (BBU), or wireless fidelity (Wi-Fi) access point (AP). In addition, in a network structure, the network equipment can include a central unit (CU) node, a distributed unit (DU) node, or a RAN device including a CU node and a DU node.

[0090] Alternatively, a RAN node can be a macro base station, micro base station, indoor base station, relay node, donor node, or a wireless controller in a cloud radio access network (CRAN) scenario. A RAN node can also be a server, wearable device, vehicle, or vehicle-mounted device. For example, the access network device in vehicle-to-everything (V2X) technology can be a roadside unit (RSU).

[0091] In another possible scenario, multiple RAN nodes collaborate to assist the terminal in achieving wireless access, and different RAN nodes respectively implement part of the functions of the base station. For example, the RAN node can be a CU, DU, CU-control plane (CP), CU-user plane (UP), or radio unit (RU). The CU and DU can be set separately, or they can be included in the same network element, such as a baseband unit (BBU). The RU can be included in a radio frequency device or radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU), a radio head (RH), or a remote radio head (RRH).

[0092] 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 access network (open RAN, O-RAN or 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. For the convenience of description, this application takes CU, CU-CP, CU-UP, DU and RU as examples for description. Any unit of CU (or CU-CP, CU-UP), DU and RU in this application can be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

[0093] The communication between the access network device and the terminal device follows a certain protocol layer structure. The protocol layer may include a control plane protocol layer and a user plane protocol layer. The control plane protocol layer may include at least one of the following: a radio resource control (RRC) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, a media access control (MAC) layer, or a physical (PHY) layer. The user plane protocol layer may include at least one of the following: a service data adaptation protocol (SDAP) layer, a PDCP layer, an RLC layer, a MAC layer, or a physical layer.

[0094] For the correspondence between network elements in the ORAN system and their achievable protocol layer functions, please refer to Table 1 below.

[0095] Table 1

[0096] ORAN network elements 3GPP protocol layer functions O-CU-CP RRC+PCDP-Control Plane (PDCP-C) O-CU-UP SDAP+PCDP-User Plane (PDCP-U) O-DU RLC+MAC+PHY-high O-RU PHY-low

[0097] The network device may be any other device that provides wireless communication functionality to the terminal device. The embodiments of this application do not limit the specific technology and device form used by the network device. For ease of description, the embodiments of this application do not limit this.

[0098] The network equipment may also include core network equipment, which may include, for example, a mobility management entity (MME), a home subscriber server (HSS), a serving gateway (S-GW), a policy and charging rules function (PCRF), and a public data network gateway (PDN gateway or P-GW) in a fourth generation (4G) network; and network elements such as an access and mobility management function (AMF), a user plane function (UPF), or a session management function (SMF) in a 5G network. In addition, the core network equipment may also include other core network equipment in a 5G network and a next generation network of a 5G network.

[0099] In an embodiment of the present application, the above-mentioned network device may also have a network node with AI capabilities, which can provide AI services for terminals or other network devices. For example, it can be an AI node on the network side (access network or core network), a computing power node, a RAN node with AI capabilities, a core network element with AI capabilities, etc.

[0100] In the embodiments of the present application, the apparatus for implementing the function of the network device may be the network device, or may be a device capable of supporting the network device in implementing the function, such as a chip system, which may be installed in the network device. In the technical solutions provided in the embodiments of the present application, the technical solutions provided in the embodiments of the present application are described by taking the network device as an example.

[0101] (3) Configuration and pre-configuration: In this application, configuration and pre-configuration are used simultaneously. Configuration refers to the network device / server sending some parameter configuration information or parameter values ​​to the terminal through messages or signaling, so that the terminal can determine the communication parameters or resources during transmission based on these values ​​or information. Pre-configuration is similar to configuration, and can be parameter information or parameter values ​​pre-negotiated between the network device / server and the terminal device, or parameter information or parameter values ​​used by the base station / network device or terminal device as specified in the standard protocol, or parameter information or parameter values ​​pre-stored in the base station / server or terminal device. This application does not limit this.

[0102] Furthermore, these values ​​and parameters can be changed or updated.

[0103] (4) The terms "system" and "network" in the embodiments of the present application can be used interchangeably. "Multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, "at least one of A, B and C" includes A, B, C, AB, AC, BC or ABC. In addition, unless otherwise specified, the ordinal numbers such as "first" and "second" mentioned in the embodiments of the present application are used to distinguish multiple objects, and are not used to limit the order, timing, priority or importance of multiple objects.

[0104] (5) “Sending” and “receiving” in the embodiments of the present application indicate the direction of signal transmission. For example, “sending information to XX” can be understood as the destination of the information being XX, which can include direct sending through the air interface, as well as indirect sending through the air interface by other units or modules. “Receiving information from YY” can be understood as the source of the information being YY, which can include direct receiving from YY through the air interface, as well as indirect receiving from YY through the air interface from other units or modules. “Sending” can also be understood as the “output” of the chip interface, and “receiving” can also be understood as the “input” of the chip interface.

[0105] In other words, sending and receiving can be performed between devices, for example, between a network device and a terminal device, or can be performed within a device, for example, sending or receiving between components, modules, chips, software modules or hardware modules within the device through a bus, wiring or interface.

[0106] It is understandable that information may be processed between the source and destination of information transmission, such as coding, modulation, etc., but the destination can understand the valid information from the source. Similar expressions in this application can be understood similarly and will not be repeated.

[0107] (6) In the embodiments of the present application, "indication" may include direct indication and indirect indication, and may also include explicit indication and implicit indication. The information indicated by a certain information (such as the indication information described below) is called information to be indicated. In the specific implementation process, there are many ways to indicate the information to be indicated, such as but not limited to, directly indicating the information to be indicated, such as the information to be indicated itself or the index of the information to be indicated. The information to be indicated may also be indirectly indicated by indicating other information, wherein the other information is associated with the information to be indicated; or only a part of the information to be indicated may be indicated, while the other part of the information to be indicated is known or agreed in advance. For example, the indication of specific information may be achieved by means of the arrangement order of each information agreed in advance (such as predefined by the protocol), thereby reducing the indication overhead to a certain extent. The present application does not limit the specific method of indication. It is understandable that for the sender of the indication information, the indication information can be used to indicate the information to be indicated, and for the receiver of the indication information, the indication information can be used to determine the information to be indicated.

[0108] In this application, unless otherwise specified, the same or similar parts between the various embodiments can refer to each other. In the various embodiments of this application, and the various methods / designs / implementations in each embodiment, if there is no special explanation and logical conflict, the terms and / or descriptions between different embodiments and the various methods / designs / implementations in each embodiment are consistent and can be referenced to each other. The technical features in different embodiments and the various methods / designs / implementations in each embodiment can be combined to form new embodiments, methods, or implementations according to their inherent logical relationships. The following description of the implementation methods of this application does not constitute a limitation on the scope of protection of this application.

[0109] The present application can be applied to a long term evolution (LTE) system, a new radio (NR) system, or a communication system evolved after 5G (such as 6G, etc.). The communication system includes at least one network device and / or at least one terminal device.

[0110] See also Figure 1a , which is a schematic diagram of the communication system in this application. Figure 1aIn the example, a network device and 6 terminal devices are shown, and the 6 terminal devices are terminal device 1, terminal device 2, terminal device 3, terminal device 4, terminal device 5 and terminal device 6. Figure 1a In the example shown, terminal device 1 is a smart teacup, terminal device 2 is a smart air conditioner, terminal device 3 is a smart gas pump, terminal device 4 is a vehicle, terminal device 5 is a mobile phone, and terminal device 6 is a printer.

[0111] like Figure 1a As shown, the sending entity of the AI ​​configuration information can be a network device. The receiving entity of the AI ​​configuration information can be terminal devices 1-6. In this case, the network device and terminal devices 1-6 form a communication system. In this communication system, terminal devices 1-6 can send data to the network device, and the network device needs to receive data sent by terminal devices 1-6. At the same time, the network device can send configuration information to terminal devices 1-6.

[0112] For example, in Figure 1a In the example, terminal devices 4 and 6 can also form a communication system. Terminal device 5 acts as a network device, i.e., the sending entity of AI configuration information; terminal devices 4 and 6 act as terminal devices, i.e., the receiving entities of AI configuration information. For example, in a connected vehicle system, terminal device 5 sends AI configuration information to terminal devices 4 and 6, respectively, and receives data from terminal devices 4 and 6. Correspondingly, terminal devices 4 and 6 receive AI configuration information from terminal device 5 and send data to terminal device 5.

[0113] by Figure 1a Taking the communication system shown as an example, in addition to executing communication-related services, different devices (including between network devices and network devices, between network devices and terminal devices, and / or between terminal devices and terminal devices) may also execute AI-related services.

[0114] like Figure 1b As shown, taking the network device as a base station as an example, the base station can perform communication-related services and AI-related services with one or more terminal devices, and communication-related services and AI-related services can also be performed between different terminal devices.

[0115] like Figure 1c As shown, taking terminal devices including TV and mobile phone as an example, communication-related services and AI-related services can also be performed between the TV and the mobile phone.

[0116] The technical solution provided by this application can be applied to wireless communication systems (such as Figure 1a 、 Figure 1b or Figure 1c The system shown in the figure) is a system in which an AI network element can be introduced into the communication system provided in the present application to implement some or all AI-related operations. The AI ​​network element can also be called an AI node, an AI device, an AI entity, an AI module, an AI model, or an AI unit, etc. The AI ​​network element can be a network element built into the communication system. For example, the AI ​​network element can be an AI module built into: an access network device, a core network device, a cloud server, or an operation, administration and maintenance (OAM) to implement AI-related functions. The OAM can serve as a network manager for a core network device and / or as a network manager for an access network device. Alternatively, the AI ​​network element can also be a network element independently set up in the communication system. Optionally, the terminal or the chip built into the terminal can also include an AI entity to implement AI-related functions.

[0117] The following is a brief introduction to the AI ​​that may be involved in this application.

[0118] AI can imbue machines with human intelligence. For example, it can use computer hardware and software to simulate certain intelligent human behaviors. Machine learning methods can be used to achieve artificial intelligence. In machine learning, a machine uses training data to learn (or train) a model. This model represents the mapping from input to output. The learned model can be used for inference (or prediction), meaning that the model can be used to predict the output corresponding to a given input. This output can also be called an inference result (or prediction result).

[0119] Machine learning can include supervised learning, unsupervised learning, and reinforcement learning. Among them, unsupervised learning can also be called unsupervised learning.

[0120] Supervised learning uses machine learning algorithms to learn the mapping relationship between sample values ​​and sample labels based on collected sample values ​​and sample labels, and then expresses this learned mapping relationship using an AI model. The process of training a machine learning model is the process of learning this mapping relationship. During training, sample values ​​are input into the model to obtain the model's predicted values. The model parameters are optimized by calculating the error between the model's predicted values ​​and the sample labels (ideal values). Once the mapping relationship is learned, the learned mapping can be used to predict new sample labels. The mapping relationship learned by supervised learning can include linear mappings or nonlinear mappings. Based on the type of label, the learning task can be divided into classification tasks and regression tasks.

[0121] Unsupervised learning uses algorithms to discover inherent patterns in collected sample values. One type of unsupervised learning algorithm uses the samples themselves as supervisory signals, meaning the model learns the mapping from one sample to another. This is called self-supervised learning. During training, the model parameters are optimized by calculating the error between the model's predictions and the samples themselves. Self-supervised learning can be used in signal compression and decompression recovery applications. Common algorithms include autoencoders and generative adversarial networks.

[0122] Reinforcement learning, unlike supervised learning, is a type of algorithm that learns problem-solving strategies through interaction with the environment. Unlike supervised and unsupervised learning, reinforcement learning problems lack explicit label data for "correct" actions. Instead, the algorithm must interact with the environment to obtain reward signals from the environment, and then adjust its decision-making actions to maximize the reward signal value. For example, in downlink power control, the reinforcement learning model adjusts the downlink transmit power of each user based on the overall system throughput fed back by the wireless network, hoping to achieve higher system throughput. The goal of reinforcement learning is also to learn the mapping between environmental states and optimal (e.g., optimal) decision-making actions. However, because the labels for "correct actions" cannot be obtained in advance, network optimization cannot be achieved by calculating the error between actions and "correct actions." Reinforcement learning training is achieved through iterative interaction with the environment.

[0123] A neural network (NN) is a specific model in machine learning technology. According to the universal approximation theorem, NNs can theoretically approximate any continuous function, enabling them to learn arbitrary mappings. Traditional communication systems require extensive expert knowledge to design communication modules. However, deep learning communication systems based on neural networks can automatically discover implicit patterns in massive data sets and establish mapping relationships between data, achieving performance superior to traditional modeling methods.

[0124] The idea of ​​a neural network is derived from the neuronal structure of the brain. For example, each neuron performs a weighted sum operation on its input values ​​and outputs the result through an activation function.

[0125] like Figure 2a As shown in Figure 1, it is a schematic diagram of the neuron structure. Assume that the input of the neuron is x = [x0, x1, ..., x n ], and the weights corresponding to each input are w=[w0,w1,...,w n ], where n is a positive integer, w i and x i It can be a decimal, an integer (such as 0, a positive integer or a negative integer, etc.), or a complex number. i As x i The weight of xi Weighted. The bias of weighted summation of input values ​​according to 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, if the activation function of a neuron is: y = f(z) = z, then the output of the neuron is: b can be a decimal, an integer (eg, 0, a positive integer, or a negative integer), or a complex number, etc. The activation functions of different neurons in a neural network can be the same or different.

[0126] Furthermore, neural networks generally include multiple layers, each of which may include one or more neurons. Increasing the depth and / or width of a neural network can improve its expressive power, providing more powerful information extraction and abstract modeling capabilities for complex systems. The depth of a neural network can refer to the number of layers it comprises, and the number of neurons in each layer can be referred to as the width of that layer. In one implementation, a neural network includes an input layer and an output layer. The input layer processes the input information received by the neural network through neurons, passing the processing results to the output layer, which then obtains the output of the neural network. In another implementation, a neural network includes an input layer, a hidden layer, and an output layer. The input layer processes the input information received by the neural network through neurons, passing the processing results to an intermediate hidden layer. The hidden layer performs calculations on the received processing results to obtain a calculation result, which is then passed to the output layer or the next adjacent hidden layer, which ultimately obtains the output of the neural network. A neural network can include one hidden layer or multiple hidden layers connected in sequence, without limitation.

[0127] The neural network is, for example, a deep neural network (DNN). Depending on how the network is constructed, DNNs can include feedforward neural networks (FNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

[0128] Figure 2b This is a schematic diagram of a FNN network. FNNs are characterized by full connectivity between neurons in adjacent layers. This characteristic typically requires a large amount of storage space and results in high computational complexity.

[0129] CNN is a type of neural network specifically designed to process data with a grid-like structure. For example, time series data (e.g., discrete sampling on the time axis) and image data (e.g., discrete sampling on the two-dimensional axis) can both be considered grid-like data. CNNs do not utilize all input information at once for computation. Instead, they use a fixed-size window to intercept a portion of the information for convolution operations, significantly reducing the computational complexity of model parameters. Furthermore, depending on the type of information intercepted by the window (e.g., people and objects in an image represent different types of information), each window can use a different convolution kernel, enabling CNNs to better extract features from the input data.

[0130] RNNs are a type of DNN that utilizes feedback time series information. RNN inputs include the current input value and its own output value at the previous moment. RNNs are suitable for capturing temporally correlated sequence features and are particularly well-suited for applications such as speech recognition and channel coding.

[0131] During the machine learning model training process, a loss function can be defined. This function describes the gap or discrepancy between the model's output and the desired target value. Loss functions can be expressed in a variety of forms, and their specific form is not restricted. The model training process can be viewed as adjusting some or all of the model's parameters to keep the loss function below a threshold or meet the target.

[0132] A model may also be referred to as an AI model, rule, or other name. An AI model can be considered a specific method for implementing an AI function. An AI model represents a mapping relationship or function between the input and output of a model. AI functions may include one or more of the following: data collection, model training (or model learning), model information release, model inference (or model reasoning, inference, or prediction, etc.), model monitoring or model verification, or inference result release, etc. AI functions may also be referred to as AI (related) operations, or AI-related functions.

[0133] The following is an illustrative description of the implementation process of a fully connected neural network, also known as a multilayer perceptron (MLP), with reference to the accompanying drawings.

[0134] like Figure 2c As shown in the figure, an MLP consists of an input layer (left), an output layer (right), and multiple hidden layers (center). Each layer of the MLP contains several nodes, called neurons. Neurons in adjacent layers are connected to each other.

[0135] Alternatively, considering neurons in two adjacent layers, the output h of the neurons in the next layer is the weighted sum of all neurons x in the previous layer connected to it and passes through the activation function, which can be expressed as:

[0136] h=f(wx+b).

[0137] Among them, w is the weight matrix, b is the bias vector, and f is the activation function.

[0138] Alternatively, the output of the neural network can be recursively expressed as:

[0139] y=f n (w n f n-1 (...)+b n ).

[0140] Where n is the index of the neural network layer, n is greater than or equal to 1, and n is less than or equal to N, where N is the total number of neural network layers.

[0141] In other words, a neural network can be understood as a mapping from an input data set to an output data set. Neural networks are typically initialized randomly, and the process of obtaining this mapping from random w and b using existing data is called neural network training.

[0142] Optionally, a specific method of training is to use a loss function to evaluate the output results of the neural network.

[0143] like Figure 2d As shown, the error can be back-propagated, and the neural network parameters (including w and b) can be iteratively optimized by the gradient descent method until the loss function reaches the minimum value, that is, Figure 2d The “better point (e.g., optimal point)” in Figure 2d The neural network parameters corresponding to the “better point (e.g., optimal point)” in the training can be used as the neural network parameters in the trained AI model information.

[0144] Alternatively, the gradient descent process can be expressed as:

[0145]

[0146] Among them, θ is the parameter to be optimized (including w and b), L is the loss function, and η is the learning rate, which controls the step size of gradient descent. represents the derivative operation, represents the derivative of θ with respect to L.

[0147] Optionally, the backpropagation process utilizes the chain rule for partial derivatives.

[0148] like Figure 2eAs shown, the gradient of the previous layer parameters can be recursively calculated from the gradient of the next layer parameters, which can be expressed as:

[0149]

[0150] Among them, w ij is the weight of node j connecting to node i, s i is the weighted sum of the inputs to node i.

[0151] The technical solution provided by this application can be applied to wireless communication systems (such as Figure 1a or Figure 1b or Figure 1c In wireless communication systems, communication nodes generally possess both signal transceiver and computing capabilities. For example, network devices with computing capabilities primarily provide computing power to support signal transceiver capabilities (e.g., processing both sending and receiving signals), enabling communication between the network device and other communication nodes.

[0152] The technical solution provided by this application can be applied to wireless communication systems (such as Figure 1a or Figure 1b In wireless communication systems, communication nodes generally possess both signal transceiver and computing capabilities. For example, network devices with computing capabilities primarily provide computing power to support signal transceiver capabilities (e.g., processing both transmit and receive signals) to enable communication between the network device and other communication nodes. Furthermore, in addition to processing communication signals within the communication network, communication devices may also handle other communication tasks (e.g., channel prediction, beam management, resource scheduling, etc.).

[0153] However, in communication networks, communication nodes may have excess computing power beyond just supporting the aforementioned communication tasks. Therefore, how to utilize this computing power is a pressing technical issue.

[0154] In one possible implementation, a communication node can serve as a participating node in an AI learning system, applying its computing power to a specific component of the system. Generally speaking, in an AI learning system, AI nodes can establish datasets based on tasks, complete model training offline, and then perform inference after online deployment. However, after long periods of continuous learning, AI models may experience catastrophic forgetting and loss of plasticity, resulting in increasingly poor performance. Therefore, there are currently no technical solutions to improve the performance of AI models deployed on communication devices.

[0155] In order to solve the above problems, the present application provides a communication method and related devices, which will be described in detail below with reference to the accompanying drawings.

[0156] See also Figure 3 , is a schematic diagram of an implementation of the communication method provided in this application, and the method includes the following steps.

[0157] It should be noted that in the following, Figure 3 The method is illustrated by taking a first communication device and another communication device (e.g., a second communication device) as the execution subjects of the interaction diagram as an example, but the present application does not limit the execution subjects of the interaction diagram. For example, the communication device can be a communication device (e.g., a terminal device or a network device), or a chip, a baseband chip, a modem chip, a system on chip (SoC) chip including a modem core, a system in package (SIP) chip, a communication module, a chip system, a processor, a logic module, or software in the communication device.

[0158] S301. A second communication device sends first information, and the first communication device receives the first information. The first information is used to determine a first distillation loss for a first output and a second output, where the first output is obtained by processing a first AI model based on a first input, and the second output is obtained by processing a second AI model based on a second input.

[0159] S302. The first communication device updates the first AI model based on the first distillation loss.

[0160] In this application, AI model can be replaced by other terms, such as neural network, neural network model, AI neural network model, machine learning model, or AI processing model.

[0161] It should be understood that the first communication device can determine the first distillation loss through the first information. Generally, the distillation loss can describe / indicate / reflect the difference between the first output of the first AI model and the second output of the second AI model during the model compression or knowledge transfer process. For example, the distillation loss can include the mean squared error (MSE), KL divergence (Kullback-Leibler divergence), or other distance / similarity metric functions used to characterize the difference between the first output and the second output.

[0162] The first information can be implemented in multiple ways.

[0163] For example, since the first communication device can obtain the first output through the first AI model, the first information may include the second output, so that the first communication device can obtain the second output through the first information and determine the first distillation loss based on the second output and the first output. For details, see the implementation process of implementation method one below.

[0164] For another example, the first information may include the value of the first distillation loss (or a quantized value, an index value, etc.), so that the first communication device can determine the first distillation loss through the first information. For details, see the implementation process of the second implementation method below.

[0165] It should be noted that the first AI model can implement a model update process through knowledge distillation based on the second output provided by the second AI model. In this process, the first AI model can obtain the knowledge of the second AI model. Accordingly, the second AI model can be a larger or more complex model, and the first AI model can be a smaller or simpler model. Thus, in the case that the first communication device may not have strong AI capabilities, the model knowledge of the larger or more complex model can be deployed in the first AI model in the first communication device through knowledge distillation, which can improve the efficiency of model deployment.

[0166] Optionally, the first AI model and the second AI model can be described in a variety of ways. For example, the first AI model can be called a student model, and the second AI model can be called a teacher model; or, the first AI model can be called a knowledge learning model, and the second AI model can be called a knowledge transfer model; for example, the first AI model can be called a small model, and the second AI model can be called a large model.

[0167] For example, the first AI model is a small model, and the second AI model is a large model. The large model can refer to a machine learning model with a large number of parameters and a complex structure, capable of processing massive amounts of data and completing various complex tasks, such as natural language processing, computer vision, and speech recognition.

[0168] Alternatively, large models are typically built from deep neural networks, with billions or even hundreds of billions of parameters.

[0169] Optionally, the purpose of designing a large model can be to improve the model's expressiveness and predictive performance, and to be able to handle more complex tasks and data.

[0170] Alternatively, large models can learn complex patterns and features by training on massive amounts of data, have stronger generalization capabilities, and can make accurate predictions on unprocessed data.

[0171] In contrast, a small model can refer to a model with fewer parameters and fewer layers. Generally speaking, compared to small models, large models usually have more parameters and deeper layers, and have stronger expressive power and higher accuracy, but also require more computing resources and time for training and inference. They are suitable for scenarios with large data volumes and abundant computing resources, such as cloud computing, high-performance computing, and artificial intelligence.

[0172] Alternatively, small models have the advantages of being lightweight, efficient, and easy to deploy, and are suitable for scenarios with small data volumes and limited computing resources, such as mobile applications, embedded devices, and the Internet of Things.

[0173] based on Figure 3 In the scheme shown, the first communication device can determine the first distillation loss of the first output and the second output through the first information received in S301, wherein the first output can be obtained by processing based on the first input by the first AI model deployed on the first communication device, and the second output can be obtained by processing based on the second input by the second AI model deployed on other communication devices (such as the second communication device). Thereafter, in S302, the first communication device can update the first AI model based on the first distillation loss. In this way, while the computing power of the communication node can be applied to the processing of the AI ​​model, the first communication device can also obtain the knowledge of the second AI model by updating the model through the distillation loss to improve the model performance.

[0174] In addition, data privacy is relatively strong in wireless communication scenarios. In the above process, different communication devices can update the AI ​​model through knowledge distillation without transmitting their respective input data. While improving model performance, it can also protect the privacy of the data of each communication device to enhance user experience.

[0175] exist Figure 3 In one possible implementation of the illustrated solution, in S302, the first communication device updates the first AI model based on the first distillation loss, including: the first communication device updates the first AI model based on the first distillation loss and a task loss, wherein the task loss is determined based on the labels corresponding to the first output and the first input. In other words, during the knowledge distillation process, the first communication device may update the first AI model based on not only the first distillation loss but also the task loss, thereby further improving the performance of the updated first AI model obtained based on the knowledge distillation process.

[0176] Generally, task loss can be obtained by using a loss function specifically designed for the AI ​​task to optimize the performance of the AI ​​model when the AI ​​model performs the AI ​​task. For example, in a classification task, the loss function used to determine the task loss can be a cross-entropy loss function, which can be used to evaluate the difference between the probability distribution predicted by the AI ​​model and the true label. For another example, in a regression task, the loss function used to determine the task loss can include MSE, which can be used to measure the average squared difference between the predicted value and the true value.

[0177] exist Figure 3 In one possible implementation of the illustrated solution, the first input is different from the second input. Specifically, the first input of the first AI model in the first communication device and the second input of the second AI model in the second communication device may be different. For example, the second input may contain more information than the first input. In this way, the second output obtained by the second communication device based on the second input can obtain more knowledge to improve the performance of the knowledge distillation implemented based on the second output.

[0178] As an implementation example, Figure 4 As shown, Figure 3 The processing of the first AI model and the second AI model of the method shown can be achieved by Figure 4 The architecture shown is implemented.

[0179] exist Figure 4 In the example, the first AI model can be the smaller or simpler model (e.g., the student model) and the second AI model can be the larger or more complex model (e.g., the teacher model). Figure 4 The processing of each module shown includes:

[0180] The first AI model is configured to obtain a first output based on a first input and provide the first output to the distillation loss determination module.

[0181] The second AI model is configured to obtain a second output based on the second input, and provide the second output to the distillation loss determination module.

[0182] A distillation loss determination module is used to determine a distillation loss based on the first output and the second output, and provide the distillation loss to an external device (e.g., a first communication device where the first AI model is located).

[0183] A task loss determination module (optional module) is used to determine the task loss based on the first output and the label corresponding to the task performed by the first AI model, and provide the task loss to the outside (for example, the first communication device where the first AI model is located).

[0184] The distillation loss provided by the distillation loss determination module can be used in the model update process of the first AI model. Optionally, the model update process can also be based on the task loss provided by the task loss determination module. Thus, through the knowledge distillation process, the knowledge of the second AI model is transferred to the first AI model to improve the performance of the first AI model.

[0185] In addition, as we can see from the previous article, Figure 4 The first AI model shown can be deployed on a first communication device, Figure 4 The second AI model shown can be deployed in the second communication device. Figure 4 The other modules shown can be deployed in various ways, which will be described below with some implementation examples.

[0186] Implementation method 1, such as Figure 5a As shown, the distillation loss determination module can be deployed in the first communication device.

[0187] In implementation manner 1, the first information received by the first communication device in S301 may be used to indicate the second output.

[0188] In which, the first communication device can obtain a second output through the received first information, and accordingly, the first communication device can determine the first distillation loss based on the second output and the first output of the first AI model, so that the first communication device updates the first AI model based on the first distillation loss.

[0189] In a possible implementation of the first implementation, the method further includes: the first communication device receives indication information indicating that the characteristic dimension of the second output is the same as the characteristic dimension of the first output. Specifically, the first AI model and the second AI model can be deployed in different AI models respectively, and accordingly, the characteristic dimensions of the first output obtained based on the first AI model and the second output obtained based on the second AI model may be different or the same. In the above process, the first communication device can determine that the characteristic dimension of the second output is the same as the characteristic dimension of the first output through the above indication information, so that the first communication device can determine the first distillation loss based on the second output and the first output obtained based on the first information, without performing characteristic dimension processing, which can reduce the implementation complexity.

[0190] Optionally, different feature dimensions may refer to the dimensions of the output vector of the intermediate feature layer of the AI ​​model (and / or the vector of the output of the AI ​​model). For example, if the intermediate feature layer of the AI ​​model is MLP and has a width of 256, then the feature dimension of the intermediate feature layer of the AI ​​model is 256. For another example, if the AI ​​model is a transformer and the embedding dimension of the output of the AI ​​model is 512, and the sequence length is 100, then the feature dimension of the output of the AI ​​model is 51200 (i.e., 512*100=51200).

[0191] In another possible implementation of the first implementation, the first distillation loss is determined by processing the first output and a first processing result obtained by processing the second output using a characteristic dimension, wherein the first processing result and the first output have the same characteristic dimension. In the above process, the first communication device can obtain the second output from the received first information. Accordingly, the first communication device can perform characteristic dimension processing on the second output to obtain the first processing result. This allows the first communication device to determine the first distillation loss based on the first processing result and the first output having the same characteristic dimension, thus avoiding the situation where the distillation loss cannot be determined due to different characteristic dimensions.

[0192] Optionally, the feature dimension processing can be performed by an internal processing module of the first AI model, i.e., the first AI model has the feature dimension processing function. Alternatively, the feature dimension processing can be performed independently of the processing of an external processing module of the first AI model.

[0193] Implementation method 2, such as Figure 5b As shown, the distillation loss determination module can be deployed in the second communication device.

[0194] In the second implementation, the first information received by the first communication device in S301 may indicate the first distillation loss. Furthermore, before S301, the method further includes: the first communication device sending second information, where the second information is used to indicate the first output.

[0195] The first communication device may also transmit second information indicating the first output, enabling a recipient of the second information (e.g., the second communication device) to determine the first distillation loss based on the second information. Accordingly, the first information transmitted by the recipient to the first communication device may indicate the first distillation loss, allowing the first communication device to update the first AI model based on the first distillation loss, thereby reducing implementation complexity of the first communication device.

[0196] Optionally, the feature dimension processing can be performed by an internal processing module of the second AI model, i.e., the second AI model has the feature dimension processing function. Alternatively, the feature dimension processing can be performed independently of the processing of an external processing module of the second AI model.

[0197] In one possible implementation of the second implementation, the characteristic dimension of the second output is the same as the characteristic dimension of the first output. Specifically, the characteristic dimension of the first output indicated by the second information sent by the first communication device may be the same as the characteristic dimension of the second output. In this way, the implementation complexity of the recipient of the second information (e.g., the second communication device) can be reduced.

[0198] Optionally, if the characteristic dimension of the second output differs from the characteristic dimension of the first output, after the second communication device obtains the first output via the second information, the second communication device may determine the first distillation loss based on the second output and a second processing result obtained by processing the characteristic dimension of the first output, where the second processing result and the characteristic dimension of the second output are the same. In this manner, the second communication device can determine the first distillation loss based on the second processing result and the second output having the same characteristic dimension, thereby avoiding the situation where the distillation loss cannot be determined due to different characteristic dimensions.

[0199] In a possible implementation of the second implementation, before S301, the method further includes: the first communication device sends third information, and the third information is used to determine the second input; wherein the third information is associated with the first input. Specifically, the first communication device may also send third information for determining the second input, and the third information is associated with the first input of the first AI model. In this way, the first AI model and the second AI model can determine the corresponding first output and second output based on the associated first input and second input, so as to improve the performance of the knowledge distillation implemented based on the first output and the second output.

[0200] Optionally, in addition to implementation method one and implementation method two, the distillation loss determination module may also include other implementation methods. For example, the distillation loss determination module may be deployed on a third communication device, wherein the third communication device is different from the first communication device and the second communication device. Thus, by determining the distillation loss based on the first output and the second output by the third communication device, the implementation complexity of the first communication device and the second communication device can be reduced. Moreover, in the case where the third communication device is a trusted third-party device, the risk of privacy leakage caused by the direct interaction between the output of the AI ​​model (such as the first output and / or the second output) between the first communication device and the second communication device can also be avoided.

[0201] Optionally, in Implementation Methods 1 and 2, if the knowledge distillation process also requires the participation of a task loss determination module, the task loss determination module can similarly be deployed in a variety of ways. For example, the task loss determination module can be deployed on the first communication device to reduce the transmission overhead of the task loss. In another example, the task loss determination module can be deployed on the second communication device to reduce the implementation complexity of the first communication device.

[0202] exist Figure 3 In one possible implementation of the method shown, before the first communication device receives the first information in S301, the method further includes: the first communication device receives the configuration information of the first AI model; wherein the configuration information of the first AI model includes at least one of the following: the task identifier corresponding to the first AI model (for example, the task ID in Table 2 below), the identifier of the configuration information of the first AI model (for example, the configuration ID in Table 2 below), the network structure parameters of the first AI model, the training parameters of the first AI model, and the distillation parameters of the first AI model. Thus, the first communication device and the second communication device can implement the deployment of the first AI model in the first communication device through the interaction process of the configuration information of the first AI model. In addition, the configuration information of the first AI model may include at least one of the above items to enhance the flexibility of the solution implementation.

[0203] In addition, when the configuration information of the first AI model includes the distillation parameters of the first AI model, the first communication device is enabled to implement a subsequent knowledge distillation process based on the distillation parameters of the first AI model.

[0204] Optionally, the configuration information of the first AI model may be pre-configured.

[0205] As an implementation example, the configuration information of one or more AI models can be implemented as described in Table 2 below.

[0206] Table 2

[0207] Task ID Configuration ID Network structure parameters Training parameters Distillation parameters 0 0 {…} {…} {…} 0 1 {…} {…} {…} 0 2 {…} {…} {…} 0 … … … … 1 0

[0208] In Table 2, the network structure parameters include {neural network type, necessary parameters of the neural network structure}, which can be specific model components or the names of publicly known or open-source models. This means that a specific neural network can be restored from this field. Examples include {MLP, width 256, depth 10, input 100, output 2}, {Resnet-18, input 3x32x32, number of classes 10}, and {Transformer Encoder, 256-dimensional, 12 layers, 8 heads, 4 patches}.

[0209] In Table 2, the training parameters include basic parameters for model training, such as learning rate and scheduling method, batch size, number of training rounds, loss function, optimizer, and data preprocessing method, such as {learning rate 1e-3, cosine scheduler, batch size 256, epoch 200, CE loss, AdamW, imagenet_default}.

[0210] In Table 2, distillation parameters include one or more of the following: feature location and distribution weight, distillation loss function weight, and feature adaptation network type and location, such as {distilled features {0, 1, 3} weights {1, 1, 1}, distillation loss weight 0.9, adaptation method {linear projection} BS side}.

[0211] It is understandable that the first communication device can obtain the configuration information of one or more AI models shown in Table 2 above through pre-configuration or configuration by another communication device (e.g., a second communication device). Thereafter, through the instruction of the other communication device, the first communication device can determine that the configuration information of one of the AI ​​models is the configuration information of the first AI model.

[0212] exist Figure 3 In a possible implementation of the method shown, the method further includes: the first communication device receives fourth information, the fourth information is used to indicate one or more distillation loss configuration information corresponding to the second AI model; the first communication device sends fifth information, the fifth information is used to indicate the first distillation loss configuration information among the one or more distillation loss parameters, and the first distillation loss configuration information is used to determine the first distillation loss. Specifically, the first communication device and the second communication device can implement a process of selecting or indicating distillation loss configuration information through the interaction process of the fourth information and the fifth information, so that the first communication device can implement a subsequent knowledge distillation process based on the distillation loss configuration adapted to the first communication device (and / or the first AI model).

[0213] Optionally, in the one or more distillation loss configuration information, each distillation loss configuration information includes at least one of the following: an identifier of each distillation loss configuration information (for example, the "level" in Table 3 below), an indication that the parameter type of the distillation loss is a hard label, a soft label, or intermediate features of some or all layers of the AI ​​model (for example, the "open interface" in Table 3 below), or a layer index in the second AI model used to output the second output (for example, the "shared layer index" in Table 3 below).

[0214] It should be understood that when the distillation loss configuration information includes information indicating that the parameter type of the distillation loss is a hard label, a soft label, or intermediate features of some or all layers of the AI ​​model, the first communication device can use this information to clarify that the first distillation loss is determined using the hard label, the soft label, or intermediate features of some or all layers of the AI ​​model. In other words, this information can clarify that the first output and / or second output used to determine the first distillation loss is a hard label, a soft label, or intermediate features of some or all layers of the AI ​​model.

[0215] For example, in the above-mentioned implementation method 1, through the indication information, the first communication device can determine that the second output indicated by the first information is a hard label, a soft label, or an intermediate feature of part or all layers of the AI ​​model.

[0216] For example, in the second implementation method mentioned above, through the indication information, the first communication device can determine that the first output indicated by the second information is a hard label, a soft label, or an intermediate feature of part or all layers of the AI ​​model.

[0217] Generally, hard labels can be used to deterministically represent the class a sample belongs to. For example, in a binary classification problem, a sample either belongs to class 1 and can be labeled as 1 or True, or belongs to class 0 and can be labeled as 0 or False. In a multi-classification problem, if a sample belongs to a specific class, its corresponding hard label is 1, and the labels of all other classes are 0.

[0218] Correspondingly, the soft label can be an uncertainty representation of the sample category, and can represent the possibility of the sample belonging to each category in the form of a probability distribution. For example, in a multi-classification task, the soft label of a sample may be a probability vector, such as [0.3, 0.5, 0.2], which means that the sample has a 30% probability of belonging to the first category, a 50% probability of belonging to the second category, and a 20% probability of belonging to the third category. In addition, in knowledge distillation, the teacher model (such as the second AI model) generally generates soft labels for the samples. These labels contain rich relationship information between categories, and the student model (such as the first AI model) performs knowledge distillation (or knowledge transfer) by learning these soft labels.

[0219] As an implementation example, one or more distillation loss configuration information may be implemented as described in Table 3 below.

[0220] Table 3

[0221] grade Open interface Shared layer index 0 Shared hard labels A 1 Shared soft labels B 2 Sharing specific layer intermediate features C 3 Share intermediate features across all layers D

[0222] In Table 3, the fourth information may include one or more rows of information in Table 3, respectively corresponding to the above-mentioned one or more distillation loss configuration information.

[0223] In Table 3, the number of neural network layers of the second AI model is N (N is a positive integer) as an example. That is, the second AI model may include N neural network layers, and one or more of the N neural network layers are used to output the second output. In other words, any value of A / B / C / D can indicate a layer set, and the layer set can include the one or more layers.

[0224] As an example, if only the last layer is shared, any value of A / B / C / D can be expressed as {N} or {-1}, where N represents the maximum layer index.

[0225] As another example, any value among A / B / C / D may include the index of each layer in the layer set. For example, the level 3 index set in Table 3 may be represented as {1, 2, 3, ..., N}.

[0226] As another example, the relationship between the value of the shared layer index and the corresponding layer set can be preconfigured, that is, the layer set corresponding to any value of A / B / C / D can be preconfigured. For example, the layer set is the last layer (which can be expressed as {-1}) corresponds to the shared layer index value 0, the layer set includes all layers (which can be expressed as {1,2,3,…,N}) corresponds to the shared layer index value 1, the layer set includes the 1st, 2nd, and 3rd layers (which can be expressed as {1,2,3}) corresponds to the shared layer index value 2, and so on.

[0227] Optionally, in addition to the columns of information shown in Table 3, in the one or more distillation loss configuration information, each distillation loss configuration information may also include the architecture information of the model (for example, the architecture information is used to indicate that the model architecture of the second AI model is MLP, CNN, or Transformer, etc.). This information helps the first communication device to design a more efficient distillation structure / algorithm.

[0228] exist Figure 3 In a possible implementation of the method shown, before S301, the method further includes: the first communication device sends sixth information, and the sixth information is used to indicate AI capabilities and / or AI tasks; wherein the sixth information is used to determine the configuration information of the first AI model and / or the sixth information is used to determine the second AI model. Specifically, the first communication device may send sixth information for indicating AI capabilities and / or AI tasks, so that the recipient of the sixth information (e.g., the second communication device) can determine the first AI model and / or the second AI model based on the sixth information. In this way, the first AI model and the second AI model used in the subsequent knowledge distillation process can be adapted to the AI ​​capabilities and / or AI tasks of the first communication device.

[0229] See also Figure 6 , an embodiment of the present application provides a communication device 600, which can implement the functions of the first communication device (or second communication device) in the above method embodiment, and thus can also achieve the beneficial effects of the above method embodiment. In the embodiment of the present application, the communication device 600 can be the first communication device (or second communication device), or it can be an integrated circuit or component inside the first communication device (or second communication device), such as a chip, a baseband chip, a modem chip, a SoC chip including a modem core, a system-in-package (SIP) chip, a communication module, a chip system, a processor, etc.

[0230] It should be noted that the transceiver unit 602 may include a sending unit and a receiving unit, which are respectively used to perform sending and receiving.

[0231] In a possible implementation, when the device 600 is used to execute Figure 3 And when the method is executed by the first communication device in the related embodiment, the device 600 includes a processing unit 601 and a transceiver unit 602; the transceiver unit 602 is used to receive first information, and the first information is used to determine the first distillation loss of the first output and the second output; wherein the first output is obtained by the first AI model based on the first input, and the second output is obtained by the second AI model based on the second input; the processing unit 601 updates the first AI model based on the first distillation loss.

[0232] In a possible implementation, when the device 600 is used to execute Figure 3 And when the method is executed by the second communication device in the related embodiment, the device 600 includes a processing unit 601 and a transceiver unit 602; the processing unit 601 is used to determine the first information, and the first information is used to determine the first distillation loss of the first output and the second output; wherein, the first output is obtained by the first AI model based on the first input, and the second output is obtained by the second AI model based on the second input; the first AI model is a student model, and the second AI model is a teacher model; the first distillation loss is used to update the first AI model; the transceiver unit 602 is used to send the first information to the terminal device.

[0233] In one possible design, when the communication device 600 is a terminal device or a communication module in a terminal, the functions of the processing unit 601 can be implemented by one or more processors. Specifically, the processor can include a modem chip, or a SoC chip or SIP chip containing a modem core. The functions of the transceiver unit 602 can be implemented by a transceiver circuit.

[0234] In one possible design, when the communication device 600 is a circuit or chip responsible for communication functions in a terminal, such as a modem chip or a SoC chip or SIP chip containing a modem core, the functions of the processing unit 601 can be implemented by a circuit system including one or more processors or processor cores in the above chip. The functions of the transceiver unit 602 can be implemented by an interface circuit or data transceiver circuit on the above chip.

[0235] It should be noted that, for details on the information execution process of the units of the above-mentioned communication device 600, please refer to the description in the method embodiment shown above in this application, and no further details will be given here.

[0236] See also Figure 7 , is another schematic structural diagram of a communication device 700 provided in this application, wherein the communication device 700 includes a logic circuit 701 and an input / output interface 702. The communication device 700 may be a chip or an integrated circuit.

[0237] in, Figure 6 The transceiver unit 602 shown may be a communication interface, which may be Figure 7 The input / output interface 702 in the embodiment may include an input interface and an output interface. Alternatively, the communication interface may be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0238] In a possible implementation, when the device 700 is used to execute Figure 3 And when the method is executed by the first communication device in the related embodiments, the input-output interface 702 is used to receive first information, and the first information is used to determine the first distillation loss of the first output and the second output; wherein the first output is obtained by the first AI model based on the first input, and the second output is obtained by the second AI model based on the second input; the logic circuit 701 updates the first AI model based on the first distillation loss.

[0239] In a possible implementation, when the device 700 is used to execute Figure 3 And when the method is executed by the second communication device in the related embodiments, the logic circuit 701 is used to determine the first information, and the first information is used to determine the first distillation loss of the first output and the second output; wherein, the first output is obtained by the first AI model based on the first input, and the second output is obtained by the second AI model based on the second input; the first AI model is a student model, and the second AI model is a teacher model; the first distillation loss is used to update the first AI model; the input and output interface 702 is used to send the first information to the terminal device.

[0240] The logic circuit 701 and the input / output interface 702 may also execute other steps executed by the first communication device or the second communication device in any embodiment and achieve corresponding beneficial effects, which will not be described in detail here.

[0241] In one possible implementation, Figure 6 The processing unit 601 shown can be Figure 7 Logic circuit 701 in.

[0242] Optionally, the logic circuit 701 may be a processing device, and the functions of the processing device may be partially or entirely implemented by software. The functions of the processing device may be partially or entirely implemented by software.

[0243] Optionally, the processing device may include a memory and a processor, wherein the memory is used to store a computer program, and the processor reads and executes the computer program stored in the memory to perform corresponding processing and / or steps in any one of the method embodiments.

[0244] Alternatively, the processing device may include only a processor. A memory for storing the computer program is located outside the processing device, and the processor is connected to the memory via circuits / wires to read and execute the computer program stored in the memory. The memory and processor may be integrated or physically separate.

[0245] Optionally, the processing device may be one or more chips, or one or more integrated circuits. For example, the processing device may be one or more field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), system-on-chips (SoCs), central processor units (CPUs), network processors (NPs), digital signal processors (DSPs), microcontroller units (MCUs), programmable logic controllers (PLDs), or other integrated chips, or any combination of the above chips or processors.

[0246] See also Figure 8 , is the communication device 800 involved in the above embodiment provided in the embodiment of the present application, and the communication device 800 may specifically be the communication device as the terminal device in the above embodiment, Figure 8The example shown is implemented by a terminal device (or a component in the terminal device).

[0247] Herein, a possible logical structure diagram of the communication device 800 is shown. The communication device 800 may include but is not limited to at least one processor 801 and a communication port 802 .

[0248] in, Figure 6 The transceiver unit 602 shown may be a communication interface, which may be Figure 8 The communication port 802 may include an input interface and an output interface. Alternatively, the communication port 802 may also be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0249] Further optionally, the device may also include at least one of a memory 803 and a bus 804. In an embodiment of the present application, the at least one processor 801 is used to control and process the actions of the communication device 800.

[0250] Furthermore, the processor 801 may be a central processing unit (CPU), a general-purpose processor (GPPC), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic device (PLD), a transistor logic device (TLD), a hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the systems, devices, and units described above may refer to the corresponding processes in the aforementioned method embodiments and will not be further described herein.

[0251] It should be noted that Figure 8 The communication device 800 shown can be specifically used to implement the steps implemented by the terminal device in the aforementioned method embodiment and achieve the corresponding technical effects of the terminal device. Figure 8 The specific implementation of the communication device shown can refer to the description in the aforementioned method embodiment, and will not be repeated here.

[0252] See also Figure 9 , is a structural diagram of a communication device 900 involved in the above embodiment provided in an embodiment of the present application. The communication device 900 may specifically be a communication device as a network device in the above embodiment. Figure 9 The example shown is that the network device is implemented by the network device (or a component in the network device), wherein the structure of the communication device can refer to Figure 9 The structure shown.

[0253] The communication device 900 includes at least one processor 911 and at least one network interface 914. Further optionally, the communication device also includes at least one memory 912, at least one transceiver 913 and one or more antennas 915. The processor 911, the memory 912, the transceiver 913 and the network interface 914 are connected, for example, via a bus. In an embodiment of the present application, the connection may include various interfaces, transmission lines or buses, etc., which are not limited in this embodiment. The antenna 915 is connected to the transceiver 913. The network interface 914 is used to enable the communication device to communicate with other communication devices through a communication link. For example, the network interface 914 may include a network interface between the communication device and the core network device, such as an S1 interface, and the network interface may include a network interface between the communication device and other communication devices (such as other network devices or core network devices), such as an X2 or Xn interface.

[0254] in, Figure 6 The transceiver unit 602 shown may be a communication interface, which may be Figure 9 The network interface 914 in the embodiment may include an input interface and an output interface. Alternatively, the network interface 914 may also be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0255] Processor 911 is primarily used to process communication protocols and communication data, control the entire communication device, execute software programs, and process software program data, for example, to support the communication device in performing the actions described in the embodiments. The communication device may include a baseband processor and a central processing unit. The baseband processor is primarily used to process communication protocols and communication data, while the central processing unit is primarily used to control the entire terminal device, execute software programs, and process software program data. Figure 9 The processor 911 in the embodiment can integrate the functions of a baseband processor and a central processing unit. Those skilled in the art will appreciate that the baseband processor and the central processing unit can also be independent processors interconnected via a bus or other technology. Those skilled in the art will appreciate that a terminal device can include multiple baseband processors to adapt to different network standards, and a terminal device can include multiple central processing units to enhance its processing capabilities. The various components of the terminal device can be connected via various buses. The baseband processor can also be expressed as a baseband processing circuit or a baseband processing chip. The central processing unit can also be expressed as a central processing circuit or a central processing chip. The function of processing communication protocols and communication data can be built into the processor or stored in a memory in the form of a software program, which is executed by the processor to implement the baseband processing function.

[0256] The memory is primarily used to store software programs and data. Memory 912 can exist independently and be connected to processor 911. Alternatively, memory 912 and processor 911 can be integrated together, for example, within a single chip. Memory 912 can store program code for executing the technical solutions of the embodiments of the present application, and execution is controlled by processor 911. The various computer program codes executed can also be considered drivers for processor 911.

[0257] Figure 9 Only one memory and one processor are shown. In an actual terminal device, multiple processors and multiple memories may exist. Memory may also be referred to as a storage medium or storage device. The memory may be a storage element on the same chip as the processor, i.e., an on-chip storage element, or an independent storage element, which is not limited in the embodiments of the present application.

[0258] The transceiver 913 can be used to support the reception or transmission of radio frequency signals between the communication device and the terminal, and the transceiver 913 can be connected to the antenna 915. The transceiver 913 includes a transmitter Tx and a receiver Rx. Specifically, one or more antennas 915 can receive radio frequency signals. The receiver Rx of the transceiver 913 is used to receive the radio frequency signal from the antenna, convert the radio frequency signal into a digital baseband signal or a digital intermediate frequency signal, and provide the digital baseband signal or digital intermediate frequency signal to the processor 911 so that the processor 911 can further process the digital baseband signal or digital intermediate frequency signal, such as demodulation and decoding. In addition, the transmitter Tx in the transceiver 913 is also used to receive a modulated digital baseband signal or digital intermediate frequency signal from the processor 911, convert the modulated digital baseband signal or digital intermediate frequency signal into a radio frequency signal, and send the radio frequency signal through one or more antennas 915. Specifically, the receiver Rx can selectively perform one or more stages of down-mixing and analog-to-digital conversion on the RF signal to obtain a digital baseband signal or a digital intermediate frequency signal. The order of the down-mixing and analog-to-digital conversion processes is adjustable. The transmitter Tx can selectively perform one or more stages of up-mixing and digital-to-analog conversion on the modulated digital baseband signal or digital intermediate frequency signal to obtain a RF signal. The order of the up-mixing and digital-to-analog conversion processes is adjustable. The digital baseband signal and the digital intermediate frequency signal may be collectively referred to as digital signals.

[0259] The transceiver 913 may also be referred to as a transceiver unit, a transceiver, a transceiver device, etc. Optionally, a device in the transceiver unit that implements a receiving function may be referred to as a receiving unit, and a device in the transceiver unit that implements a transmitting function may be referred to as a transmitting unit. That is, the transceiver unit includes a receiving unit and a transmitting unit. The receiving unit may also be referred to as a receiver, an input port, a receiving circuit, etc., and the transmitting unit may be referred to as a transmitter, a transmitter, or a transmitting circuit, etc.

[0260] It should be noted that Figure 9 The communication device 900 shown can be specifically used to implement the steps implemented by the network device in the aforementioned method embodiment and achieve the corresponding technical effects of the network device. Figure 9 The specific implementation of the communication device 900 can refer to the description in the aforementioned method embodiment, which will not be repeated here.

[0261] See also Figure 10 , which is a structural diagram of the communication device involved in the above embodiments provided in the embodiments of the present application.

[0262] It can be understood that the communication device 100 includes, for example, modules, units, elements, circuits, or interfaces, which are appropriately configured together to implement the technical solutions provided in this application. The communication device 100 can be the terminal device or network device described above, or a component (such as a chip) in these devices, used to implement the method described in the following method embodiment. The communication device 100 includes one or more processors 101. The processor 101 can be a general-purpose processor or a dedicated processor. For example, it can be a baseband processor or a central processing unit. The baseband processor can be used to process communication protocols and communication data, and the central processing unit can be used to control the communication device (such as a RAN node, terminal, or chip, etc.), execute software programs, and process data of software programs.

[0263] Optionally, in one design, the processor 101 may include a program 103 (sometimes also referred to as code or instruction), and the program 103 may be executed on the processor 101 to enable the communication device 100 to perform the method described in the following embodiments. In another possible design, the communication device 100 includes a circuit ( Figure 10 not shown).

[0264] Optionally, the communication device 100 may include one or more memories 102 on which a program 104 (sometimes also referred to as code or instructions) is stored. The program 104 can be run on the processor 101, so that the communication device 100 executes the method described in the above method embodiment.

[0265] Optionally, the processor 101 and / or the memory 102 may include AI modules 107 and 108, which are used to implement AI-related functions. The AI ​​module may be implemented through software, hardware, or a combination of software and hardware. For example, the AI ​​module may include a wireless intelligent control (RIC) module. For example, the AI ​​module may be a near-real-time RIC or a non-real-time RIC.

[0266] Optionally, data may be stored in the processor 101 and / or the memory 102. The processor and the memory may be provided separately or integrated together.

[0267] Optionally, the communication device 100 may further include a transceiver 105 and / or an antenna 106. The processor 101 may also be sometimes referred to as a processing unit, and controls the communication device (e.g., a RAN node or terminal). The transceiver 105 may also be sometimes referred to as a transceiver unit, a transceiver, a transceiver circuit, or a transceiver, and is configured to implement the transceiver functions of the communication device through the antenna 106.

[0268] in, Figure 6 The illustrated processing unit 601 may be the processor 101 . Figure 6 The transceiver unit 602 shown may be a communication interface, which may be Figure 10 The transceiver 105 in the embodiment may include an input interface and an output interface. Alternatively, the transceiver 105 may be a transceiver circuit, which may include an input interface circuit and an output interface circuit.

[0269] An embodiment of the present application further provides a computer-readable storage medium, which is used to store one or more computer-executable instructions. When the computer-executable instructions are executed by a processor, the processor executes the method described in the possible implementation methods of the first communication device or the second communication device in the aforementioned embodiment.

[0270] An embodiment of the present application also provides a computer program product (or computer program). When the computer program product is executed by the processor, the processor executes the method that may be implemented by the above-mentioned first communication device or second communication device.

[0271] An embodiment of the present application also provides a chip system, which includes at least one processor for supporting a communication device to implement the functions involved in the possible implementation methods of the above-mentioned communication device. Optionally, the chip system also includes an interface circuit, which provides program instructions and / or data to the at least one processor. In one possible design, the chip system may also include a memory, which is used to store the necessary program instructions and data for the communication device. The chip system can be composed of chips, or it can include chips and other discrete devices, wherein the communication device can specifically be the first communication device or the second communication device in the aforementioned method embodiment.

[0272] An embodiment of the present application further provides a communication system, wherein the network system architecture includes the first communication device and / or the second communication device in any of the above embodiments.

[0273] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. Whether a function is performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0274] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0275] In addition, the functional units in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the contributing part or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Claims

1. A communication method, characterized in that: The method is applied to a terminal device or a module in the terminal device, and includes: Receive first information, where the first information is used to determine a first distillation loss of a first output and a second output; wherein the first output is obtained by processing a first artificial intelligence (AI) model based on a first input, and the second output is obtained by processing a second AI model based on a second input; The first AI model is updated based on the first distillation loss.

2. The method according to claim 1, characterized in that The first information is used to indicate the second output.

3. The method according to claim 2, characterized in that The method further comprises: Indication information indicating that a feature dimension of the second output is the same as a feature dimension of the first output is received.

4. The method according to claim 2, characterized in that The first distillation loss is determined by performing feature dimension processing on the first output and a first processing result obtained by performing feature dimension processing on the second output, wherein the first processing result and the feature dimension of the first output are the same.

5. The method according to claim 1, characterized in that The first information is used to indicate the first distillation loss; the method further includes: Second information is sent, where the second information is used to indicate the first output.

6. The method according to claim 5, characterized in that The feature dimension of the second output is the same as the feature dimension of the first output.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Sending third information, where the third information is used to determine the second input; wherein the third information is associated with the first input.

8. The method according to any one of claims 1 to 7, characterized in that The first input is different from the second input.

9. The method according to any one of claims 1 to 8, characterized in that Before receiving the first information, the method further includes: Receive configuration information of the first AI model; The configuration information of the first AI model includes at least one of the following: The task identifier corresponding to the first AI model, the identifier of the configuration information of the first AI model, the network structure parameters of the first AI model, the training parameters of the first AI model, and the distillation parameters of the first AI model.

10. The method according to any one of claims 1 to 9, characterized in that The method further comprises: receiving fourth information, where the fourth information is used to indicate one or more distillation loss configuration information corresponding to the second AI model; Fifth information is sent, where the fifth information is used to indicate first distillation loss configuration information among the one or more distillation loss parameters, where the first distillation loss configuration information is used to determine the first distillation loss.

11. The method according to claim 10, characterized in that In the one or more distillation loss configuration information, each distillation loss configuration information includes at least one of the following: The identifier of each distillation loss configuration information, the parameter type of the distillation loss is a hard label, a soft label, an indication of the intermediate features of some or all layers of the AI ​​model, or the layer index in the second AI model used to output the second output.

12. The method according to any one of claims 1 to 11, characterized in that The method further comprises: Send sixth information, where the sixth information is used to indicate AI capabilities and / or AI tasks; wherein the sixth information is used to determine configuration information of the first AI model and / or the sixth information is used to determine the second AI model.

13. A communication method, characterized in that: include: Determine first information, where the first information is used to determine a first distillation loss for a first output and a second output; wherein the first output is obtained by processing a first artificial intelligence (AI) model based on a first input, and the second output is obtained by processing a second AI model based on a second input; the first AI model is a student model, and the second AI model is a teacher model; and the first distillation loss is used to update the first AI model; Send the first information to the terminal device.

14. The method according to claim 13, characterized in that The first information is used to indicate the second output.

15. The method according to claim 14, characterized in that The method further comprises: Indication information is sent indicating that the feature dimension of the second output is the same as the feature dimension of the first output.

16. The method according to claim 14, characterized in that The first distillation loss is determined by performing feature dimension processing on the first output and a first processing result obtained by performing feature dimension processing on the second output, wherein the first processing result and the feature dimension of the first output are the same.

17. The method according to claim 13, wherein The first information is used to indicate the first distillation loss; the method further includes: Second information is received, where the second information is used to indicate the first output.

18. The method according to claim 17, characterized in that The feature dimension of the second output is the same as the feature dimension of the first output.

19. The method according to any one of claims 13 to 18, characterized in that The method further comprises: Receive third information, where the third information is used to determine the second input; wherein the third information is associated with the first input.

20. The method according to any one of claims 13 to 19, characterized in that The first input is different from the second input.

21. The method according to any one of claims 13 to 20, characterized in that Before receiving the first information, the method further includes: Sending configuration information of the first AI model; The configuration information of the first AI model includes at least one of the following: The task identifier corresponding to the first AI model, the identifier of the configuration information of the first AI model, the network structure parameters of the first AI model, the training parameters of the first AI model, and the distillation parameters of the first AI model.

22. The method according to any one of claims 13 to 21, characterized in that The method further comprises: Sending fourth information, where the fourth information is used to indicate one or more distillation loss configuration information corresponding to the second AI model; Fifth information is received, where the fifth information is used to indicate first distillation loss configuration information among the one or more distillation loss parameters, where the first distillation loss configuration information is used to determine the first distillation loss.

23. The method according to claim 22, characterized in that In the one or more distillation loss configuration information, each distillation loss configuration information includes at least one of the following: The identifier of each distillation loss configuration information, the parameter type of the distillation loss is a hard label, a soft label, an indication of the intermediate features of some or all layers of the AI ​​model, or the layer index in the second AI model used to output the second output.

24. The method according to any one of claims 13 to 23, characterized in that The method further comprises: Receive sixth information, where the sixth information is used to indicate AI capabilities and / or AI tasks; wherein the sixth information is used to determine configuration information of the first AI model and / or the sixth information is used to determine the second AI model.

25. A communication device, characterized in that: Comprising means for performing the method according to any one of claims 1 to 24.

26. A communication device, characterized in that: The method comprises at least one processor coupled to a memory; the at least one processor is configured to execute the method according to any one of claims 1 to 24.

27. The communication device according to claim 26, characterized in that The communication device is a chip or a chip system.

28. A readable storage medium, characterized in that The storage medium stores a computer program or instruction, and when the computer program or instruction is executed by the communication device, the method according to any one of claims 1 to 24 is implemented.

29. A computer program product, characterized in that The method comprises instructions or a computer program, which, when executed on a computer, causes the computer to perform the method according to any one of claims 1 to 24.

Images