Communication method and apparatus

Abstract

A communication method and apparatus. The method comprises: a first device determining the degree of matching between training data and inference data on the basis of distribution information of the training data and distribution information of the inference data; and the first device determining, on the basis of the degree of matching between the training data and the inference data, whether a first model can be applied to the inference data, wherein the first model is obtained by means of performing training using the training data.

Description

Communication method and device Technical Field

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

[0002] Artificial intelligence (AI) technology learning involves offline learning. Offline learning involves pre-training a model using training data to obtain a model for inference. After training, the trained model is applied to the inference data to obtain inference results.

[0003] The inference data used in actual model applications often differs from the training data. For example, in wireless communication scenarios, inference data can change as the device moves. Furthermore, inference data may include sensory data, which can vary with environmental factors such as temperature and humidity. Consequently, the difference between inference data and training data can increase over time.

[0004] When the difference between the inference data and the training data is large, if the model obtained based on the training data is applied to the inference data, the model may not achieve good performance, resulting in poor generalization performance of the model.

[0005] Summary of the Invention

[0006] The embodiments of the present application provide a communication method and device, which are conducive to improving the generalization of the model.

[0007] In a first aspect, an embodiment of the present application provides a communication method that can be performed by a first device. The first device herein can refer to the first device itself or to a processor, module, chip, or chip system in the first device that implements the method. In this method, the first device determines the degree of match between the training data and the inference data based on distribution information of the training data and distribution information of the inference data. Based on the degree of match between the training data and the inference data, the first device determines whether a first model can be applied to the inference data, where the first model is trained using the training data.

[0008] As can be seen, the first device determines the degree of match between the training data and the inference data, and based on this degree of match, determines whether the first model obtained based on the training data can be applied to the inference data. This facilitates the first device determining that the first model can be applied to the inference data when the degree of match between the training data and the inference data is high, thereby improving the generalization of the first model.

[0009] In one possible implementation, the degree of matching between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data; alternatively, the degree of matching between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data; alternatively, the degree of matching between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data. Thus, the degree of matching between the training data and the inference data can be flexibly determined in a variety of ways.

[0010] Among them, the difference between the distribution information of the training data and the distribution information of the inference data indicates the degree of difference between the distribution information of the training data and the distribution information of the inference data; the similarity between the distribution information of the training data and the distribution information of the inference data indicates the degree of similarity between the distribution information of the training data and the distribution information of the inference data; the distance between the distribution information of the training data and the distribution information of the inference data indicates the distance between the distribution information of the training data and the distribution information of the inference data.

[0011] In one possible implementation, the first model is an artificial intelligence AI model or a machine learning ML model.

[0012] In one possible implementation, the training data and the inference data include multiple features, such that the distribution information of the training data includes distribution information of the multiple features in the training data, and the distribution information of the inference data includes distribution information of the multiple features in the inference data, and the multiple features are of the same type in the training data and the inference data. In this case, the first device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, including: determining the degree of matching corresponding to each feature based on the distribution information of each feature in the training data and the distribution information of the inference data, where the degree of matching corresponding to each feature is the degree of matching between the training data and the inference data; and determining the degree of matching between the training data and the inference data based on the matching degree corresponding to each feature.

[0013] It can be seen that when the training data and the inference data include multiple features, the first device uses the matching degree corresponding to each feature in the multiple features to determine the matching degree between the training data and the inference data, which can improve the accuracy of the matching degree between the training data and the inference data.

[0014] In one possible implementation, the first device determines whether the first model can be applied to the inference data based on a degree of match between the training data and the inference data, including: determining that the first model can be applied to the inference data when the degree of match between the training data and the inference data is greater than a first threshold; or determining that the first model cannot be applied to the inference data when the degree of match between the training data and the inference data is less than or equal to the first threshold. The first threshold may be preset or configurable.

[0015] It is understood that when the degree of match between the training data and the inference data is greater than the first threshold, it indicates that the difference between the distribution information of the training data and the distribution information of the inference data is small, or that the similarity between the distribution information of the training data and the distribution information of the inference data is large, or that the distance between the distribution information of the training data and the distribution information of the inference data is small. If the first model is applied to the inference data, the first model can have good generalization, and the first device determines that the first model can be applied to the inference data. When the degree of match between the training data and the inference data is less than or equal to the first threshold, it indicates that the difference between the distribution information of the training data and the distribution information of the inference data is large, or that the similarity between the distribution information of the training data and the distribution information of the inference data is small, or that the distance between the distribution information of the training data and the distribution information of the inference data is large. If the first model is applied to the inference data, the first model cannot have good generalization, and the first device determines that the first model cannot be applied to the inference data. In this manner, the first device determines that the first model can be applied to the inference data while ensuring that the first model can have good generalization, thereby improving the generalization of the first model.

[0016] In one possible implementation, the first device is a terminal device, i.e., the method is performed by the terminal device. In this case, when the degree of match between the training data and the inference data exceeds a first threshold, the terminal device may further perform the following steps: sending an inference request message to a network device, the inference request message being used to request a first model for inference; receiving the first model from the network device; and obtaining an inference result based on the first model and the inference data.

[0017] As can be seen, when the degree of match between the training data and the inference data is high, the terminal device requests the network device to obtain the first model for inference and determines the inference result based on the obtained first model and the inference data. In this approach, the inference result is obtained by applying the first model to the inference data when the degree of match between the training data and the inference data is high, thereby improving the generalization of the first model.

[0018] In another possible implementation, when the first device is a terminal device and the degree of match between the training data and the inference data is less than or equal to a first threshold, the terminal device may further perform the following steps: sending updated training data to the network device; receiving a second model from the network device, where the second model is obtained by the network device training the first model using the updated training data; and obtaining an inference result based on the second model and the inference data.

[0019] As can be seen, when the degree of match between the training data and the inference data is low, the terminal device sends the updated training data to the network device, causing the network device to update the first model. The terminal device then determines the inference result based on the second model updated by the network device based on the first model and the inference data. In this approach, the second model used by the terminal device during inference is obtained by updating the first model based on the updated training data when the degree of match between the training data and the inference data is low. Thus, the terminal device applies the second model to the inference data, which can improve the generalization of the second model and the accuracy of the inference result.

[0020] In another possible implementation, when the first device is a terminal device and the degree of match between the training data and the inference data is less than or equal to a first threshold, the terminal device may further perform the following steps: sending a model update request message for updating the first model to the network device; receiving a third model from the network device, where the third model is a model after the network device updates the first model; and obtaining an inference result based on the third model and the inference data.

[0021] As can be seen, when the degree of match between the training data and the inference data is low, the terminal device requests the network device to update the first model and obtains an inference result based on the third model and the inference data after the network device updates the first model. In this approach, the third model used by the terminal device during inference is the model obtained by the network device when the degree of match between the training data and the inference data is low. Therefore, the terminal device applies the third model to the inference data, which can improve the generalization of the third model and the accuracy of the inference result.

[0022] In one possible implementation, when the first device is a terminal device, before the terminal device determines the matching degree corresponding to each feature based on the distribution information of each feature in the training data and the distribution information in the inference data, the terminal device may also perform the following steps: determine a first parameter, the first parameter including one or more of the following: the number of features included in the training data and the inference data, the type of each feature in the multiple features, the value range of each feature, the sampling interval of each feature, and the probability density codebook of each feature; based on the first parameter, determine the distribution information of each feature in the training data and the distribution information in the inference data.

[0023] In a possible implementation, the first parameter further includes one or more of the following: a representation of the degree of matching between the training data and the inference data, a weight of the degree of matching corresponding to each feature, and a first threshold.

[0024] In another possible implementation, the first device is a network device, meaning that the method is performed by the network device. In this case, when the degree of match between the training data and the inference data exceeds a first threshold, the network device also sends the first model to the terminal device, causing the terminal device to apply the first model to the inference data. In this approach, the network device causes the terminal device to apply the first model to the inference data when the degree of match between the training data and the inference data is high, thereby improving the generalization of the first model.

[0025] In another possible embodiment, the first device is a network device, and when the matching degree between the training data and the inference data is less than or equal to a first threshold, the network device further performs the following steps: sending indication information to the terminal device indicating that the matching degree between the training data and the inference data is less than or equal to the first threshold; receiving updated training data from the terminal device; updating the first model based on the updated training data to obtain a second model; and sending the second model to the terminal device.

[0026] As can be seen, when the degree of match between the training data and the inference data is low, the network device indicates this to the terminal device, prompting the terminal device to report updated training data. Consequently, the network device updates the first model based on the updated training data reported by the terminal device, obtaining a second model. Transmitting the second model to the terminal device facilitates the terminal device's use of the more generalizable second model for inference, thereby improving the accuracy of the inference results.

[0027] In another possible implementation, the first device is a network device, and the matching degree between the training data and the inference data is less than or equal to the first threshold. The network device may also perform the following steps: updating the first model to obtain a third model; and sending the third model to the terminal device.

[0028] As can be seen, when the degree of match between the training data and the inference data is low, the network device can automatically update the first model to obtain a third model and send the third model to the terminal device. This helps the terminal device use the third model with better generalization for inference, thereby improving the accuracy of the inference results.

[0029] In one possible implementation, when the first device is a network device, the network device also receives distribution information of the inference data from the terminal device before determining the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0030] In a second aspect, the present application further provides a communication device. The communication device has the ability to implement some or all of the functions described in any of the first aspects above. For example, the functions of the communication device may include the functions of some or all of the embodiments described in the first aspect of the present application, or may include the functions of implementing any one of the embodiments of the present application alone. The functions may be implemented by hardware, or by hardware executing corresponding software implementations. The hardware or software includes one or more units or modules corresponding to the above functions.

[0031] In one possible design, the communication device may include a processing unit and a communication unit. The processing unit is configured to support the communication device in performing the corresponding functions of the above method. The communication unit is configured to support the transmission and reception of signals. The communication device may also include a storage unit, coupled to the processing unit and the communication unit, which stores program instructions and data necessary for the communication device.

[0032] In one embodiment, the communication device includes a communication unit and a processing unit, wherein the communication unit is used to send and receive signals / signaling;

[0033] The processing unit is configured to determine a degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data;

[0034] The processing unit is further configured to determine whether a first model can be applied to the inference data based on a degree of matching between the training data and the inference data, wherein the first model is obtained by training using the training data.

[0035] In addition, in this aspect, the communication device can also execute the methods in various possible implementations of the first aspect above. For other possible implementations, please refer to the relevant content of the first aspect above, which will not be described in detail here.

[0036] As an example, the communication unit may be a transceiver, the storage unit may be a memory, and the processing unit may be a processor.

[0037] In one embodiment, the device includes a transceiver and a processor, wherein the transceiver is used to transmit and receive signals / signaling;

[0038] The processor is configured to determine a degree of matching between the training data and the inference data based on distribution information of the training data and distribution information of the inference data;

[0039] The processor is further configured to determine whether a first model can be applied to the inference data based on a degree of matching between the training data and the inference data, wherein the first model is obtained by training using the training data.

[0040] In addition, in this aspect, other possible implementations of the communication device can refer to the relevant content of the first aspect above and will not be described in detail here.

[0041] In another embodiment, the communication device is a chip or a chip system. The processing unit may also be embodied as a processing circuit or a logic circuit; and the transmitting unit may be an input / output interface, an interface circuit, an output circuit, an input circuit, a pin, or a related circuit on the chip or chip system.

[0042] During implementation, the processor can be used to perform, for example, but not limited to, baseband-related processing, and the transceiver can be used to perform, for example, but not limited to, radio frequency transceiver. The above-mentioned devices can be respectively arranged on independent chips, or at least partially or entirely arranged on the same chip. For example, the processor can be further divided into an analog baseband processor and a digital baseband processor. Among them, the analog baseband processor can be integrated with the transceiver on the same chip, and the digital baseband processor can be arranged on a separate chip. With the continuous development of integrated circuit technology, more and more devices can be integrated on the same chip. For example, a digital baseband processor can be integrated with multiple application processors (such as, but not limited to, a graphics processor, a multimedia processor, etc.) on the same chip. Such a chip can be called a system on a chip (SoC). Whether each device is independently arranged on different chips or integrated on one or more chips often depends on the needs of product design. The embodiments of the present application do not limit the implementation form of the above-mentioned devices.

[0043] In a third aspect, the present application also provides a processor for executing the various methods described above. In the process of executing these methods, the process of sending the above-mentioned information and receiving the above-mentioned information in the above-mentioned methods can be understood as the process of the processor outputting the above-mentioned information and the process of the processor receiving the above-mentioned information input. When outputting the above-mentioned information, the processor outputs the above-mentioned information to the transceiver so that the transceiver can transmit it. After being output by the processor, the above-mentioned information may also need to undergo other processing before reaching the transceiver. Similarly, when the processor receives the above-mentioned information input, the transceiver receives the above-mentioned information and inputs it into the processor. Furthermore, after the transceiver receives the above-mentioned information, the above-mentioned information may need to undergo other processing before being input into the processor.

[0044] For the sending and receiving operations involved in the processor, unless otherwise specified, or unless they conflict with their actual functions or internal logic in the relevant descriptions, they can be more generally understood as processor output, reception, input and other operations, rather than sending and receiving operations directly performed by the RF circuit and antenna.

[0045] During implementation, the processor may be a processor specifically configured to execute these methods, or may be a processor that executes computer instructions in a memory to execute these methods, such as a general-purpose processor. The memory may be a non-transitory memory, such as a read-only memory (ROM), which may be integrated with the processor on the same chip or disposed on separate chips. The embodiments of the present application do not limit the type of memory or the configuration of the memory and the processor.

[0046] Fourthly, the present application further provides a communication system, which includes a network device and a terminal device. In another possible design, the system may further include other devices that interact with the network device and the terminal device. The first device in the embodiments of the present application is a terminal device in the system, or a network device in the system, which is not limited in the embodiments of the present application.

[0047] In a fifth aspect, the present application provides a computer-readable storage medium for storing instructions, which, when executed by a computer, implements the method described in any one of the first aspects above.

[0048] In a sixth aspect, the present application also provides a computer program product comprising instructions, which, when executed on a computer, implements the method described in any one of the first aspects above.

[0049] In the seventh aspect, the present application provides a chip system, which includes a processor and an interface, wherein the interface is used to obtain a program or instruction, and the processor is used to call the program or instruction to implement or support the first device to implement the functions involved in the first aspect. For example, determining or processing at least one of the data and information involved in the above method. In one possible design, the chip system also includes a memory, which is used to store program instructions and data necessary for the terminal. The chip system can be composed of a chip, or it can include a chip and other discrete devices.

[0050] In an eighth aspect, the present application provides a communication device comprising a processor for executing a computer program or executable instructions stored in a memory, wherein when the computer program or executable instructions are executed, the device executes a method in any possible implementation of the first aspect.

[0051] In one possible implementation, the processor and the memory are integrated together; in another possible implementation, the memory is located outside the communication device.

[0052] The beneficial effects of the second to eighth aspects can refer to the beneficial effects of the first aspect and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a schematic diagram of a system architecture;

[0054] Figure 2 is a schematic diagram of another system architecture;

[0055] FIG3 is a schematic diagram of fitting of training data and a model;

[0056] FIG4 is a flow chart of a communication method provided in an embodiment of the present application;

[0057] FIG5 is an interactive diagram of a communication method provided in an embodiment of the present application;

[0058] FIG6 is a schematic diagram of a probability density distribution function of a feature provided in an embodiment of the present application;

[0059] FIG7 is an interactive diagram of another communication method provided in an embodiment of the present application;

[0060] FIG8 is an interactive diagram of another communication method provided in an embodiment of the present application;

[0061] FIG9 is an interactive diagram of another communication method provided in an embodiment of the present application;

[0062] FIG10 is a schematic structural diagram of a communication device provided in an embodiment of the present application;

[0063] FIG11 is a schematic structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application are described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0065] In order to better understand the communication method disclosed in the embodiment of the present application, a communication system applicable to the embodiment of the present application is described.

[0066] The embodiments of the present application can be applied to systems evolved after 5G, such as long-term evolution (LTE) systems, fifth-generation (5G) mobile communication systems, and sixth-generation (6G) mobile communication systems, as well as satellite communications and short-range wireless communication systems. The system architecture is shown in Figure 1. A wireless communication system may include one or more network devices and one or more terminal devices. A wireless communication system may also perform point-to-point communication, such as communication between multiple terminal devices.

[0067] It can be understood that the wireless communication systems mentioned in the embodiments of the present application include but are not limited to: narrowband Internet of Things (NB-IoT) system, long term evolution (LTE) system, three major application scenarios of 5G / 6G mobile communication systems: enhanced mobile broadband (eMBB), ultra reliable low latency communication (URLLC) and massive machine type of communication (mMTC), wireless fidelity (WiFi) system, etc.

[0068] In an embodiment of the present application, the network device is a device with wireless transceiver functions, which is used to communicate with a terminal device. It can be an evolved base station (evolved Node B, eNB or eNodeB) in LTE, or a base station in a 5G / 6G network or a base station in a future evolved public land mobile network (public land mobile network, PLMN), a broadband network service gateway (BNG), an aggregation switch or a non-third generation partnership project (3GPP) access device, etc. Optionally, the network devices in the embodiments of the present application may include various forms of base stations, such as: macro base stations, micro base stations (also known as small stations), relay stations, access points, devices that realize base station functions in the future, access points (AP) in WiFi systems, transmitting and receiving points (TRP), transmitting points (TP), mobile switching centers, and devices that perform base station functions in device-to-device (D2D), vehicle-to-everything (V2X), and machine-to-machine (M2M) communications, devices that realize base station functions in communication systems evolved after 5G, integrated access and backhaul (IAB), and may also include centralized units (CU) and distributed units (DU) in cloud radio access networks (C-RAN) systems, non-terrestrial communication networks (non-terrestrial) networks, and mobile switching centers. The network equipment in the NTN communication system can be deployed on a high-altitude platform or a satellite, or can be various devices constituting an access node, such as an active antenna unit (AAU) and a baseband unit (BBU), etc., which are not specifically limited in the embodiments of the present application.

[0069] Network devices can communicate and interact with core network devices to provide communication services to terminal devices. Core network devices are, for example, devices in the 5G core network (CN). As a bearer network, the core network provides an interface to the data network, providing communication connections, authentication, management, policy control, and data service delivery for terminals.

[0070] In the embodiments of the present application, the terminal device may include various handheld devices, vehicle-mounted devices, wearable devices, computing devices, or other processing devices connected to a wireless modem with wireless communication capabilities. The terminal device may also be referred to as a terminal. Terminal equipment may also refer to user equipment (UE), access terminal, subscriber unit, user agent, cellular phone, smart phone, wireless data card, personal digital assistant (PDA), tablet computer, wireless modem, handheld device, laptop computer, smart point of sale (POS), customer-premises equipment (CPE), machine type communication (MTC) terminal, communication equipment carried by high-altitude aircraft, wearable device, drone, robot, terminal in D2D, terminal in vehicle to everything (V2X), virtual reality (VR) terminal device, augmented reality (AR) terminal device, wireless terminal in industrial control, wireless terminal in self-driving, wireless terminal in remote medical, wireless terminal in smart grid, wireless terminal in transportation security, etc. This application does not limit the wireless terminals in smart cities, smart homes, or terminal devices in future communication networks.

[0071] Please refer to Figure 2, which is another schematic diagram of a system architecture. The embodiment of the present application can be applied to a network between a base station and a terminal device as shown in Figure 2, or to a network between base stations as shown in Figure 2, or to a network between terminal devices as shown in Figure 2. Among them, the base stations can communicate with each other through a backhaul link, which can be a wired backhaul link (such as optical fiber, copper cable) or a wireless backhaul link (such as microwave). The terminal device can communicate with the corresponding base station through a wireless link.

[0072] Base stations are used to provide wireless access services to terminal devices. Specifically, each base station corresponds to a service coverage area (also called a cell, as shown in the elliptical area in Figure 2). Terminal devices entering this area communicate with the base station via wireless signals to receive the wireless access services provided by the base station. The service coverage areas of base stations may overlap, and terminal devices in the overlapping areas can receive wireless signals from multiple base stations, so multiple base stations can provide services to the terminal device simultaneously.

[0073] In the embodiments of the present application, the first device is the aforementioned network device, or the aforementioned terminal device, which is not limited in the embodiments of the present application. For example, the first device is a terminal device in the system architecture shown in Figure 1 or Figure 2. For another example, the first device is a network device in the system architecture shown in Figure 1 or Figure 2.

[0074] In the embodiments of the present application, the terminal device may also have artificial intelligence (AI) processing capabilities or machine learning (ML) capabilities, and the network device may also have AI processing capabilities or ML capabilities. For example, the terminal device or network device may have neural network training capabilities, reasoning capabilities, etc.

[0075] The embodiments disclosed herein will present various aspects, embodiments, or features of the present invention centered around a system comprising multiple devices, components, modules, etc. It should be understood that each system may include additional devices, components, modules, etc., and / or may not include all of the devices, components, modules, etc. discussed in conjunction with the accompanying drawings. Furthermore, combinations of these solutions may also be used.

[0076] Currently, 5G communications have introduced technologies such as eMBB, URLLC, and mMTC. Future 6G wireless communication systems will evolve towards higher throughput, lower latency, higher reliability, increased connections, and higher spectrum utilization. To meet the performance requirements of future wireless communications, artificial intelligence (AI) or machine learning (ML) is being researched in a variety of areas, including modeling complex and unknown environments, learning, channel prediction, intelligent signal generation and processing, network status tracking and intelligent scheduling, and network optimization and deployment.

[0077] Model learning can be categorized into online and offline learning. Online learning dynamically updates the model as data continues to arrive. Offline learning pre-trains the model using training data to obtain a model for inference. Online learning is real-time, quickly adapting to changes in the characteristics of new data and avoiding the need to regenerate the model in offline learning. However, online learning requires real-time computing, which can impact system performance and may not result in optimal model performance in the early stages of learning.

[0078] In offline learning, after training, the trained model is applied to inference data to obtain inference results. However, training data and inference data often differ. Even if the inference data initially matches the training data, the difference between the two data may increase over time. When there are significant differences between the training data and inference data, the model obtained based on the training data may match the training data well, but may not achieve good performance when applied to inference data, resulting in poor generalization performance.

[0079] The generalization of a model is defined as the performance of a model trained on the training data on inference data. If a trained model achieves good performance when applied to inference data, the model has good generalization. If a trained model does not achieve good performance when applied to inference data, the model has poor generalization. Alternatively, if a trained model produces high-accuracy inference results when applied to inference data, the model has good generalization; if a trained model produces low-accuracy inference results when applied to inference data, the model has poor generalization.

[0080] See Figure 3, which illustrates the fitting of training data to the model. As shown in Figure 3, in the initial stages of model training, when the model cannot fit the training data, the relationship between the model and the training data is called underfitting. As training progresses, the model gradually fits the training data. When the model overfits the training data, the relationship between the model and the training data becomes overfitting. In this overfitting state, the model overfits the training data. When there are discrepancies between the training data and the inference data, the model cannot fit the inference data. Using this model for inference results in decreased performance, meaning that the model's generalization is poor.

[0081] In the embodiment of the present application, the training data is the data used when training the model, and the inference data is the data used when actually using the model for inference.

[0082] In the embodiment of the present application, the degree of match between the training data and the inference data can represent the generalization of the model. It is understandable that when the degree of match between the training data and the inference data is large, it indicates that the difference between the training data and the inference data is small, or that the similarity between the training data and the inference data is large, or that the distance between the training data and the inference data is small. If the model obtained based on the training data and the inference data are used for inference, the accuracy of the inference result can be high, and thus the generalization of the model is good. When the degree of match between the training data and the inference data is small, it indicates that the difference between the training data and the inference data is large, or that the similarity between the training data and the inference data is small, or that the distance between the training data and the inference data is large. If the model obtained based on the training data and the inference data are used for inference, the accuracy of the inference result obtained is low, and thus the generalization of the model is poor.

[0083] It can be seen that the greater the matching degree between the training data and the inference data, the better the generalization of the model obtained based on the training data; the smaller the matching degree between the training data and the inference data, the worse the generalization of the model obtained based on the training data.

[0084] In the embodiment of the present application, the degree of matching between the training data and the inference data can also be replaced by one of the following, including but not limited to: the degree of adaptation between the training data and the inference data, the degree of conformity between the training data and the inference data, the consistency between the training data and the inference data, and the degree of harmony between the training data and the inference data. It can also be understood that the embodiment of the present application does not limit the naming of "the degree of matching between the training data and the inference data". When the degree of matching between the training data and the inference data is named by other names, it can also express the generalization of the model. The form of expressing the generalization of the model can refer to the form of expressing the generalization of the model when the name of matching is adopted, and will not be repeated here.

[0085] In the embodiments of the present application, the first model is an artificial intelligence (AI) model, or the first model is a machine learning (ML) model. The embodiments of the present application do not limit the specific form of the first model. For example, the first model may also be a model other than an AI model or an ML model.

[0086] In an embodiment of the present application, the first model can be applied to inference data, which can be understood as: the first device can use the first model and inference data to obtain an inference result with higher accuracy; or it can be understood as: the first device uses the inference data as the input of the first model, and can obtain an inference result with higher accuracy output by the first model.

[0087] An embodiment of the present application provides a communication method 100. In this communication method 100, a first device determines a degree of match between the training data and the inference data based on distribution information of the training data and distribution information of the inference data. Based on the degree of match between the training data and the inference data, the first device determines whether a first model obtained based on the training data can be applied to the inference data. This approach facilitates the first device determining that the first model can be applied to the inference data when the degree of match between the training data and the inference data is high, thereby improving the generalization of the first model.

[0088] The embodiment of the present application also takes the first device as a terminal device, the training data and the inference data both come from the terminal device, the terminal device determines the inference result, and the network device trains the model as an example, and proposes a communication method 200. In the communication method 200, the terminal device sends training data to the network device. The network device trains a model based on the training data to obtain a first model. The terminal device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data. When the degree of matching between the training data and the inference data is greater than a first threshold, the terminal device requests the network device to obtain the first model for inference, thereby determining the inference result based on the first model and the inference data sent by the network device. When the first model can be applied to the inference data, the terminal device applies the first model to the inference data to obtain the inference result, thereby improving the generalization of the first model.

[0089] Furthermore, when the degree of match between the training data and the inference data is less than or equal to a first threshold, the terminal device sends the updated training data to the network device, causing the network device to update the first model based on the updated training data. Consequently, the terminal device obtains an inference result based on the second model updated by the network device from the first model and the inference data. The second model used by the terminal device for inference is obtained by updating the first model based on the updated training data when the degree of match between the training data and the inference data is low, thereby ensuring the accuracy of the inference result.

[0090] This embodiment of the present application also proposes a communication method 300, assuming that the first device is a terminal device, the training data comes from a network device, the inference data comes from the terminal device, the terminal device determines the inference result, and the network device trains a model. This communication method 300 differs from communication method 200 in that the training data used to train the network device's model comes from the network device itself. Furthermore, when the degree of match between the training data and the inference data is less than or equal to a first threshold, the terminal device sends a model update request message to the network device, requesting the network device to update the first model. The network device then updates the first model to obtain a third model. This third model is obtained by updating the first model using the updated training data determined by the network device. The network device sends the third model to the terminal device. The terminal device obtains an inference result based on the third model and the inference data. The third model used by the terminal device for inference is obtained by updating the first model when the degree of match between the training data and the inference data is low, thereby ensuring the accuracy of the inference result.

[0091] This embodiment of the present application also proposes a communication method 400, assuming that the first device is a network device, both training data and inference data are sent to a terminal device, the terminal device determines the inference result, and the network device trains a model. This communication method 400 differs from communication method 200 in that the terminal device also sends distribution information of the inference data to the network device. Based on the distribution information of the training data and the inference data, the network device determines the degree of match between the training data and the inference data. When the degree of match between the training data and the inference data exceeds a first threshold, the network device sends the first model to the terminal device, causing the terminal device to apply the first model to the inference data, thereby improving the generalization of the first model. Furthermore, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the network device indicates to the terminal device that the degree of match between the training data and the inference data is low, causing the terminal device to report updated training data. Consequently, the network device updates the first model based on the updated training data reported by the terminal device, obtaining a second model. The terminal device obtains the inference result based on the second model and the inference data from the network device. This approach ensures the accuracy of the inference result.

[0092] The embodiment of the present application also proposes a communication method 500, taking the first device as a network device, the training data from the network device, the inference data from the terminal device, the terminal device determining the inference result, and the network device training model as an example. This communication method 500 differs from communication method 400 in that when the degree of match between the training data and the inference data is less than or equal to a first threshold, the network device automatically updates the first model to obtain a third model. The updated training data used by the network device to update the first model is determined by the network device itself, rather than reported by the terminal device. Thus, the terminal device obtains the inference result based on the third model and the inference data from the network device. This approach can ensure the accuracy of the inference result.

[0093] The present application embodiment provides a communication method 100. FIG4 is a flow chart of the communication method 100. The communication method 100 is described from the perspective of the first device. The communication method 100 includes but is not limited to the following steps:

[0094] S101. The first device determines a matching degree between the training data and the inference data based on distribution information of the training data and distribution information of the inference data.

[0095] The distribution information of the training data may be the probability distribution of the training data, and the distribution information of the inference data may be the probability distribution of the inference data. The embodiments of the present application do not limit the specific forms of the distribution information of the training data and the distribution information of the inference data.

[0096] In one optional implementation, the degree of match between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data. Alternatively, it can be understood that the first device determines the difference between the distribution information of the training data and the distribution information of the inference data as the degree of match between the training data and the inference data. Alternatively, it can be understood that the first device determines the difference between the distribution information of the training data and the distribution information of the inference data as the measurement criterion for the degree of match between the training data and the inference data.

[0097] The difference between the distribution information of the training data and the distribution information of the inference data indicates the degree of difference between the distribution information of the training data and the distribution information of the inference data. For example, the first device uses KL (Kullback–Leibler) divergence to determine the degree of match between the training data and the inference data.

[0098] For example, p(x i ) represents the probability density corresponding to the probability distribution P of the training data, q(x i) represents the probability density corresponding to the probability distribution Q of the inference data, and KL(p||q) represents the difference between the probability distribution of the training data and the probability distribution of the inference data. KL(p||q) is calculated as shown in formula (1):

[0099] It can be seen that the first device determines the difference between the probability distribution of the training data and the probability distribution of the inference data based on the probability density corresponding to the probability distribution P of the training data and the probability density corresponding to the probability distribution Q of the inference data, thereby determining the degree of matching between the training data and the inference data.

[0100] It can be understood that the smaller the value of KL(p||q), the smaller the KL divergence, the smaller the difference between the training data and the inference data, and thus the greater the matching degree between the training data and the inference data; the larger the value of KL(p||q), the larger the KL divergence, the greater the difference between the training data and the inference data, and thus the smaller the matching degree between the training data and the inference data.

[0101] In another optional embodiment, the degree of match between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data. Alternatively, it can be understood that the first device determines the similarity between the distribution information of the training data and the distribution information of the inference data as the degree of match between the training data and the inference data. Alternatively, it can be understood that the first device determines the similarity between the distribution information of the training data and the distribution information of the inference data as a measure of the degree of match between the training data and the inference data. The similarity between the distribution information of the training data and the distribution information of the inference data indicates the degree of similarity between the distribution information of the training data and the distribution information of the inference data.

[0102] For example, the first device determines the JS (Jensen–Shannon) divergence as the degree of match between the training data and the inference data. JS divergence is a variant based on KL divergence. JS divergence can solve the asymmetric problem of KL divergence, that is, KL(p||q)! = KL(q||p), while JS divergence is symmetric. In addition, the value of KL can be infinite, and the value of JS divergence is between 0 and 1, so JS divergence has a normalization effect. The calculation method of JS divergence is as shown in formula (2):

[0103] Where p represents the probability density corresponding to the probability distribution P of the training data, q represents the probability density corresponding to the probability distribution Q of the inference data, and JS(p||q) represents the similarity between the probability distribution P of the training data and the probability distribution Q of the inference data. is the difference between P and (P+Q) / 2, is the difference between Q and (P+Q) / 2.

[0104] It can be seen that the first device determines the similarity between the probability distribution of the training data and the probability distribution of the inference data based on the probability density corresponding to the probability distribution P of the training data and the probability density corresponding to the probability distribution Q of the inference data, thereby determining the matching degree between the training data and the inference data.

[0105] It can be understood that the smaller the value of JS(p||q), the smaller the JS divergence, the smaller the similarity between the training data and the inference data, and thus the smaller the matching degree between the training data and the inference data; the larger the value of JS(p||q), the larger the KL divergence, the greater the similarity between the training data and the inference data, and thus the greater the matching degree between the training data and the inference data.

[0106] In another optional embodiment, the degree of match between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data. Alternatively, it can be understood that the first device determines the distance between the distribution information of the training data and the distribution information of the inference data as the degree of match between the training data and the inference data. Alternatively, it can be understood that the first device determines the distance between the distribution information of the training data and the distribution information of the inference data as a measure of the degree of match between the training data and the inference data. The distance between the distribution information of the training data and the distribution information of the inference data indicates the proximity of the distribution information of the training data to the distribution information of the inference data.

[0107] For example, the first device determines the Wasserstein distance as the degree of match between the training data and the inference data. If the distribution information of the training data is far from the distribution information of the inference data, the KL divergence value is infinite and the JS divergence value is constant. In this case, both the KL divergence value and the JS divergence value are meaningless. In this case, even if there is no overlap between the distribution information of the training data and the distribution information of the inference data, the Wasserstein distance can still reflect the distance between the distribution information of the training data and the distribution information of the inference data, and thus the Wasserstein distance can be determined as the degree of match between the training data and the inference data.

[0108] The calculation method of Wasserstein distance is as follows: W(p,q)=(inf γ∈Γ(p,q) ∫‖pq‖ k dγ(p,q)) 1 / k (3)

[0109] Where p represents the probability density corresponding to the probability distribution P of the training data, q represents the probability density corresponding to the probability distribution Q of the inference data, W(p,q) represents the distance between the distribution information of the training data and the distribution information of the inference data, Γ(p,q) is the set of all possible joint distributions of the two distributions p and q, γ obeys the joint distribution Γ(p,q), ‖pq‖k is the k-order matrix of the two distributions p and q, ∫‖pq‖ k dγ(p,q) represents the integration of a k-order matrix, which can also be understood as the expectation of a k-order matrix. Inf represents the lower bound, and 1 / k represents the square root of the lower bound of the expectation of the k-order moment. Therefore, W(p,q) is the lower bound of the expectation of the distance ||pq|| under the joint distribution Γ(p,q), which is used to measure the distance between two distributions.

[0110] It can be seen that the first device can determine the difference, or similarity, or distance between the distribution information of the training data and the distribution information of the inference data based on the probability distribution of the training data and the probability distribution of the inference data, and determine the difference, or similarity, or distance between the distribution information of the training data and the distribution information of the inference data as the matching degree between the training data and the inference data.

[0111] In addition to the above-mentioned KL divergence, JS divergence, and Wasserstein distance, the first device can also use the Kolmogorov-Smirnov test, or the population stability index, or the Hellinger distance to determine the degree of matching between the training data and the inference data. Among them, the Kolmogorov-Smirnov test is a non-parametric test, which is often used to determine whether the sample is consistent with a predetermined distribution, or whether the probability distributions of two samples are different, that is, it can be used to determine whether the sample of the inference data conforms to the distribution of the training data, or to determine whether the distribution of the inference data is consistent with the distribution of the training data. The population stability index is used to judge the difference between the two distributions, that is, it can be used to judge the difference between the distribution of the inference data and the distribution of the training data. The Hellinger distance is used to determine the distance between the two distributions, that is, it can be used to determine the distance between the distribution of the inference data and the distribution of the training data.

[0112] In one optional embodiment, the training data and the inference data include a single feature. In this case, the distribution information of the training data is the distribution information of the feature in the training data, and the distribution information of the inference data is the distribution information of the feature in the inference data. Thus, the first device can directly determine the degree of match between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0113] In another optional embodiment, the training data and the inference data include multiple features. In this case, the first device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, including: determining the degree of matching corresponding to each feature based on the distribution information of each feature in the training data and the distribution information of each feature in the inference data, where the degree of matching corresponding to each feature is the degree of matching between the training data and the inference data; and determining the degree of matching between the training data and the inference data based on the degree of matching corresponding to each feature.

[0114] The implementation method for determining the degree of matching corresponding to each feature of the first device based on the distribution information of each feature in the training data and the distribution information of each feature in the inference data can be referred to the implementation method for determining the degree of matching between the training data and the inference data described above, and will not be repeated here. For example, the first device determines the KL divergence corresponding to each feature of the first device based on the distribution information of each feature in the training data and the distribution information of each feature in the inference data, and the determined KL divergence is used to represent the degree of matching between the training data and the inference data.

[0115] In addition, the first device determines the degree of match between the training data and the inference data based on the degree of match corresponding to each feature, including determining the degree of match between the training data and the inference data based on the degree of match corresponding to each feature and the weight of the degree of match corresponding to each feature. In other words, the degree of match between the training data and the inference data is the weighted sum of the degree of match corresponding to multiple features.

[0116] For example, the training data and the inference data are channel data, and the first device determines the distribution information of the angle and amplitude in the training data, as well as the distribution information of the angle and amplitude in the inference data. The first device determines the degree of matching of the angle between the training data and the inference data based on the angle distribution information in the training data and the angle distribution information in the inference data; the first device determines the degree of matching of the amplitude between the training data and the inference data based on the amplitude distribution information in the training data and the amplitude distribution information in the inference data. The first device determines the degree of matching between the training data and the inference data based on the degree of matching of the angle between the training data and the inference data and the degree of matching of the amplitude between the training data and the inference data, as well as the weight of the degree of matching of the angle between the training data and the inference data and the weight of the degree of matching of the amplitude between the training data and the inference data.

[0117] For another example, the training data and the inference data include feature X, feature Y, and feature Z. The first device determines the distribution information of feature X, feature Y, and feature Z in the training data and the distribution information of feature Z in the inference data, respectively. The first device determines the matching degree X between feature X in the training data and the inference data based on the distribution information of feature X, feature Y, and feature Z in the training data and the distribution information of feature Z in the inference data, respectively. KL , the matching degree Y between the feature Y training data and the inference data KL , the matching degree Z of feature Z between training data and inference data KL In addition, the weights of the matching degrees corresponding to feature X, feature Y, and feature Z are w respectively. X , w Y , w Z , then the first device determines the matching degree between the training data and the inference data as: w X *X KL +w Y *Y KL +w Z *Z KL (4)

[0118] It can be seen that if the training data and the inference data include multiple features, the first device considers the matching degree of each feature between the training data and the inference data when determining the matching degree between the training data and the inference data, thereby improving the accuracy of the matching degree between the training data and the inference data.

[0119] In one optional embodiment, the first device determines a first parameter before determining the degree of match between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data. The first device determines the distribution information of each feature in the training data and the distribution information of each feature in the inference data based on the first parameter.

[0120] Among them, the first parameter includes one or more of the following: the number of features included in the training data and the inference data, the type of each feature in the multiple features, the value range of each feature, the sampling interval of each feature, and the probability density codebook of each feature.

[0121] It is understood that when the first device is a terminal device, the terminal device determines the first parameter, which may be received by the terminal device from the network device, or the terminal device may determine the first parameter on its own. When the first device is a network device, the network device determines the first parameter, which may be understood as the network device determining the first parameter on its own.

[0122] After determining the first parameter, the first device determines, based on the first parameter, the distribution information of each feature in the training data and the distribution information of each feature in the inference data. For example, when the first parameter includes a value range and a sampling interval, the first device assumes that the training data and the inference data include one feature, and uniformly samples the training data and the inference data by default, to obtain a probability density distribution of the training data and a probability density distribution of the inference data. Then, based on the probability density distribution of the training data and the probability density distribution of the inference data, the first device determines the degree of match between the training data and the inference data.

[0123] For another example, when the first parameter includes a value range, a sampling interval and a probability density codebook, the first device defaults to that the training data and the inference data include a feature, and the first device defaults to performing non-uniform sampling on the training data and the inference data based on the value range, the sampling interval and the probability density codebook to obtain the probability density distribution of the training data and the probability density distribution of the inference data, and then determine the degree of matching between the training data and the inference data based on the probability density distribution of the training data and the probability density distribution of the inference data.

[0124] For another example, if the first parameter includes the number of features in the training data and inference data being 2, and the value range of each of the two features and the sampling interval of each feature, the first device uniformly samples each feature based on the value range and sampling interval of each feature to obtain a probability density distribution of each of the two features in the training data and inference data. Furthermore, the first device determines the matching degree corresponding to each feature based on the probability density distribution of each feature in the training data and inference data, and then determines the matching degree between the training data and inference data based on the matching degree corresponding to each feature.

[0125] It is understandable that when the first device determines the KL divergence or JS divergence as the degree of match between the training data and the inference data, it is necessary to calculate the probability density difference of the same numerical value in the training data and the inference data. If the value range and sampling frequency used are inconsistent, invalid KL divergence or JS divergence may result. Therefore, the first parameter used by the first device to determine the distribution information of each feature in the training data is the same as the first parameter used to determine the distribution information of each feature in the inference data. It can also be understood that the first device needs to align the first parameter used to determine the distribution information of each feature in the training data and the first parameter used to determine the distribution information of each feature in the inference data.

[0126] S102. The first device determines whether a first model can be applied to the inference data based on a degree of matching between the training data and the inference data, where the first model is obtained by training using the training data.

[0127] It can be understood that the first device determines whether the first model can be applied to the inference data based on the matching degree between the training data and the inference data, including: when the matching degree between the training data and the inference data is greater than the first threshold, determining that the first model can be applied to the inference data; or when the matching degree between the training data and the inference data is less than or equal to the first threshold, determining that the first model cannot be applied to the inference data.

[0128] The first threshold may be preset or configured. For example, when the first device is a terminal device, the first threshold is configured by the network device for the terminal device. For another example, when the first device is a network device, the first threshold is pre-set by the network device.

[0129] In addition, when the degree of matching between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data, when the difference between the distribution information of the training data and the distribution information of the inference data is less than or equal to a first preset value, the degree of matching between the training data and the inference data is greater than a first threshold; when the difference between the distribution information of the training data and the distribution information of the inference data is greater than the first preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold. The first preset value can be preset or can be configured. For example, the first device determines the KL divergence as the degree of matching between the training data and the inference data, and when the KL divergence is less than or equal to the first preset value, the degree of matching between the training data and the inference data is greater than the first threshold; when the KL divergence is greater than the first preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold.

[0130] When the degree of matching between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data, when the similarity between the distribution information of the training data and the distribution information of the inference data is greater than a second preset value, the degree of matching between the training data and the inference data is greater than a first threshold; when the similarity between the distribution information of the training data and the distribution information of the inference data is less than or equal to the second preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold. The second preset value may be preset or may be configured. For example, the first device determines the JS divergence as the degree of matching between the training data and the inference data, and when the JS divergence is greater than the second preset value, the degree of matching between the training data and the inference data is greater than the first threshold; when the JS divergence is less than or equal to the second preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold.

[0131] When the degree of matching between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data, when the distance between the distribution information of the training data and the distribution information of the inference data is less than or equal to a third preset value, the degree of matching between the training data and the inference data is greater than a first threshold; when the distance between the distribution information of the training data and the distribution information of the inference data is greater than the third preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold. The third preset value can be preset or can be configured. For example, the first device determines the Wasserstein distance as the degree of matching between the training data and the inference data, and when the Wasserstein distance is less than or equal to the third preset value, the degree of matching between the training data and the inference data is greater than the first threshold; when the Wasserstein distance is greater than the third preset value, the degree of matching between the training data and the inference data is less than or equal to the first threshold.

[0132] It can be seen that when the matching degree between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data, the smaller the difference between the distribution information of the training data and the distribution information of the inference data, the greater the matching degree between the training data and the inference data; when the matching degree between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data, the greater the similarity between the distribution information of the training data and the distribution information of the inference data, the greater the matching degree between the training data and the inference data; when the matching degree between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data, the smaller the distance between the distribution information of the training data and the distribution information of the inference data, the greater the matching degree between the training data and the inference data.

[0133] Therefore, when the degree of match between the training data and the inference data is greater than the first threshold, it indicates that the difference between the distribution information of the training data and the distribution information of the inference data is small, or that the similarity between the distribution information of the training data and the distribution information of the inference data is large, or that the distance between the distribution information of the training data and the distribution information of the inference data is small. If the first model obtained based on the training data is applied to the inference data, a high-accuracy inference result can be obtained. Then, the first device determines that the first model can be applied to the inference data, which can improve the generalization of the first model. When the degree of match between the training data and the inference data is less than or equal to the first threshold, it indicates that the difference between the distribution information of the training data and the distribution information of the inference data is large, or that the similarity between the distribution information of the training data and the distribution information of the inference data is small, or that the distance between the distribution information of the training data and the distribution information of the inference data is large. If the first model obtained based on the training data is applied to the inference data, a high-accuracy inference result cannot be obtained, that is, the generalization of the first model is poor. Then, the first device determines that the first model cannot be applied to the inference data.

[0134] It can be seen that when the degree of match between the training data and the inference data is large, the first device determines that the first model can be applied to the inference data, which can improve the generalization of the first model; when the degree of match between the training data and the inference data is small, it determines that the first model cannot be applied to the inference data, which can ensure that the first model is not used when the generalization of the first model is poor.

[0135] In an embodiment of the present application, the first device determines the degree of match between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data. Based on the degree of match between the training data and the inference data, the first device determines whether a first model obtained based on the training data can be applied to the inference data. This approach is advantageous in that when the first device determines that the first model can be applied to the inference data when the degree of match between the training data and the inference data is high, it can improve the generalization of the first model.

[0136] Furthermore, the first device uses proactive lifecycle management to determine the generalizability of the first model by determining the degree of match between training data and inference data before applying the first model for inference. This avoids the latency and performance degradation caused by reactive lifecycle management and improves the efficiency of the first model's lifecycle management.

[0137] The present embodiment also proposes a communication method 200, taking the first device as a terminal device, the training data and inference data both coming from the terminal device, the terminal device determining the inference result, and the network device training the model as an example. FIG5 is an interaction diagram of the communication method 200. The communication method 200 is described from the perspective of the interaction between the terminal device and the network device. The communication method 200 includes but is not limited to the following steps:

[0138] S201. The terminal device sends training data to the network device. Correspondingly, the network device receives the training data from the terminal device.

[0139] It is understood that the terminal device determines training data and reports the training data to the network device. For example, if the terminal device receives a reference signal from the network device and the training data is the reference signal received by the terminal device, the terminal device reports the received reference signal to the network device. Furthermore, the terminal device sending training data to the network device can be understood as the terminal device sending distribution information of the training data to the network device.

[0140] In one optional implementation, the terminal device quantizes the distribution information of the training data, obtains the quantized distribution information, and then reports the quantized distribution information to the network device. This approach reduces the overhead of indicating the distribution information and improves system spectrum efficiency. Furthermore, the terminal device quantizing the distribution information of the training data can be understood as quantizing the distribution information of each feature included in the training data.

[0141] In an optional implementation, when the distribution information of a feature in the training data is a continuous value, the terminal device may quantize the distribution information of the feature using uniform or non-uniform sampling. If the terminal device quantizes the distribution information of the feature using uniform sampling, it is necessary to determine the value range and sampling interval of the feature. Therefore, the terminal device sends the value range and sampling interval of each feature in the training data to the network device. Based on the value range and sampling interval of each feature in the training data, the network device may determine the value sampling points for each feature in the training data.

[0142] If the terminal device uses non-uniform sampling to quantize the distribution information of the feature, in addition to determining the numerical range and sampling interval of the feature, it is also necessary to determine a bitmap used to indicate the sampling points. Therefore, the terminal device sends the numerical range of each feature in the training data, the sampling interval of each feature, and the bitmap used to indicate the sampling points of each feature to the network device. Based on the numerical range of each feature in the training data, the sampling interval of each feature, and the bitmap of each feature sampling point, the network device can determine the numerical sampling point of each feature in the training data for the terminal device.

[0143] For example, Figure 6 is a schematic diagram of the probability density distribution function of a feature, which is a feature in the training data. If the terminal device uniformly samples the values ​​of the feature in Figure 6, and the sampled values ​​are 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, then the sampling interval for the terminal device to sample the value of the feature is 2. In addition, as shown in Figure 6, the value range of the feature is 0 to 20. The terminal device sends the value range of the feature to the network device, which is 0 to 20, with a sampling interval of 2. The network device can determine that the terminal device's sampling points for the value of the feature are 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20.

[0144] If the terminal device performs non-uniform sampling on the values ​​of the feature in Figure 6, it must also use a bitmap to indicate the sampling points in addition to uniform sampling. For example, if the terminal device samples the values ​​0, 2, 8, 10, and 20 of the feature in Figure 6, it can use 1100 1100 001 to represent the sampling points. The position index of 1 in 1100 1100 001 corresponds to the sampling point in uniform sampling, that is, the uniform sampling point corresponding to the position index of 1 is the sampling point in non-uniform sampling. The terminal device sends the network device a value range of 0 to 20 for this feature, a sampling interval of 2, and a bitmap representing the sampling points of 1100 1100 001. The network device can then determine that the terminal device's sampling points for this feature are 0, 2, 8, 10, and 20.

[0145] Similarly, when the probability density of the feature in the training data is also a continuous value, the terminal device needs to quantize the probability density of the feature and determine the probability density codebook of the feature. For example, as shown in Figure 6, the numerical range of the probability density of the feature is 0 to 0.2, and the corresponding probability density codebook can be as shown in Table 1 below. The probability density codebook includes 11 codewords, and each codeword requires 4 bits for quantization. When the terminal device uses uniform sampling to sample the characteristic value of the probability density, and the sampling interval is 2, the probability density in Figure 6 requires a total of 11*4=44 bits; if the terminal device uses non-uniform sampling to sample the characteristic value of the probability density, when the bitmap is 1100 1100 001, the probability density in Figure 6 requires a total of 11+5*4=31 bits, where 5 indicates that there are a total of 5 1s in the bitmap. Therefore, the terminal device can send the probability density codebook of the feature to the network device, and the network device determines the sampling of the probability density of the feature by the terminal device based on the probability density codebook of the feature. The network device combines the numerical sampling of the feature by the terminal device and the sampling of the probability density of the feature to determine the probability density distribution of the feature.

[0146] Table 1

[0147] As can be seen, after the terminal device quantizes the distribution information of each feature in the training data using uniform or non-uniform sampling, it then sends the parameters used to quantize the distribution information of each feature when sending the training data to the network device. The network device then determines the distribution information of each feature based on the quantization parameters corresponding to the distribution information of each feature. Compared to directly sending the distribution information of each feature in the training data to the network device, this approach reduces the overhead of indicating the distribution information and improves the system spectrum efficiency.

[0148] S202. The network device trains a model based on the training data to obtain a first model.

[0149] The embodiment of the present application does not limit the method used by the network device to train the model. For example, the network device uses a gradient descent method to train the model based on the training data to obtain a first model.

[0150] S203. The terminal device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0151] It can be understood that the implementation method of the terminal device determining the matching degree between training data and inference data based on the distribution information of the training data and the distribution information of the inference data can be referred to in the above S101, where the first device determines the matching degree between training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, and no further details are given.

[0152] Optionally, before determining the degree of match between the training data and the inference data, the terminal device determines a first parameter based on the distribution information of the training data and the distribution information of the inference data; and determines the distribution information of each feature in the training data and the distribution information of each feature in the inference data based on the first parameter. The first parameter is described in S101 above and is not further described.

[0153] In addition, the terminal device determines the first parameter, including: the terminal device receives the first parameter from the network device. In other words, the first parameter may be configured by the network device to the terminal device.

[0154] Optionally, the terminal device determines the first parameter independently. For example, before sending training data to the network device, the terminal device determines the first parameter when quantizing the distribution information of each feature in the training data, such as the first parameter being the value range and sampling interval of the feature.

[0155] Optionally, after the terminal device determines the first parameter by itself, it sends the first parameter to the network device. The network device determines whether the first parameter determined by the terminal device by itself is reasonable. If the network device determines that the first parameter determined by the terminal device by itself is reasonable, it sends an indication message to the terminal device allowing the use of the first parameter, so that the terminal device adopts the first parameter to determine the distribution information of each feature in the training data and the distribution information in the inference data; or, the network device does not send an indication message to the terminal device allowing the use of the first parameter, and the terminal device adopts the first parameter by default to determine the distribution information of each feature in the training data and the distribution information in the inference data.

[0156] Optionally, if the network device determines that the first parameter determined by the terminal device is unreasonable, the network device sends information indicating the recommended first parameter to the terminal device, so that the terminal device adopts the first parameter recommended by the network device to determine the distribution information of each feature in the training data and the distribution information in the inference data.

[0157] Optionally, the first parameter also includes one or more of the following: a representation of the degree of match between the training data and the inference data, a weight assigned to the degree of match for each feature, and a first threshold. For example, the first parameter also includes a representation of the degree of match between the training data and the inference data, so that the terminal device determines the degree of match between the training data and the inference data based on the representation of the degree of match between the training data and the inference data indicated in the first parameter. For another example, the first parameter also includes a first threshold, so that the terminal device subsequently determines whether the first model can be applied to the inference data based on the first threshold.

[0158] In the embodiment of the present application, when the matching degree between the training data and the inference data is greater than the first threshold, the terminal device and the network device may execute the following S2041 to S2043:

[0159] S2041. When the degree of match between the training data and the inference data is greater than a first threshold, the terminal device sends an inference request message to the network device, where the inference request message is used to request acquisition of a first model for inference. In response, the network device receives the inference request message from the terminal device.

[0160] It is understandable that the implementation method of the matching degree between the training data and the inference data being greater than the first threshold can be referred to in the above S102 and will not be repeated here.

[0161] As described in S102 above, when the degree of match between the training data and the inference data is greater than a first threshold, the first model can be applied to the inference data. Therefore, when the degree of match between the training data and the inference data is greater than the first threshold, the terminal device sends an inference request message to the network device to request the first model for inference.

[0162] S2042: The network device sends the first model to the terminal device. Correspondingly, the terminal device receives the first model from the network device.

[0163] S2043. The terminal device obtains an inference result based on the first model and the inference data.

[0164] It is understandable that the terminal device inputs the inference data into the first model and obtains the inference result output by the first model.

[0165] As can be seen, when the degree of match between the training data and the inference data exceeds the first threshold, the terminal device requests the network device to obtain the first model for inference through inference request information, and then obtains the inference result based on the first model and inference data issued by the network device. When the degree of match between the training data and the inference data exceeds the first threshold, the terminal device applies the first model to the inference data, thereby improving the generalization of the first model and ensuring the accuracy of the inference result.

[0166] In another optional implementation, when the degree of matching between the training data and the inference data is less than or equal to the first threshold, the terminal device and the network device may execute the following S2051 to S2053:

[0167] S2051. When the matching degree between the training data and the inference data is less than or equal to the first threshold, the terminal device sends the updated training data to the network device. Correspondingly, the network device receives the updated training data from the terminal device.

[0168] It is understandable that the implementation method of the matching degree between the training data and the inference data is less than or equal to the first threshold can be referred to the above S102 and will not be repeated here.

[0169] As described in S102 above, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the first model cannot be applied to the inference data. Therefore, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device sends updated training data to the network device, causing the network device to update the first model based on the updated training data. This approach facilitates timely updates of the first model by the network device, thereby extending the lifecycle of the first model.

[0170] Optionally, before sending the updated training data to the network device, the terminal device determines a degree of match between the updated training data and the inference data based on the distribution information of the updated training data and the distribution information of the inference data. The implementation method for the terminal device to determine the degree of match between the updated training data and the inference data can be referenced to the implementation method for the first device to determine the degree of match between the training data and the inference data in S101, and will not be further described.

[0171] Optionally, the terminal device sends the updated training data to the network device, including: when the degree of match between the updated training data and the inference data is greater than a first threshold, the terminal device sends the updated training data to the network device. This approach enables the network device to update the first model based on the updated training data to obtain a model with better generalization.

[0172] S2052. The network device sends a second model to the terminal device, where the second model is obtained by training the first model using the updated training data. Correspondingly, the terminal device receives the second model from the network device.

[0173] The network device receives updated training data from the terminal device, trains the first model based on the updated training data to obtain a second model, and sends the second model to the terminal device so that the terminal device applies the second model to the inference data.

[0174] S2053. The terminal device obtains an inference result based on the second model and the inference data.

[0175] It is understandable that the terminal device applies the second model to the inference data, that is, uses the inference data as the input of the second model to obtain the inference result output by the second model.

[0176] As can be seen, if the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device sends the updated training data to the network device. The network device then updates the first model based on the updated training data to obtain a second model. Furthermore, the terminal device obtains an inference result based on the second model and the inference data.

[0177] The second model is obtained by the network device by updating the first model based on the updated training data. Therefore, the terminal device can obtain more accurate inference results by applying the second model to the inference data. Furthermore, updating the first model based on the updated training data can improve the model's generalizability and the efficiency of its lifecycle management.

[0178] Optionally, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device does not execute the above S2051 to S2053 and directly falls back to the non-AI method. The non-AI method can be understood as using a method different from the AI ​​method to determine the result of a certain function. Or it can be understood as: when the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device uses the method currently used in the communication system to obtain the result of a certain function. For example, the training data and the inference data are both reference signals received by the terminal device. If the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device directly performs channel estimation based on the received reference signal to obtain channel information.

[0179] In an embodiment of the present application, the terminal device directly determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, which can reduce the overhead of signaling interaction between the terminal device and the network device. In addition, when the degree of matching between the training data and the inference data is greater than a first threshold, the terminal device requests the network device to obtain a first model for inference, and then applies the first model to the inference data to obtain an inference result, which can improve the generalization of the first model. When the degree of matching between the training data and the inference data is less than or equal to the first threshold, the terminal device reports the updated training data to the network device, and obtains a second model that the network device updates based on the updated training data to the first model, so that the terminal device applies the second model to the inference data to obtain an inference result, which can improve the accuracy of the inference result.

[0180] The present embodiment also proposes a communication method 300, taking the first device as a terminal device, the training data from the network device, the inference data from the terminal device, the terminal device determining the inference result, and the network device training model as an example. FIG7 is an interaction diagram of the communication method 300. The communication method 300 is also described from the perspective of the interaction between the terminal device and the network device. The communication method 300 includes but is not limited to the following steps:

[0181] S301. The network device trains a model based on the training data to obtain a first model.

[0182] The training data may be historical data previously collected by the network device. Furthermore, the present embodiment does not limit the method used by the network device to train the model. For example, the network device may use a gradient descent method to train the model based on the training data to obtain a first model.

[0183] S302. The network device sends training data to the terminal device.

[0184] Optionally, the network device quantizes the distribution information of each feature in the training data to obtain the quantized distribution information of each feature. The network device sends the quantized distribution information of each feature to the terminal device to reduce indication overhead and improve system spectrum efficiency. The implementation of the network device quantizing the distribution information of each feature in the training data is similar to the implementation of the terminal device quantizing the distribution information of each feature in the training data in S201 above, and is not further described.

[0185] S303. The terminal device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0186] It is understandable that the implementation of S303 can refer to the implementation of S203 above and will not be repeated here.

[0187] In the embodiment of the present application, when the matching degree between the training data and the inference data is greater than the first threshold, the terminal device and the network device may execute the following S3041 to S3043:

[0188] S3041. When the degree of match between the training data and the inference data is greater than a first threshold, the terminal device sends an inference request message to the network device, where the inference request message is used to request acquisition of a first model for inference. In response, the network device receives the inference request message from the terminal device.

[0189] S3042. The network device sends the first model to the terminal device. Correspondingly, the terminal device receives the first model from the network device.

[0190] S3043. The terminal device obtains an inference result based on the first model and the inference data.

[0191] It can be understood that the implementation of S3041 to S3043 can refer to the implementation of S2041 to S2043 mentioned above and will not be repeated here.

[0192] In another optional implementation, when the degree of matching between the training data and the inference data is less than or equal to the first threshold, the terminal device and the network device may execute the following S3051 to S3053:

[0193] S3051. When the matching degree between the training data and the inference data is less than or equal to the first threshold, the terminal device sends a model update request message for updating the model to the network device. Correspondingly, the network device receives the model update request message for updating the model from the terminal device.

[0194] It is understandable that the implementation method of the matching degree between the training data and the inference data is less than or equal to the first threshold can be referred to the above S102 and will not be repeated here.

[0195] As described in S102 above, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the first model cannot be applied to the inference data. Therefore, the terminal device sends a model update request message to the network device for updating the model, causing the network device to update the model used for inference, i.e., causing the network device to update the first model.

[0196] S3052. The network device sends a third model to the terminal device, where the third model is obtained by training the first model using the updated training data. Correspondingly, the terminal device receives the second model from the network device.

[0197] After receiving a model update request from a terminal device, the network device learns that it needs to update the first model used for inference. Consequently, the network device independently determines updated training data. This updated training data is not reported to the network device by the terminal device but rather originates from the network device. For example, the network device may use data collected during a period closest to the current time as updated training data. The network device then uses the updated training data to update the first model, obtaining a third model. The network device then sends the third model to the terminal device, allowing the terminal device to determine an inference result based on the third model.

[0198] Optionally, before using the updated training data to update the first model, the network device determines the degree of match between the updated training data and the inference data. The determination method can refer to the first device determining the degree of match between the training data and the inference data in S101. Thus, when the degree of match between the updated training data and the inference data exceeds a first threshold, the network device updates the first model based on the updated training data. This approach can ensure that the second model, which is the result of the network device updating the first model, has better generalization.

[0199] S3053. The terminal device obtains an inference result based on the third model and the inference data.

[0200] After receiving the third model from the network device, the terminal device applies the third model to the inference data to obtain an inference result. In other words, the terminal device uses the inference data as input to the third model and obtains an inference result output by the third model.

[0201] When the terminal device determines that the third model used in the inference result is a model obtained by the network device updating the first model using the updated training data when the degree of match between the training data and the inference data is less than or equal to the first threshold. Thus, the terminal device applies the third model to the inference data, ensuring a high degree of accuracy in the inference result.

[0202] Optionally, when the matching degree between the training data and the inference data is less than or equal to the first threshold, the terminal device does not execute the above S3051 to S3053 and directly falls back to the non-AI mode. The specific implementation method can be found in the above communication method 200 and will not be repeated here.

[0203] In an embodiment of the present application, the training data used by the network device training model comes from the network device itself, and the terminal device determines the degree of match between the training data and the inference data. When the degree of match between the training data and the inference data is greater than a first threshold, the terminal device requests the network device to obtain a first model for inference, and then obtains an inference result based on the first model and the inference data, thereby improving the generalization of the first model. When the degree of match between the training data and the inference data is less than or equal to the first threshold, the terminal device sends a model update request message for updating the model to the network device, thereby obtaining a third model after the network device updates the first model, and then applies the third model to the inference data to obtain an inference result, thereby improving the accuracy of the inference result.

[0204] The present embodiment also takes the first device as a network device, the training data and inference data both come from a terminal device, the terminal device determines the inference result, and the network device trains the model as an example, and proposes a communication method 400. FIG8 is an interaction diagram of the communication method 400. The communication method 400 is also described from the perspective of the interaction between the terminal device and the network device. The communication method 400 includes but is not limited to the following steps:

[0205] S401. The terminal device sends training data to the network device. Correspondingly, the network device receives the training data from the terminal device.

[0206] S402. The network device trains a model based on the training data to obtain a first model.

[0207] It is understandable that the implementation of S401 and S402 can refer to the above-mentioned S201 and S202 and will not be repeated here.

[0208] S403. The terminal device sends a model inference request and distribution information of inference data to the network device. Correspondingly, the network device receives the model inference request and distribution information of inference data from the terminal device.

[0209] The model inference request is used to request inference based on the model. Furthermore, the terminal device can quantize the distribution information of each feature in the inference data, obtain the quantized distribution information of each feature, and report the quantized distribution information of each feature to the network device. This approach can reduce the indication overhead of the distribution information and improve the system spectrum efficiency. The implementation method of the terminal device quantizing the distribution information of each feature in the inference data can be referred to the implementation method of the terminal device quantizing the distribution information of each feature in the training data in S201 above, and will not be repeated here.

[0210] S404. The network device determines the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0211] As can be understood, the network device receives a model inference request from a terminal device and learns that the terminal device requests inference based on the model. Consequently, the network device determines the degree of match between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data. Furthermore, based on the degree of match between the training data and the inference data, the network device determines whether the first model is applicable to the inference data.

[0212] Among them, the implementation method of the network device determining the matching degree between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data can be referred to the above S102, where the first device determines the matching degree between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, and no further details are given.

[0213] In the embodiment of the present application, when the matching degree between the training data and the inference data is greater than the first threshold, the terminal device and the network device may execute the following S4051 and S4052:

[0214] S4051. When the matching degree between the training data and the inference data is greater than a first threshold, the network device sends the first model to the terminal device. Correspondingly, the terminal device receives the first model from the network device.

[0215] It is understandable that the implementation method of the matching degree between the training data and the inference data being greater than the first threshold can be referred to in the above S102 and will not be repeated here.

[0216] As described in S102 above, when the degree of match between the training data and the inference data is greater than a first threshold, the first model can be applied to the inference data. Therefore, when the degree of match between the training data and the inference data is greater than the first threshold, the network device sends the first model to the terminal device, so that the terminal device performs inference based on the first model.

[0217] S4052. The terminal device obtains an inference result based on the first model and the inference data.

[0218] It is understandable that the terminal device applies the first model to the inference data, that is, uses the inference data as the input of the first model to obtain the inference result output by the first model.

[0219] As can be seen, when the degree of match between the training data and the inference data exceeds the first threshold, the network device sends the first model applicable to the inference data to the terminal device. The terminal device then obtains an inference result based on the first model and the inference data. When the network device determines that the first model is applicable to the inference data, the terminal device applies the first model to the inference data, thereby improving the generalization of the first model and ensuring the accuracy of the inference result.

[0220] In another optional implementation, when the degree of matching between the training data and the inference data is less than or equal to the first threshold, the terminal device and the network device may execute the following S4061 to S4065:

[0221] S4061. When the degree of match between the training data and the inference data is less than or equal to a first threshold, the network device transmits, to the terminal device, indication information indicating that the degree of match between the training data and the inference data is less than or equal to the first threshold. In response, the terminal device receives, from the network device, indication information indicating that the degree of match between the training data and the inference data is less than or equal to the first threshold.

[0222] As described in S102 above, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the first model cannot be applied to the inference data. Therefore, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the network device sends an indication to the terminal device indicating that the degree of match between the training data and the inference data is less than or equal to the first threshold, thereby informing the terminal device that the degree of match between the training data and the inference data is low. This approach facilitates the terminal device reporting updated training data to the network device.

[0223] S4062. The terminal device sends the updated training data to the network device. Correspondingly, the network device receives the updated training data from the terminal device.

[0224] After receiving an indication from the network device indicating that the degree of match between the training data and the inference data is less than or equal to a first threshold, the terminal device determines that the first model cannot be applied to the inference data. To improve the lifecycle of the first model, the terminal device sends updated training data to the network device, causing the network device to update the first model based on the updated training data.

[0225] Optionally, before sending the updated training data to the network device, the terminal device determines the degree of match between the updated training data and the inference data based on the distribution information of the updated training data and the distribution information of the inference data. The method for determining the degree of match between the training data and the inference data by the first device can be found in S101 above and will not be further described. The terminal device sends the updated training data to the network device, including: when the degree of match between the updated training data and the inference data is greater than a first threshold, sending the updated training data to the network device. This method enables the network device to update the first model based on the updated training data, obtaining a model with better generalization.

[0226] S4063. The network device updates the first model based on the updated training data to obtain a second model.

[0227] S4064. The network device sends the second model to the terminal device. Correspondingly, the terminal device receives the second model from the network device.

[0228] S4065. The terminal device obtains an inference result based on the second model and the inference data.

[0229] After obtaining the updated training data, the network device updates the first model based on the updated training data to obtain a second model. The network device then sends the second model to the terminal device. The terminal device then obtains an inference result based on the second model and the inference data. The second model is obtained by the network device based on the updated training data. Therefore, the terminal device can improve the generalization of the second model by applying the second model to the inference data.

[0230] In an embodiment of the present application, both the training data and the inference data come from a terminal device, and the network device determines the degree of match between the training data and the inference data. When the degree of match between the training data and the inference data is greater than a first threshold, the network device sends a first model obtained based on the training data to the terminal device. The terminal device obtains an inference result based on the first model and the inference data, which can improve the generalization of the first model. When the degree of match between the training data and the inference data is less than or equal to the first threshold, the network device obtains updated training data from the terminal device, updates the first model based on the updated training data, and obtains a second model. The terminal device applies the second model obtained from the network device to the inference data, which can ensure the accuracy of the inference result.

[0231] The present embodiment also takes the first device as a network device, the training data comes from the network device, the inference data comes from the terminal device, the terminal device determines the inference result, and the network device trains the model as an example to propose a communication method 500. Figure 9 is an interactive diagram of the communication method 500. The communication method 500 is also explained from the perspective of the interaction between the terminal device and the network device. The communication method 500 includes but is not limited to the following steps:

[0232] S501. The network device trains a model based on the training data to obtain a first model.

[0233] The training data may be historical data previously collected by the network device. Furthermore, the present embodiment does not limit the method used by the network device to train the model. For example, the network device may use a gradient descent method to train the model based on the training data to obtain a first model.

[0234] S502: The terminal device sends a model inference request and distribution information of inference data to the network device. Correspondingly, the network device receives the model inference request and distribution information of inference data from the terminal device.

[0235] S503. The network device determines the matching degree between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0236] S5041. When the matching degree between the training data and the inference data is greater than a first threshold, the network device sends the first model to the terminal device. Correspondingly, the terminal device receives the first model from the network device.

[0237] S5042. The terminal device obtains an inference result based on the first model and the inference data.

[0238] It is understandable that the implementation of S502 to S5042 can refer to the above-mentioned S403 to S4052 and will not be repeated here.

[0239] In another optional implementation, when the degree of matching between the training data and the inference data is less than or equal to the first threshold, the terminal device and the network device may execute the following S5051 to S5053:

[0240] S5051. When the matching degree between the training data and the inference data is less than or equal to the first threshold, the network device updates the first model to obtain a third model.

[0241] As described in S102 above, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the first model cannot be applied to the inference data. Therefore, when the degree of match between the training data and the inference data is less than or equal to the first threshold, the network device updates the first model to obtain a third model. Specifically, the network device independently determines the updated training data. This means that the updated training data is not reported to the network device by the terminal device but rather originates from the network device.

[0242] Optionally, before using the updated training data to update the first model, the network device determines the degree of match between the updated training data and the inference data. The method for determining this degree of match between the training data and the inference data by the first device in S101 is described in detail and will not be further described. Optionally, the network device uses the updated training data to update the first model, including: when the degree of match between the updated training data and the inference data is greater than a first threshold, using the updated training data to update the first model to obtain a third model. This method enables the network device to obtain a third model after updating the first model with better generalization.

[0243] S5052: The network device sends the third model to the terminal device. Correspondingly, the terminal device receives the third model from the network device.

[0244] After obtaining the third model, the network device sends the third model to the terminal device, so that the terminal device uses the third model to determine the inference result.

[0245] S5053. The terminal device obtains an inference result based on the third model and the inference data.

[0246] After receiving the third model from the network device, the terminal device applies the third model to the inference data to obtain an inference result. In other words, the terminal device uses the inference data as input to the third model and obtains an inference result output by the third model.

[0247] When the terminal device determines that the third model used in the inference result is a model obtained by the network device updating the first model using the updated training data when the degree of match between the training data and the inference data is less than or equal to the first threshold. Thus, the terminal device applies the third model to the inference data, ensuring a high degree of accuracy in the inference result.

[0248] Optionally, when the matching degree between the training data and the inference data is less than or equal to the first threshold, the terminal device does not execute the above S5051 to S5053 and directly falls back to the non-AI mode. The specific implementation method can be found in the above communication method 200 and will not be repeated here.

[0249] In this embodiment of the present application, the training data used by the network device training model comes from the network device itself, and the network device determines the degree of match between the training data and the inference data. When the degree of match between the training data and the inference data exceeds a first threshold, the network device transmits a first model derived from the training data to the network device. The terminal device obtains an inference result based on the first model and the inference data, thereby improving the generalization of the first model and obtaining a high-performance inference result.

[0250] Furthermore, when the degree of match between the training data and the inference data is less than or equal to a first threshold, the network device automatically updates the first model to obtain a third model. Consequently, the terminal device obtains an inference result based on the third model and the inference data. The third model used by the terminal device to determine the inference result is the model obtained by updating the first model when the degree of match between the training data and the inference data is less than or equal to the first threshold. Therefore, by applying the third model to the inference data, the terminal device can obtain an inference result with better performance.

[0251] It is understandable that the network device in the above-mentioned communication methods 200 to 500 can be replaced with a terminal device different from the terminal device in the communication methods 200 to 500. For example, in the communication method 200, the first device is terminal device a, and the network device can be replaced with a terminal device b different from terminal device a, so that the interaction between terminal device a and the network device is replaced by the interaction between terminal device a and terminal device b.

[0252] To implement the various functions of the methods provided in the embodiments of the present application, the terminal device and the network device may include hardware structures and / or software modules, and implement the aforementioned functions in the form of hardware structures, software modules, or a combination of hardware structures and software modules. Whether a particular one of the aforementioned functions is implemented in the form of hardware structures, software modules, or a combination of hardware structures and software modules depends on the specific application and design constraints of the technical solution.

[0253] As shown in Figure 10, an embodiment of the present application provides a communication device 1000. The communication device 1000 may be a component of a first device (e.g., an integrated circuit, a chip, etc.) and is configured to implement the method of the method embodiment of the present application. The communication device 1000 may include: a communication unit 1001 and a processing unit 1002. Optionally, a storage unit 1003 may also be included.

[0254] In one possible design, one or more units shown in FIG10 may be implemented by one or more processors, or by one or more processors and memories, or by one or more processors and transceivers, or by one or more processors, memories, and transceivers, which are not limited in this embodiment of the present application. The processors, memories, and transceivers may be provided separately or integrated.

[0255] The communication device 1000 has the function of implementing the first device described in the embodiment of the present application. For example, the communication device 1000 includes a module or unit or means (means) corresponding to the first device performing the steps involved in the first device described in the embodiment of the present application. The function or unit or means (means) can be implemented by software, or by hardware, or by hardware executing the corresponding software implementation, or by a combination of software and hardware. For details, please refer to the corresponding description in the aforementioned corresponding method embodiment.

[0256] In one possible design, a communication device 1000 may include: a processing unit 1002 and a communication unit 1001, wherein the communication unit is configured to transmit and receive signals / signaling;

[0257] The processing unit 1002 is configured to determine the matching degree between the training data and the inference data based on the distribution information of the training data and the inference data;

[0258] The processing unit 1002 is further configured to determine whether a first model can be applied to the inference data based on a degree of matching between the training data and the inference data, where the first model is obtained by training using the training data.

[0259] In an optional implementation, the degree of matching between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data; or, the degree of matching between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data; or, the degree of matching between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data.

[0260] In an optional implementation, the first model is an artificial intelligence AI model or a machine learning ML model.

[0261] In an optional implementation, the training data and the inference data include multiple features; the processing unit 1002 determines the matching degree between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, specifically for: determining the matching degree corresponding to each feature based on the distribution information of each feature in the multiple features in the training data and the distribution information in the inference data, the matching degree corresponding to each feature being the matching degree of the feature between the training data and the inference data; determining the matching degree between the training data and the inference data based on the matching degree corresponding to each feature.

[0262] In an optional implementation, the processing unit 1002 determines whether the first model can be applied to the inference data based on the matching degree between the training data and the inference data, specifically for: when the matching degree between the training data and the inference data is greater than a first threshold, determining that the first model can be applied to the inference data; or, when the matching degree between the training data and the inference data is less than or equal to the first threshold, determining that the first model cannot be applied to the inference data.

[0263] In an optional implementation, the communication device 1000 is applied to a terminal device, the matching degree between the training data and the inference data is greater than the first threshold, and the processing unit 1002 is also used to: send inference request information to the network device, the inference request information is used to request to obtain the first model for inference; receive the first model from the network device; and obtain an inference result based on the first model and the inference data.

[0264] In another optional implementation, the communication device 1000 is applied to a terminal device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the processing unit 1002 is also used to: send inference request information to the network device, the inference request information is used to request to obtain the first model for inference; receive the first model from the network device; and obtain an inference result based on the first model and the inference data.

[0265] In another optional implementation, the communication device 1000 is applied to a terminal device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the processing unit 1002 is also used to: send updated training data to the network device; receive a second model from the network device, the second model is obtained by training the first model using the updated training data; and obtain an inference result based on the second model and the inference data.

[0266] In an optional implementation, the communication device 1000 is applied to a terminal device, and the processing unit 1002 is further used to: determine a first parameter before determining the matching degree corresponding to each feature based on the distribution information of each feature in the multiple features in the training data and the distribution information of the inference data, the first parameter includes one or more of the following: the number of features included in the training data and the inference data, the type of each feature in the multiple features, the value range of each feature, the sampling interval of each feature, and the probability density codebook of each feature; determine the distribution information of each feature in the training data and the distribution information of each feature in the inference data according to the first parameter.

[0267] In an optional implementation, the first parameter further includes one or more of the following: a representation of the degree of matching between the training data and the inference data, a weight of the degree of matching corresponding to each feature, and the first threshold.

[0268] In an optional implementation, the communication device 1000 is applied to a network device, the matching degree between the training data and the inference data is greater than the first threshold, and the processing unit 1002 also sends the first model to the terminal device.

[0269] In another optional implementation, the communication device 1000 is applied to a network device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the processing unit 1002 is further used to: send an indication information to the terminal device indicating that the matching degree between the training data and the inference data is less than or equal to the first threshold; receive updated training data from the terminal device; update the first model based on the updated training data to obtain a second model; and send the second model to the terminal device.

[0270] In another optional implementation, the communication device 1000 is applied to a network device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the processing unit 1002 is also used to: update the first model to obtain a third model; and send the third model to the terminal device.

[0271] In another optional implementation, the communication device 1000 is applied to a network device, and the processing unit 1002 receives the distribution information of the inference data from the terminal device before determining the matching degree between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data.

[0272] The embodiments of the present application and the above-mentioned method embodiments are based on the same concept, and the technical effects they bring are also the same. For the specific principles, please refer to the description of the above-mentioned embodiments, which will not be repeated here.

[0273] The present application also provides a communication device 1100. FIG11 is a schematic diagram of the structure of the communication device 1100. The communication device 1100 may be a first device, or a chip, chip system, or processor that supports the first device in implementing the above-mentioned method. The device may be used to implement the method described in the above-mentioned method embodiment. For details, please refer to the description of the above-mentioned method embodiment.

[0274] The communication device 1100 may include one or more processors 1101. The processor 1101 may be a general-purpose processor or a dedicated processor. For example, it may be a baseband processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, or a central processing unit (CPU). The baseband processor may be used to process communication protocols and communication data, and the central processing unit may be used to control a communication device (e.g., a base station, a baseband chip, a terminal, a terminal chip, a distributed unit (DU) or a centralized unit (CU), etc.), execute software programs, and process data of the software programs.

[0275] Optionally, the communication device 1100 may include one or more memories 1102, on which instructions 1104 may be stored. The instructions may be executed on the processor 1101, causing the communication device 1100 to perform the method described in the above method embodiment. Optionally, the memory 1102 may also store data. The processor 1101 and memory 1102 may be provided separately or integrated together.

[0276] The memory 1102 may include, but is not limited to, non-volatile memories such as a hard disk drive (HDD) or a solid-state drive (SSD), random access memory (RAM), erasable programmable ROM (EPROM), ROM or portable read-only memory (CD-ROM), etc.

[0277] Optionally, the communication device 1100 may further include a transceiver 1105 and an antenna 1106. The transceiver 1105 may be referred to as a transceiver unit, a transceiver, or a transceiver circuit, etc., and is configured to implement transceiver functions. The transceiver 1105 may include a receiver and a transmitter. The receiver may be referred to as a receiver or a receiving circuit, etc., and is configured to implement a receiving function; the transmitter may be referred to as a transmitter or a transmitting circuit, etc., and is configured to implement a transmitting function.

[0278] The communication apparatus 1100 is a first device: the processor 1101 is configured to execute S101 and S102 in the communication method 100 .

[0279] The communication device 1100 is a terminal device: the processor 1101 is used to execute S203, S2043, S2053 in the above communication method 200, S303, S3043, S3053 in the communication method 300, S4052, S4065 in the communication method 400, and S503, S5042, S5053 in the communication method 500; the transceiver 1105 is used to execute the above communication method 200; S201, S2041, S2042, S2051, S2052 in method 200, S302, S3041, S3042, S3051, S3052 in communication method 300, S401, S403, S4051, S4061, S4062, S4064 in communication method 400, S502, S5041, S5052 in communication method 500.

[0280] The communication device 1100 is a network device: the processor 1101 is used to execute S202 in the above-mentioned communication method 200, S301 in the communication method 300, S402, S404, S4063 in the communication method 400, and S501 and S5051 in the communication method 500; the transceiver 1105 is used to execute S201, S2041, S2042, S2051, S2052 in the above-mentioned communication method 200, S302, S3041, S3042, S3051, S3052 in the communication method 300, S401, S403, S4051, S4061, S4062, S4064 in the communication method 400, and S502, S5041, S5042 in the communication method 500.

[0281] In another possible design, processor 1101 may include a transceiver for implementing receiving and transmitting functions. For example, the transceiver may be a transceiver circuit, an interface, or an interface circuit. The transceiver circuit, interface, or interface circuit for implementing the receiving and transmitting functions may be separate or integrated. The transceiver circuit, interface, or interface circuit may be used for reading and writing code / data, or the transceiver circuit, interface, or interface circuit may be used for transmitting or delivering signals.

[0282] In another possible design, processor 1101 may optionally store instructions 1103. Instructions 1103, when executed on processor 1101, may cause communication device 1100 to perform the method described in the above method embodiment. Instructions 1103 may be fixed in processor 1101. In this case, processor 1101 may be implemented by hardware.

[0283] In another possible design, the communication device 1100 may include a circuit that can implement the functions of sending, receiving, or communicating in the aforementioned method embodiments. The processor and transceiver described in the embodiments of the present application can be implemented in an integrated circuit (IC), an analog IC, a radio frequency integrated circuit (RFIC), a mixed signal IC, an application specific integrated circuit (ASIC), a printed circuit board (PCB), an electronic device, etc. The processor and transceiver can also be manufactured using various IC process technologies, such as complementary metal oxide semiconductor (CMOS), N-type metal oxide semiconductor (nMetal-oxide-semiconductor, NMOS), P-type metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (Bipolar Junction Transistor, BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.

[0284] The scope of the communication device described in the embodiments of the present application is not limited thereto, and the structure of the communication device may not be limited to FIG11. The communication device may be an independent device or may be part of a larger device. For example, the communication device may be:

[0285] (1) An independent integrated circuit (IC), or chip, or chip system or subsystem;

[0286] (2) A set of one or more ICs, optionally including a storage component for storing data and instructions;

[0287] (3) ASIC, such as a modem;

[0288] (4) Modules that can be embedded in other devices;

[0289] In the embodiments of the present application, the communication device and chip may also execute the implementation method described in the above-mentioned communication device 1000. Those skilled in the art may also understand that the various illustrative logical blocks and steps listed in the embodiments of the present application can be implemented by electronic hardware, computer software, or a combination of the two. Whether such functions are implemented by hardware or software depends on the specific application and the design requirements of the entire system. Those skilled in the art may use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the scope of protection of the embodiments of the present application.

[0290] The embodiments of the present application and the method embodiments shown in the above-mentioned communication methods 100 to 500 are based on the same concept and bring the same technical effects. For the specific principles, please refer to the description of the embodiments shown in the above-mentioned communication methods 100 to 500 and will not be repeated here.

[0291] The present application also provides a computer-readable storage medium for storing computer software instructions, which, when executed by a communication device, implements the functions of any of the above method embodiments.

[0292] The present application also provides a computer program product for storing computer software instructions, which, when executed by a communication device, implements the functions of any of the above method embodiments.

[0293] The present application also provides a computer program that, when executed on a computer, implements the functions of any of the above method embodiments.

[0294] The present application also provides a communication system, which includes one or more network devices and one or more terminal devices. In another possible design, the system may also include other devices that interact with the network devices and terminal devices in the solution provided by the present application.

[0295] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a high-density digital video disc (DVD)), or a semiconductor medium (eg, an SSD).

[0296] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A communication method, characterized in that: The method comprises: Determining a degree of matching between the training data and the inference data based on distribution information of the training data and distribution information of the inference data; Based on the matching degree between the training data and the inference data, it is determined whether a first model can be applied to the inference data, wherein the first model is obtained by training with the training data.

2. The method according to claim 1, characterized in that: The degree of matching between the training data and the inference data is determined by the difference between the distribution information of the training data and the distribution information of the inference data; or, The matching degree between the training data and the inference data is determined by the similarity between the distribution information of the training data and the distribution information of the inference data; or, The degree of matching between the training data and the inference data is determined by the distance between the distribution information of the training data and the distribution information of the inference data.

3. The method according to claim 1 or 2, characterized in that: The first model is an artificial intelligence AI model or a machine learning ML model.

4. The method according to any one of claims 1 to 3, characterized in that: The training data and the inference data include a plurality of features; and determining the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data includes: Based on the distribution information of each feature of the multiple features in the training data and the distribution information of each feature in the inference data, determining the matching degree corresponding to each feature, wherein the matching degree corresponding to each feature is the matching degree of the feature between the training data and the inference data; Based on the matching degree corresponding to each feature, the matching degree between the training data and the inference data is determined.

5. The method according to any one of claims 1 to 4, characterized in that: The determining, based on the degree of matching between the training data and the inference data, whether the first model can be applied to the inference data comprises: When the degree of match between the training data and the inference data is greater than a first threshold, determining that the first model can be applied to the inference data; or, When the degree of matching between the training data and the inference data is less than or equal to a first threshold, it is determined that the first model cannot be applied to the inference data.

6. The method according to claim 5, characterized in that The method is executed by a terminal device, the matching degree between the training data and the inference data is greater than the first threshold, and the method further includes: Sending inference request information to a network device, where the inference request information is used to request acquisition of the first model for inference; receiving the first model from the network device; Based on the first model and the inference data, an inference result is obtained.

7. The method according to claim 5, characterized in that The method is executed by a terminal device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the method further includes: Sending updated training data to network devices; Receiving a second model from a network device, where the second model is obtained by training the first model using the updated training data; Based on the second model and the inference data, an inference result is obtained.

8. The method according to claim 5, characterized in that The method is executed by a terminal device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the method further includes: Sending model update request information for updating the first model to the network device; receiving a third model from a network device, wherein the third model is a model obtained by updating the first model; Based on the third model and the inference data, an inference result is obtained.

9. The method according to any one of claims 6 to 8, characterized in that: Before determining the matching degree corresponding to each feature based on the distribution information of each feature in the training data and the distribution information of the inference data, the method further includes: Determine a first parameter, where the first parameter includes one or more of the following: the number of features included in the training data and the inference data, the type of each feature in the multiple features, the value range of each feature, the sampling interval of each feature, and the probability density codebook of each feature; According to the first parameter, distribution information of each feature in the training data and distribution information of each feature in the inference data are determined.

10. The method according to claim 9, characterized in that The first parameter also includes one or more of the following: a representation method of the matching degree between the training data and the inference data, a weight of the matching degree corresponding to each feature, and the first threshold.

11. The method according to claim 5, characterized in that The method is performed by a network device, the matching degree between the training data and the inference data is greater than the first threshold, and the method further includes: The first model is sent to a terminal device.

12. The method according to claim 5, characterized in that The method is performed by a network device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the method further includes: Sending indication information to a terminal device indicating that the degree of matching between the training data and the inference data is less than or equal to the first threshold; receiving updated training data from the terminal device; Update the first model based on the updated training data to obtain a second model; The second model is sent to the terminal device.

13. The method according to claim 5, characterized in that The method is performed by a network device, the matching degree between the training data and the inference data is less than or equal to the first threshold, and the method further includes: Updating the first model to obtain a third model; The third model is sent to the terminal device.

14. The method according to any one of claims 11 to 13, characterized in that: Before determining the degree of matching between the training data and the inference data based on the distribution information of the training data and the distribution information of the inference data, the method further includes: Receive distribution information of inference data from terminal devices.

15. A communication device, characterized in that: The device comprises a module or a unit for implementing the method according to any one of claims 1 to 14.

16. A communication device, characterized in that: The apparatus comprises a processor configured to perform the method according to any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store instructions, and when the instructions are executed on a computer, the method according to any one of claims 1 to 14 is executed.

18. A computer program product comprising instructions, characterized in that When the method is executed on a computer, the method according to any one of claims 1 to 14 is executed.

Images