Apparatus, method and system for machine learning

Abstract

Apparatuses, methods, and systems for machine learning are disclosed. An exemplary method includes obtaining a local distribution parameter set of a second device, the local distribution parameter set characterizing data distribution characteristics of a local data set of the second device; and transmitting the trained specific model to the second device, wherein the specific model is determined based at least in part on the local distribution parameter set.

Description

Technical Field

[0001] The present disclosure relates to the field of communications, and in particular, to devices, methods, and systems for machine learning. Background Art

[0002] Artificial Intelligence (AI) technology is increasingly widely used. An artificial intelligence model can be trained through machine learning. In the process of training and using an artificial intelligence model, improved devices, methods, and systems are desired. Summary of the Invention

[0003] The present disclosure provides devices, methods, and systems for machine learning.

[0004] One aspect of the present disclosure relates to an electronic device, including: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code, which when executed by the at least one processing unit, causes the electronic device to perform the following operations: obtaining a local distribution parameter set of a second device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the second device; and sending a trained specific model to the second device, where the specific model is determined at least in part based on the local distribution parameter set.

[0005] Another aspect of the present disclosure relates to an electronic device, including: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code, where the computer program code when executed by the at least one processing unit, causes the electronic device to perform the following operations: sending a local distribution parameter set to a first device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the electronic device; and receiving a trained specific model from the first device, where the specific model is determined at least in part based on the local distribution parameter set.

[0006] Another aspect of the present disclosure relates to an electronic device, including: at least one processing unit; and at least one storage unit including computer program code, wherein when the computer program code is executed by the at least one processing unit, the electronic device is caused to perform the following operations: receiving one or more sets of local distribution parameters, each set of local distribution parameters in the one or more sets of local distribution parameters characterizing the data distribution characteristics of the local data set of a corresponding client among one or more clients; sending a task identifier to a model library, the task identifier identifying a training task associated with the one or more clients; receiving from the model library one or more trained candidate models corresponding to the task identifier; and sending a model list including a matching model selected from the one or more candidate models for each of the one or more clients, the matching model being determined at least in part based on the corresponding set of local distribution parameters in the one or more sets of local distribution parameters.

[0007] Another aspect of the present disclosure relates to a computer-readable storage medium storing one or more instructions, which when executed by one or more processing circuits of an electronic device, cause the electronic device to perform any method as described in the present disclosure.

[0008] Another aspect of the present disclosure relates to a computer program product including a computer program which, when executed by a processor, implements any method as described in the present disclosure.

[0009] Another aspect of the present disclosure relates to an apparatus including components for performing any method as described in the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and other objects and advantages of the present disclosure will be further described below in conjunction with specific embodiments and with reference to the accompanying drawings. In the drawings, the same or corresponding technical features or components will be denoted by the same or corresponding reference numerals.

[0011] Figure 1 A schematic diagram of an exemplary system for federated learning according to some embodiments of the present disclosure is shown.

[0012] Figure 2 An exemplary block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0013] Figure 3 An exemplary block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0014] Figure 4 An exemplary block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0015] Figure 5 A flowchart of a method according to some embodiments of the present disclosure is shown.

[0016] Figure 6 A flowchart of a method according to some embodiments of the present disclosure is shown.

[0017] Figure 7 A flowchart of a method according to some embodiments of the present disclosure is shown.

[0018] Figure 8 A schematic diagram of a process according to some embodiments of the present disclosure is shown.

[0019] Figure 9 The correspondence between a local data set and a local distribution parameter set according to some embodiments of the present disclosure is shown.

[0020] Figure 10A An example process of calculating the similarity between a local data set and a training data set according to some embodiments of the present disclosure is shown.

[0021] Figure 10B An example of calculating the feature distance between two samples of a local data set and a training data set according to some embodiments of the present disclosure is shown.

[0022] Figure 11 A block diagram showing a first example of an exemplary configuration of a gNB to which the techniques of the present disclosure can be applied.

[0023] Figure 12 A block diagram showing a second example of an exemplary configuration of a gNB to which the techniques of the present disclosure can be applied.

[0024] Figure 13 A block diagram showing an example of an exemplary configuration of a communication device to which the techniques of the present disclosure can be applied.

[0025] Figure 14 A block diagram showing an example of an exemplary configuration of a vehicle navigation device to which the techniques of the present disclosure can be applied.

[0026] Although the embodiments described in the present disclosure may be susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and are described in detail in the present disclosure. It should be understood that the drawings and the detailed description thereof are not intended to limit the embodiments to the particular forms disclosed, but on the contrary, are intended to cover all modifications, equivalents, and alternative arrangements falling within the spirit and scope of the claims. Detailed Description of Specific Embodiments

[0027] In the following, exemplary embodiments of the present disclosure will be described with reference to the accompanying drawings. For clarity and conciseness, not all features of the embodiments are described in the specification. However, it should be understood that many implementation-specific settings must be made during the implementation of the embodiments in order to achieve the specific goals of the developer, for example, to comply with those limitations related to the device and the business, and these limitations may vary with different implementations. In addition, it should also be understood that although the development work may be very complex and time-consuming, for those skilled in the art who benefit from the present disclosure, such development work is merely a routine task.

[0028] Here, it should also be noted that in order to avoid obscuring the present disclosure with unnecessary details, only the processing steps and / or device structures that are closely related to at least the solution according to the present disclosure are shown in the drawings, while other details that are not closely related to the present disclosure are omitted.

[0029] The application of artificial intelligence technology is becoming increasingly widespread. An artificial intelligence model can be trained through machine learning. Generally speaking, machine learning is iteratively performed based on a training data set. A complete training data set usually includes an input data part and an output data part, where the input data part can represent the original data, and the output data part can include labels, predictions, decisions, etc. corresponding to the original data. For example, in the example of training a classification model, the output data part can be the classification label corresponding to the original data. Another example is that in the example of training a prediction model, the output data part can be the prediction result corresponding to the original data. Another example is that in the example of training a decision model, the output data part can be the decision result corresponding to the original data. By providing the training data set including the input data part and the output data part to the model to be trained, the behavior and performance of the model can be iteratively corrected, thereby obtaining a trained model. The trained model can be used to generate new labels, predictions, or decisions based on new input data.

[0030] A data set including an input data part and a complete output data part can be called a labeled data set. A data set including an input data part but lacking a complete output data part can be called an unlabeled data set or a partially unlabeled data set. Using an unlabeled data set or a partially unlabeled data set as a training data set to train a model will result in insufficient model training, thereby reducing the performance of the trained model.

[0031] The inventors of the present disclosure have found that the training data sets held by the model training party are not always complete. Obtaining the corresponding output data part based on the input data part of the data set usually requires sufficient expertise and experience (e.g., image annotation experts, etc.) and sufficient processing time, and accordingly incurs costs. In some cases, due to limitations in knowledge, time, or cost, it is difficult for the model training party to obtain the complete output data part. This challenge is particularly prominent in the Federated Learning (FL) process described below. According to the technology of the present disclosure, it is possible to supplement the output data part corresponding to the input data part, thereby forming a complete training data set. This data supplementation process can be automatic, fast, accurate, and / or low-cost.

[0032] In addition, the inventors of the present disclosure have found that the input data part in the training data set may depend on the specific scenario. For example, the training image set associated with the first scenario may have different features from the training image set associated with the second scenario, resulting in differences between the models trained based on these two sets of image sets. When applying the trained model to a new local image set, it is beneficial to find the trained model that best matches the local image.

[0033] 1. Exemplary System

[0034] This section describes an exemplary system for federated learning. Federated learning is a distributed learning technique. In a system employing federated learning, each of multiple training devices uses its local data set as the training data set to respectively perform training on the corresponding local model. Each training device can upload the trained corresponding local model to the server. The server aggregates the multiple local models received from the multiple training devices to obtain a global model.

[0035] Figure 1 A schematic diagram of an exemplary system 100 for federated learning according to some embodiments of the present disclosure is shown. The system 100 may include a server 110. The server 110 may act as the server for federated learning. The system 100 may also include a plurality of clients 120-1 to 120-5. The plurality of clients may be uniformly represented by 120. Each client 120 may be communicatively coupled to the server 110. Each client 120 may act as a training device for federated learning.

[0036] During federated learning, the training of an artificial intelligence model can be an iterative process that includes multiple training rounds. In one training round, each client 120 can use the local dataset of that client as a training dataset to perform training, thereby obtaining a corresponding trained local model 121. The local datasets of different clients 120 may be different, so the obtained local models 121 may be different. For example, clients 120-1, 120-2, 120-3, 120-4, 120-5 can obtain trained local models 121-1, 121-2, 121-3, 121-4, 121-5 respectively. Each client 120 can upload the trained local model 121 to the server 110. The server 110 can aggregate the multiple local models 121 received from each client 120 to generate a global model 111. Then, the global model 111 can be distributed to each client 120 via the downlink between the server 110 and each client 120. Based on the received global model 111, each client 120 can update the local model 121 of that client.

[0037] In the next training round, each client 120 can continue to train the updated local model 121 of that client, thereby obtaining a trained local model 121. Once again, each client 120 can upload the trained local model 121 obtained in the current training round to the server 110. The server 110 can aggregate the multiple local models 121 received from each client 120 to generate a global model 111. The global model 111 can be distributed to each client 120 again for subsequent training rounds. This process can be repeated until an end condition is reached. The performance of the global model 111 can be improved by iteratively going through multiple training rounds.

[0038] During the process of federated learning, each client uploads a local model without having to exchange the local training dataset with the server or other training devices. Therefore, federated learning can utilize the training capabilities and training datasets of a large number of clients to train an artificial intelligence model while ensuring data security and user privacy. This helps to improve the training efficiency and the performance of the obtained artificial intelligence model.

[0039] To improve the training efficiency of federated learning and the generalization ability of the trained model, it is desirable for as many clients as possible to participate in the federated learning process. However, it is difficult to require all clients participating in the federated learning process to have sufficient experience, knowledge, processing power, or cost to obtain a complete training dataset (specifically, the complete output data part of the client's local dataset) in a conventional manner (e.g., equipped with a labeled expert with specialized knowledge). Therefore, the local dataset of one or more of these clients may not be a complete training dataset because some input data parts of the local dataset do not have corresponding output data parts. When using the local dataset with missing output data parts as the training dataset, the training of the local model by the client may be insufficient, and the resulting global model is not desirable.

[0040] On the other hand, the input data parts held by clients may vary depending on the scenario. The input data parts obtained in the first scenario may not be accurately processed based on inherent expertise and experience (e.g., a labeled expert trained based on the second scenario) to generate accurate output data parts. If the client uses a certain fixed mechanism to obtain the output data part, the output data part given by the fixed mechanism for the varying input data parts in the changing scenario may lack accuracy (e.g., labeling accuracy). In addition, if a single mechanism is applied to all clients participating in the federated learning process, it may not accurately adapt to the input data parts of each client in the corresponding scenario. The resulting output data part may not be accurate. When using the local dataset with inaccurate output data parts as the training dataset, the training of the local model by the client may be inaccurate, and the resulting global model is not desirable.

[0041] It should be understood that the federated learning process described in connection with Figure 1 is only an example of a machine learning process facing one or more technical problems. Similar technical problems may exist in other machine learning processes. Therefore, the techniques described in this disclosure are not limited to the federated learning process. For example, during the model usage (instead of model training) process, the model user also faces the problem of selecting a matching model for the local dataset in a specific scenario. In addition, there are one or more other technical problems.

[0042] 2. Exemplary Device

[0043] Figure 2 An exemplary block diagram of an electronic device 200 according to some embodiments of the present disclosure is shown. In some embodiments, the electronic device 200 may be implemented on the server side. Therefore, the electronic device 200 may be referred to as a server or a server device. For example, in Figure 1In the described system 100, the electronic device 200 may be implemented at the server 110. The electronic device 200 may be implemented as the server itself, as a part of the server, or as a control device for controlling the server. For example, the electronic device 200 may be implemented as a chip for controlling the server. In some embodiments of the present disclosure, implementing the electronic device 200 as the server itself is merely for convenience of description and is not intended to be limiting. In other embodiments, the electronic device 200 may alternatively be implemented as other devices other than the server.

[0044] According to some embodiments of the present disclosure, the electronic device 200 may include a communication unit 210, a storage unit 220, and a processing circuit 230.

[0045] The communication unit 210 of the electronic device 200 may be used to receive or send wired transmissions or radio transmissions. The communication unit 210 may perform functions such as upconversion, digital-to-analog conversion, etc. on the transmitted signal, and / or perform functions such as downconversion, analog-to-digital conversion, etc. on the received signal. Various techniques may be used to implement the communication unit 210. For example, the communication unit 210 may be implemented as communication interface components such as antenna devices, radio frequency circuits, and part of the baseband processing circuit. The communication unit 210 is drawn with a dashed line because it may alternatively be located within the processing circuit 230 or outside the electronic device 200.

[0046] The storage unit 220 of the electronic device 200 may store information generated by the processing circuit 230, information received from other devices through the communication unit 210, or information to be sent to other devices, programs, machine codes, and data for the operation of the electronic device 200, etc. The storage unit 220 may be a volatile memory and / or a non-volatile memory. For example, the storage unit 220 may include, but is not limited to, random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), read-only memory (ROM), and flash memory. The storage unit 220 is drawn with a dashed line because it may alternatively be located within the processing circuit 230 or outside the electronic device 200.

[0047] The processing circuit 230 of the electronic device 200 may be configured to perform one or more operations to provide various functions of the electronic device 200. The functions of the elements disclosed herein may be implemented using a circuit or a processing circuit 230, which includes a general - purpose processor, a dedicated processor, an integrated circuit, an ASIC (“Application - Specific Integrated Circuit”), conventional circuits, and / or a combination thereof that are configured or programmed to perform the disclosed functions. A processor is considered a processing circuit or a circuit because it includes transistors and other circuits therein. In the present disclosure, a circuit, a unit, or a device is hardware that performs or is programmed to perform the said functions. The hardware may be any hardware disclosed herein or otherwise known, which is programmed or configured to perform the said functions. When the hardware is a processor which can be considered a type of circuit, the circuit, the device, or the unit is a combination of hardware and software, and the software is used to configure the hardware and / or the processor. According to some embodiments, in response to the processing circuit 230 executing computer program code contained in the storage unit 220, the electronic device 200 may be configured to perform one or more steps performed by the server device in the present disclosure.

[0048] According to some embodiments, the electronic device 200 may be configured to obtain a set of local distribution parameters of a second device, and the set of local distribution parameters characterizes the data distribution characteristics of the local data set of the second device. The second device may be another electronic device different from the electronic device 200. For example, when the electronic device 200 is implemented as a server for machine learning, the second device may be a client associated with the machine learning. In some embodiments, the client may be a user of a model associated with the machine learning. In other embodiments, the client may be a trainer of a model associated with the machine learning. For example, the client may be a training device associated with federated learning.

[0049] According to some embodiments, the electronic device 200 may be further configured to send a trained specific model to the second device, and the specific model is determined at least in part based on the set of local distribution parameters.

[0050] The distribution parameter set can characterize the data distribution characteristics of the data set. The inventors of the present disclosure have found that different distribution parameter sets can be used to distinguish different data sets (e.g., portions of input data obtained in different scenarios). By selecting a specific model based on the local distribution parameter set of the local data set of the second device, the selected specific model can be made to match the local data set of the second device. Such a specific model can be referred to as a matching model. By applying the matching specific model to the local data set of the second device, the results output by the model have high accuracy. Additionally, by transmitting the local distribution parameter set instead of the local data set itself, the communication burden between electronic devices can be advantageously saved, since the local data set typically contains a large amount of data. Additionally, avoiding the transmission of the local data set can maintain data security and user privacy.

[0051] In some embodiments, the processing circuit 230 may include a parameter acquisition unit 231 and a model distribution unit 232, which may each include corresponding hardware circuits or software modules. The parameter acquisition unit 231 may be configured to perform the step of obtaining the local distribution parameter set. The model distribution unit 232 may be configured to perform the step of sending the specific model. According to some embodiments of the present disclosure, the electronic device 200 may be further configured to perform one or more additional steps performed by the server device, as described in further detail below. Accordingly, the processing circuit 230 may include one or more additional units (not shown) for performing the one or more additional steps.

[0052] Figure 3 An exemplary block diagram of an electronic device 300 according to some embodiments of the present disclosure is shown. In some embodiments, the electronic device 300 may be implemented on the client side. Accordingly, the electronic device 300 may be referred to as a client or a client device. For example, in Figure 1 the system 100 described above, the electronic device 300 may be implemented at the client 120. The electronic device 300 may be implemented as the client itself, as part of the client, or as a control device for controlling the client. For example, the electronic device 300 may be implemented as a chip for controlling the client. In some embodiments of the present disclosure, implementing the electronic device 300 as the client itself is merely for convenience of description and is not intended to be limiting. In other embodiments, the electronic device 300 may alternatively be implemented as other devices other than the client.

[0053] According to some embodiments of the present disclosure, the electronic device 300 may include a communication unit 310, a storage unit 320, and a processing circuit 330. The specific implementations of the communication unit 310, the storage unit 320, and the processing circuit 330 may be similar to the communication unit 210, the storage unit 220, and the processing circuit 230 described above with respect to the electronic device 200, and will not be described herein again. According to some embodiments, in response to the processing circuit 330 executing the computer program code included in the storage unit 320, the electronic device 300 may be configured to perform one or more steps executed by a client device in implementing various methods according to some embodiments of the present disclosure.

[0054] According to some embodiments, the electronic device 300 may be configured to send a local distribution parameter set to a first device. The local distribution parameter set may characterize the data distribution characteristics of the local data set of the electronic device 300. The first device may be another electronic device different from the electronic device 300. In some embodiments, when the electronic device 300 is implemented as a client associated with machine learning, the first device may include a server for the machine learning. The first device may be implemented by the electronic device 200. In other embodiments, the first device may alternatively include other devices other than the server for machine learning.

[0055] According to some embodiments, the electronic device 300 may be further configured to receive a trained specific model from the first device, where the specific model is determined at least in part based on the local distribution parameter set of the local data set of the electronic device 300.

[0056] In some embodiments, the processing circuit 330 may include a parameter sending unit 331 and a model receiving unit 332, which may each include corresponding hardware circuits or software modules. The parameter sending unit 331 may be configured to perform the step of sending the local distribution parameter set. The model receiving unit 332 may be configured to perform the step of receiving the trained specific model. According to some embodiments of the present disclosure, the electronic device 300 may be further configured to perform one or more additional steps executed by a server device, as described in further detail below. Accordingly, the processing circuit 330 may include one or more additional units (not shown) for performing the one or more additional steps.

[0057] Figure 4FIG. 0 shows an exemplary block diagram of an electronic device 400 according to some embodiments of the present disclosure. The electronic device 400 may be configured to operate collaboratively with other electronic devices (e.g., the electronic device 200) to implement the technologies described in the present disclosure. In some embodiments, the electronic device 400 may be implemented at a network element device in a cellular network, enabling the technologies described in the present disclosure to be executed in the environment of the cellular network. As a specific example, the electronic device 400 may be implemented at a network data analytics function (NWDAF). In other embodiments, the electronic device 400 may be implemented at other devices of other types of networks without limitation.

[0058] According to some embodiments of the present disclosure, the electronic device 400 may include a communication unit 410, a storage unit 420, and a processing circuit 430. The specific implementations of the communication unit 410, the storage unit 420, and the processing circuit 430 may be similar to those of the communication unit 210, the storage unit 220, and the processing circuit 230 described above with respect to the electronic device 200, and will not be elaborated herein. According to some embodiments, in response to the processing circuit 430 executing computer program code contained in the storage unit 420, the electronic device 400 may be configured to perform one or more steps of implementing various methods according to some embodiments of the present disclosure.

[0059] According to some embodiments, the electronic device 400 may be configured to receive one or more sets of local distribution parameters. Each set of local distribution parameters in the one or more sets of local distribution parameters may characterize the data distribution characteristics of the local data set of the corresponding client in one or more clients.

[0060] According to some embodiments, the electronic device 400 may be further configured to send a task identifier to a model library. The task identifier may identify a training task associated with the one or more clients.

[0061] According to some embodiments, the electronic device 400 may be further configured to receive from the model library one or more trained candidate models corresponding to the task identifier.

[0062] According to some embodiments, the electronic device 400 may be further configured to send a model list. The model list may include a matching model selected from the one or more candidate models for each of the one or more clients, and the matching model is determined at least in part based on the corresponding set of local distribution parameters in the one or more sets of local distribution parameters.

[0063] In some embodiments, the processing circuit 430 may include a parameter management unit 431 and a model matching unit 432, which may each include corresponding hardware circuits or software modules. The parameter management unit 431 may be configured to perform the step of obtaining the one or more local distribution parameter sets. The model matching unit 432 may be configured to perform the steps of sending a task identifier, receiving candidate models, and sending a model list. In some embodiments, the model matching unit 432 may include sub-units (not shown) for performing each of these steps.

[0064] According to some embodiments of the present disclosure, the electronic device 400 may be further configured to perform one or more additional steps, as described in further detail below with respect to NWDAF. Accordingly, the processing circuit 430 may include one or more additional units (not shown) for performing the one or more additional steps.

[0065] 3. Exemplary Method

[0066] Figure 5 A flowchart of a method 500 according to some embodiments of the present disclosure is shown. According to some embodiments, the method 500 may be executed at a server for machine learning. For example, the method 500 may be executed by the server 110. The server 110 may be implemented by the aforementioned electronic device 200. Accordingly, the method 500 may be executed by the processing circuit 230 of the electronic device 200.

[0067] The method 500 may start from step 510. In step 510, the server may be configured to obtain a local distribution parameter set of a client. The local distribution parameter set characterizes the data distribution characteristics of the local data set of the client.

[0068] In some embodiments, the server may be configured to directly receive the local distribution parameter set from the client. In other embodiments, the server may be configured to receive the local distribution parameter set from another device associated with the client (e.g., an intermediate device between the server and the client). In some embodiments, if a certain degree of loss of data security and user privacy can be tolerated, the server may receive the local data set (or a part of the samples of the local data set) from the client and calculate the local distribution parameter set.

[0069] The obtained local distribution parameter set may include one or more distribution parameters. The number and type of the one or more distribution parameters may vary depending on the local data set. As an example, the local data set may follow a Gaussian distribution, a gamma distribution, a lognormal distribution, or any other type of distribution. Accordingly, the local distribution parameter set may include distribution parameters corresponding to these distribution types. Optionally, the local distribution parameter set may additionally include a type indicator indicating the distribution type.

[0070] Method 500 may proceed to step 520. In step 520, the server may be configured to send the trained specific model to the client, where the specific model is determined at least in part based on the set of local distribution parameters.

[0071] In the present disclosure, a trained model refers to a model that has completed a training process based on a training data set. The trained model can be used to generate an output data portion (e.g., label, prediction, decision, etc.) based on an input data portion. Thus, the trained specific model can be applied by the client device to at least a portion of its local data set to generate an output data portion corresponding to the at least a portion.

[0072] According to some embodiments of the present disclosure, the specific model determined for the client may be selected from one or more candidate models. Among the one or more training data sets used to train the one or more candidate models, the specific training data set used to train the specific model may have the highest similarity to the local data set of the client. Selecting the specific model based on the similarity between the training data set and the local data set can make the selected specific model adapted to the local data set of the client. Accordingly, compared with other models among the one or more candidate models, the output data portion generated by the specific model based on the local data set will be more accurate.

[0073] According to some embodiments of the present disclosure, the similarity between the local data set and the specific training data set may be determined at least in part based on: (1) the set of local distribution parameters of the local data set, and (2) the set of specific distribution parameters of the specific training data set. An example calculation process for determining the similarity between two data sets based on two sets of distribution parameters is further described later.

[0074] In some embodiments, the server may be configured to implement a model provider, and the client may be configured to be a model user. In this case, the specific model determined in step 520 can be applied by the client to the local data set of the client to generate an output data portion. The client can use the output data portion to perform one or more actions.

[0075] For example, the client can be an access control device using face recognition. The server is maintained by a face recognition service provider, and the specific model is a specific face recognition model provided by the face recognition service provider. The access control device can apply the specific face recognition model to the captured user images and determine whether to allow the user to enter based on the output result of the face recognition model. When the same access control device is deployed in different scenarios (e.g., in different regions, at different times, under different lighting conditions, or at different shooting angles), the resulting local dataset (e.g., the pixel data in the user image set) may have different data distribution characteristics. Through method 500, the face recognition service provider can provide the access control device with a specific face recognition model that best matches the local dataset in the current scenario because the specific face recognition model is trained using a training dataset with similar data distribution characteristics. In this way, the accuracy of the access control device in the current scenario can be improved.

[0076] It should be understood that face recognition is only one use case of the models described in this disclosure. The techniques described in this disclosure are not limited to this specific use case.

[0077] In some other embodiments, the client can be configured to implement the model training party. For example, the server can be configured to implement the server of the federated learning process, and the client is configured to be one of the multiple clients implementing the federated learning process. In this case, the server can be configured to receive the trained local model from the client. The local model is trained by the client based on the supplemented local dataset. The supplemented local dataset includes the supplemented output data part. The supplemented output data part is generated by the specific model based on at least a part of the client's local dataset.

[0078] Specifically, the client can provide at least a part of the local dataset (e.g., at least a part lacking the corresponding output data part) as input data to the specific model to generate the supplemented output data part. The output data part can be used to supplement the client's local dataset. The supplemented local dataset can form a complete training dataset for training the client's local model.

[0079] The server can be further configured to generate a global model at least partially based on the aggregation of the local models received from the clients. For example, the server can aggregate the local model with the trained local models from other clients participating in the federated learning, as described in Figure 1 The trained local models from other clients can be trained based on the local datasets of other clients. The local datasets of other clients can be or can not be supplemented local datasets.

[0080] According to some embodiments of the present disclosure, one or more trained candidate models from which a specific model is selected can be maintained in a model library. The model library can also maintain one or more training data sets corresponding to the one or more candidate models. The model library can reside at another device different from the server (e.g., the Analytics Data Repository Function (ADRF) in a cellular network). In this case, the server can also be configured to send the set of local distribution parameters obtained in step 510 together with the client identifier of the client. Then, the server can receive a model list, which can include the specific model selected for the client.

[0081] In some embodiments, the set of local distribution parameters together with the client identifier can be sent to a remote device (e.g., electronic device 400) that communicates with the model library. Then, the server can receive a model list from the remote device. In some embodiments, the model list can include one or more matching models determined for one or more clients. The server can be configured to identify, based on the client identifier of each client, the matching model (e.g., the specific model described above) for sending to the client from the model list. Then, the identified matching model can be sent to the corresponding client, as described in step 520.

[0082] According to some embodiments of the present disclosure, the server can additionally be configured to send the set of local distribution parameters obtained in step 510 together with a task identifier. The task identifier can be used to identify a specific training task of the client in the federated learning process. The specific training task can correspond to a specific service provided by the client to the user. Then, the server can be configured to receive a specific model, where the specific model is determined at least in part based on the task identifier.

[0083] In some embodiments, the set of local distribution parameters and the task identifier can be sent by the server to a remote device (e.g., electronic device 400) that communicates with the model library. Then, the server can receive the specific model associated with the client from the remote device. For example, the remote device can obtain one or more candidate models corresponding to the task identifier from the model library. The remote device can select a specific model that matches the client from the one or more candidate models.

[0084] In embodiments of federated learning, each task identifier can be used to identify a specific training task of a group of clients during the federated learning process. For example, a first group of clients may be configured to perform a first training task (e.g., training a face recognition model based on a local image set) and be associated with a first task identifier, and a second group of clients may be configured to perform a second training task (e.g., training a channel prediction model based on a local signal data set) and be associated with a second task identifier. Each client can be associated with a task identifier. The server can identify different tasks of different clients and send the corresponding task identifier together with the set of distribution parameters associated with the client to a remote device, so as to obtain a matching model corresponding to the training task. For example, the server can obtain a set of trained face recognition models for the first group of clients. Each model in the set of face recognition models is distributed by the server to the corresponding matching client in the first group of clients to complement the local image set of the client. The server can obtain a set of trained channel prediction models for the second group of clients. Each model in the set of channel prediction models is distributed by the server to the corresponding matching client in the second group of clients to complement the local signal data set of the client. It should be understood that the above two tasks are merely exemplary, and the technology of the present disclosure is not limited thereto.

[0085] According to some embodiments, the technology of the present disclosure can be executed in a cellular network. By way of example and not limitation, the server can be deployed with an Application Function (AF) of the cellular network. The remote device can implement a Network Data Analytics Function (NWDAF) of the cellular network. The model library can reside at a Data Analytics Repository Function (ADRF) of the cellular network. In this case, if the AF deployed at the server is untrusted, the server can send the local distribution parameter set (along with an optional task identifier or client identifier) to a Network Exposure Function (NEF) of the cellular network. The NEF can forward this information to the NWDAF. Accordingly, the NWDAF can return a model list containing specific models to the server via the NEF. If the AF deployed at the server is trusted, the server can directly send the local distribution parameter set (along with an optional task identifier or client identifier) to the NWDAF. Accordingly, the NWDAF can directly return a model list containing specific models to the server without passing through the NEF.

[0086] It should be understood that the above description is only an exemplary embodiment of method 500. Details of each step of method 500 and additional or optional steps will be further described below.

[0087] Figure 6FIG. 600 is a flowchart of a method according to some embodiments of the present disclosure. According to some embodiments, method 600 may be executed at a client for machine learning. For example, method 600 may be executed by client 120. Client 120 may be implemented by the aforementioned electronic device 300. Correspondingly, method 600 may be executed by processing circuit 330 of electronic device 300.

[0088] Method 600 may start from step 610. In step 610, the client may be configured to send a set of local distribution parameters to the server, and the set of local distribution parameters characterizes the data distribution characteristics of the client's local dataset.

[0089] In the present disclosure, the local dataset of the client refers to the collection of data held by the client or that can be accessed independently. The local dataset may include various types of data. For example, the local dataset may include data captured by the client's sensors, user data associated with the client's user, communication data generated by the client's communication, and so on. The client may be configured to generate a corresponding set of local distribution parameters based on the local dataset (or a part of its samples). The client may have multiple local datasets, and may generate a set of local distribution parameters for each local dataset separately.

[0090] Method 600 may continue to step 620. In step 620, the client may be configured to receive a trained specific model from the server, where the specific model is determined at least in part based on the set of local distribution parameters of the client.

[0091] As previously described, the received specific model may be selected from one or more candidate models. The specific model may match the client's local dataset. Specifically, among one or more training datasets used to train the one or more candidate models, the specific training dataset used to train the specific model may have the highest similarity to the client's local dataset. The similarity may be determined at least in part based on the specific distribution parameters of the specific training dataset and the local distribution parameters of the client's local dataset.

[0092] According to some embodiments of the present disclosure, the client may use the received specific model to generate an output data portion based on at least a part of the local dataset. For example, the client may provide the at least a part of the local dataset as input data to the received specific model to generate an output data portion corresponding to the at least a part. Depending on the function to be implemented by the specific model, the output data portion may include at least one of labels, predictions, and / or decisions generated based on the at least a part of the local dataset.

[0093] According to some embodiments of the present disclosure, the client can be configured to implement a model user. The model user can be configured to use the output data portion generated by a specific model to perform one or more actions, such as those described in the previous example of the access control device.

[0094] According to other embodiments of the present disclosure, the client can be configured to implement a model trainer. For example, the server can be configured to implement the server of the federated learning process, and the client is configured to be one of the multiple clients of the federated learning process. In this case, the client can use the output data portion generated by a specific model to supplement the local data set. Compared with the original local data set (e.g., the local data set used to generate the local distribution parameter set in step 610), the supplemented local data set additionally includes the output data portion generated by the specific model received in step 620. The client can be further configured to use the supplemented local data set as the training data set to train the local model of the client, and send the trained local model to the server for aggregation to generate a global model for the federated learning process.

[0095] The inventors of the present disclosure recognize that using the supplemented local data set as the training data set can ensure that the local model is fully trained. Therefore, the performance of the global model obtained by aggregating the local models is also improved.

[0096] The inventors of the present disclosure also recognize that for multiple clients participating in the federated learning process, the supplemented local data set of each client is adaptively supplemented based on the data distribution characteristics of the local data set of the client, which improves the accuracy of the supplemented local data set (specifically, the output data portion), thereby improving the accuracy of training the local model. Therefore, the performance of the global model obtained by aggregating the local models is also improved.

[0097] The data supplementation process as described in the present disclosure is automatic, fast, accurate, and / or low-cost. This process does not require the client to have prior experience and knowledge. And this process can be adaptively updated based on the changing scenarios faced by the client (e.g., adaptively updating the matching model sent to the client).

[0098] As described above, the output data portion generated by a specific model may include labels, predictions, or decisions generated based on at least a portion of the local dataset. For example, for an input data portion lacking a label, the client may input the input data portion into the received specific model to generate a corresponding label. The client may associate the input data portion with the generated label to form labeled data. For an input data portion lacking a prediction result, the client may input the input data portion into the received specific model to generate a corresponding prediction result. The client may associate the input data portion with the generated prediction result. For an input data portion lacking a decision result, the client may input the input data portion into the received specific model to generate a corresponding decision result. The client may associate the input data portion with the generated decision result.

[0099] According to some embodiments of the present disclosure, during the process of complementing the local dataset, the client may not participate in one or more rounds of the federated learning process. For example, during the generation of the complemented local dataset, the client may not train the local model, or the client may not send the trained local model to the server, or the client may instruct the server not to aggregate the local model received from the client. After generating the complemented local dataset, the client may send a completion indication to the server. The client may re-participate in subsequent rounds of the federated learning process. In this way, it is possible to avoid aggregating the local model trained by the client based on an incomplete training dataset into the global model.

[0100] According to some alternative embodiments of the present disclosure, during the process of complementing the local dataset, the client may participate in the federated learning process but not use at least a non-complemented portion of the local dataset to train the local model. For example, the training dataset used by the client to train the local model may include a first input data portion in the local dataset that has a corresponding output data portion, and does not include a second input data portion that does not have a corresponding output data portion. The client may use the second input data portion to train the local model only after complementing the output data portion corresponding to the second input data portion.

[0101] It should be understood that the above description is only an exemplary embodiment of method 600. Details of each step of method 600 and additional or optional steps will be further described below.

[0102] Figure 7 A flowchart of method 700 according to some embodiments of the present disclosure is shown. According to some embodiments, method 700 may be implemented by the aforementioned electronic device 400. Correspondingly, method 700 may be executed by the processing circuit 430 of the electronic device 400.

[0103] Method 700 may start from step 720. In step 710, the electronic device 400 may be configured to receive one or more sets of local distribution parameters. Each set of local distribution parameters in the one or more sets of local distribution parameters characterizes the data distribution characteristics of the local data set of the corresponding client in one or more clients.

[0104] Method 700 may proceed to step 720. In step 720, the electronic device 400 may be configured to send a task identifier to the model library. The task identifier may identify a specific training task associated with the one or more clients. The model library maintains one or more trained models. The one or more models may be trained for different machine learning tasks.

[0105] For example, the one or more clients may include a first group of clients configured to perform a first training task (e.g., training a face recognition model based on a local image set) and associated with a first task identifier. Correspondingly, the electronic device 400 may be configured to send the first task identifier associated with the first training task to the model library.

[0106] Method 700 may proceed to step 730. In step 730, the electronic device 400 may be configured to receive from the model library one or more trained candidate models corresponding to the task identifier.

[0107] The one or more trained candidate models may be retrieved from the model library based on the task identifier for return to the electronic device 400. For example, for the aforementioned first group of clients, the one or more trained candidate models may include a set of trained face recognition models.

[0108] Method 700 may proceed to step 740. In step 740, the electronic device 400 may be configured to send a model list. The model list may include a matching model selected from the one or more candidate models for each of the one or more clients, where the matching model is determined at least in part based on the corresponding local distribution parameter set in the one or more sets of local distribution parameters.

[0109] For example, for the first client among the one or more clients, the model list may include a first matching model selected based on the first local distribution parameter set of the first local data set of the first client. For the second client among the one or more clients, the model list may include a second matching model selected based on the second local distribution parameter set of the second local data set of the second client. If there is no corresponding matching model for a specific client among the one or more clients, the model list may include a default model associated with the task identifier.

[0110] According to some embodiments of the present disclosure, each of the one or more candidate models received by the electronic device 400 is trained using a corresponding training dataset in the one or more training datasets. To determine a matching model for each client, in the one or more training datasets, the electronic device 400 may determine a specific training dataset that has the highest similarity to the local dataset of the client. Then, the electronic device 400 may determine the candidate model corresponding to the specific training dataset among the one or more candidate models as the matching model for the client.

[0111] According to some embodiments of the present disclosure, for example, the similarity may be determined at least in part based on a set of local distribution parameters of the local dataset and a set of specific distribution parameters of the specific training dataset. For example, for a first client among the one or more clients, the electronic device 400 may select a first training dataset based on a first set of local distribution parameters of the first local dataset of the first client and a set of distribution parameters of the training dataset of each candidate model, and the first training dataset has the highest similarity to the first local dataset. Then, the electronic device 400 may determine the first candidate model trained based on the first training dataset as the first matching model for the first client. For a second client among the one or more clients, the electronic device 400 may select a second training dataset based on a second set of local distribution parameters of the second local dataset of the second client and a set of distribution parameters of the training dataset of each candidate model, and the second training dataset has the highest similarity to the second local dataset. Then, the electronic device 400 may determine the second candidate model trained based on the second training dataset as the second matching model for the second client.

[0112] According to some embodiments of the present disclosure, in step 730, the electronic device 400 may additionally receive from the model library one or more training datasets associated with the one or more trained candidate models. Each of the one or more candidate models is trained based on a corresponding training dataset in the one or more training datasets. The electronic device 400 may calculate a set of distribution parameters for each training dataset based on the received one or more training datasets. The calculated set of distribution parameters may be used together with the set of local distribution parameters of the local dataset of the client to determine the similarity between the training dataset and the local dataset of the client. In some embodiments, the electronic device 400 may receive samples (instead of all) of each training dataset in the one or more training datasets from the model library, thereby saving communication and computing overhead.

[0113] According to some other embodiments of the present disclosure, the set of distribution parameters of the corresponding training dataset in the one or more training datasets is calculated and maintained by the model library. The electronic device 400 may receive the one or more sets of distribution parameters of the one or more training datasets from the model library. In this case, the electronic device 400 does not have to receive the training dataset and does not have to calculate the set of distribution parameters of the training dataset.

[0114] The set of distribution parameters described herein (e.g., the local set of distribution parameters of the local dataset of the client, or the set of distribution parameters of the training dataset) may include one or more distribution parameters. The number and type of the one or more distribution parameters may vary depending on the dataset. As an example, the dataset may follow a Gaussian distribution, a gamma distribution, a lognormal distribution, or any other type of distribution. Accordingly, the set of distribution parameters may include distribution parameters corresponding to these distribution types. In some embodiments, the dataset follows a Gaussian distribution, and the set of distribution parameters may include the mean μ and the standard deviation δ. In some embodiments, the dataset follows a gamma distribution, and the set of distribution parameters may include the shape parameter α and the inverse scale parameter β. In some embodiments, the dataset follows a lognormal distribution, and the set of distribution parameters may include the mean μ and the variance δ. In other embodiments, the dataset may follow other types of distributions, and the set of distribution parameters may include the corresponding distribution parameters.

[0115] According to some embodiments of the present disclosure, the local set of distribution parameters of the client may further include a type indicator indicating the distribution type. The electronic device 400 may calculate the set of distribution parameters of the training dataset associated with the candidate model based on the distribution type indicated by the type indicator.

[0116] According to some embodiments, in step 710, the electronic device 400 may receive the one or more local sets of distribution parameters from a server for machine learning. The server may be associated with the one or more clients. For example, the server may be a server for federated learning, and the one or more clients may be training devices participating in the federated learning.

[0117] According to some embodiments, the task identifier sent by the electronic device 400 in step 720 may be received from the server for machine learning. For example, the task identifier may be received from the server together with the local set of distribution parameters.

[0118] According to some embodiments, in step 740, the electronic device 400 may send the matching model of each of the one or more clients together with the client identifier of the client to a recipient (e.g., a server). For example, the list of models sent by the electronic device 400 may additionally include the identifier of each of the one or more clients. Each identifier may be stored in the model list in association with the matching model of the client for which the identifier is used. This enables the recipient of the model list to identify the corresponding matching model for each client.

[0119] The techniques of the present disclosure may be performed in a cellular network. For example, the electronic device 400 may be implemented at the network data analytics function (NWDAF) of the cellular network. The model library may reside at the analytics data repository function (ADRF) of the cellular network. The server for machine learning may be deployed with the application function (AF) of the cellular network. In some embodiments, the electronic device 400 at the NWDAF may communicate directly with the server. In other embodiments, the electronic device 400 at the NWDAF may communicate with the server via the network exposure function (NEF) of the cellular network. The specific receiving path may be determined based on the AF deployed at the server. For example, if the AF is trusted, the electronic device 400 may directly receive the set of distribution parameters and the task identifier from the server. Otherwise, the electronic device 400 may receive the set of distribution parameters and the task identifier from the server via the NEF.

[0120] It should be understood that the above-described are merely exemplary embodiments of the method 700. Details of the various steps of the method 700 and additional or alternative steps will be further described below.

[0121] 4. Exemplary Process

[0122] Figure 8 A schematic diagram of a process 800 according to some embodiments of the present disclosure is shown. The process 800 is performed in the context of a cellular network. The cellular network may include an ADRF, an NWDAF, an NEF, etc. It should be understood that the cellular network is merely an example of a communication environment in which the techniques of the present disclosure may be performed. In other embodiments, the techniques of the present disclosure may be performed in other communication environments without limitation.

[0123] The process 800 may involve servers and clients 1-N for machine learning. The server and the clients may be implemented by the aforementioned electronic device 200 and electronic device 300 respectively. The NWDAF may be implemented by the aforementioned electronic device 400. In this example of the cellular network, the model library may reside at the ADRF, and the server may be deployed with the AF. The client may reside at one or more user equipments (UEs). Each UE may access the cellular network through a base station (e.g., gNB) and communicate with other network element devices. In other examples, one or more of the model library, the server, and the client may reside at other suitable locations.

[0124] The process 800 may start at step 801. In step 801, the server may send a trigger notification to the client. The trigger notification may be used to initiate the execution of the process 800. In some embodiments, the server may send the trigger notification periodically. In other embodiments, the server may send the trigger notification based on a trigger request from the client. For example, if more than a threshold ratio (e.g., 50%) of the clients 1-N send a trigger request, the server may send the trigger notification. Alternatively, the server may send a trigger notification for a single client individually based on the trigger request of that client.

[0125] In some embodiments, when the client is a training device for federated learning, the client may send a trigger request based on checking the local dataset. For example, if the local dataset lacks an output data part, the client may send a trigger request to trigger data completion. Specifically, the trigger request may be sent when the completeness of the local dataset is lower than a threshold. In other embodiments, when the client is a model user, the client may send a trigger request based on the local lack of an available model or the performance of the local available model being lower than a threshold.

[0126] In step 802, in response to receiving the trigger request, the client may generate a local distribution parameter set of the local dataset.

[0127] The data sets described in this disclosure (e.g., local data sets or training data sets) may include various types of data. In some embodiments, the data set may include media data, which may include but are not limited to text data, image data, audio data, video data, tactile data, and so on. In some embodiments, the data set may include sensor data, which may include data collected by one or more sensors, and the data may describe one or more physical or chemical parameters, including but not limited to temperature, humidity, speed, acceleration, altitude, electrical parameters, magnetic parameters, purity, pH value, concentration, and so on. In some embodiments, the data set may include communication data, which may include but are not limited to measurement data or metric data related to the communication environment, communication quality, and communication status. In some embodiments, the data set may include user data, which may include but are not limited to user behavior data, user preference data, user analysis data, and so on. In some embodiments, the data set may include a combination of these data. In some embodiments, the data set may include data after preprocessing these data.

[0128] As previously mentioned, the local data set may follow various types of distributions. For example, the data set may follow a Gaussian distribution, a gamma distribution, a lognormal distribution, or any other type of distribution. Accordingly, the local distribution parameter set may include distribution parameters corresponding to these distribution types. In some embodiments, the local distribution parameter set may include the mean μ and the standard deviation δ for the Gaussian distribution. In some embodiments, the local distribution parameter set may include the shape parameter α and the inverse scale parameter β for the gamma distribution. In some embodiments, the local distribution parameter set may include the mean μ and the variance δ for the lognormal distribution. In other embodiments, the local data set may follow other types of distributions, and the local distribution parameter set may include the corresponding distribution parameters. The client may generate a local distribution parameter set characterizing the local data set based on the entire local data set or a sample thereof.

[0129] In step 803, the client may send a distribution parameter set response to the server. Each distribution parameter set response includes the corresponding local distribution parameter set of the corresponding client. In some embodiments, the corresponding local distribution parameter set further includes a type indicator corresponding to the distribution type.

[0130] In some alternative embodiments, in response to a triggered notification, instead of performing steps 802 and 803, the client may send the local data set (or a portion thereof) to the server, and the server may generate the corresponding local distribution parameter set.

[0131] In step 804, the server may send the distribution parameter list together with the task identifier to the NWDAF. For example, the server may integrate the obtained local distribution parameter sets to generate a distribution parameter list. The distribution parameter list includes one or more local distribution parameter sets associated with one or more clients. The server may also associate each local distribution parameter set with the corresponding client, thereby recording the client from which each local distribution parameter set is derived. The distribution parameter list may also include client identifiers associated with one or more clients, thereby associating each client identifier with each local distribution parameter set. In an embodiment of federated learning, the server may further associate each distribution parameter set with a specific training task (e.g., a specific task identifier) participated by the corresponding client, because a single server may serve multiple groups of clients performing multiple learning tasks.

[0132] In an embodiment of a cellular network, the server may be deployed with an AF. Depending on the attributes of the AF, step 804 may have different sub-steps. For example, if the AF is untrusted, sub-steps 804a and 804b may be performed. In sub-step 804a, the server sends the distribution parameter list together with the task identifier to the NEF. Then, in sub-step 804b, the NEF sends the distribution parameter list together with the task identifier to the NWDAF. If the AF is trusted, it is not necessary to forward the distribution parameter list together with the task identifier via the NEF. Therefore, instead of performing sub-steps 804a and 804b, sub-step 804c is performed. In step 804c, the AF sends the distribution parameter list together with the task identifier directly to the NWDAF.

[0133] In step 805, the NWDAF may send the task identifier received in step 804 to the model library. The model library may reside at the ADRF. The model library maintains one or more trained models. The one or more models may be trained for different machine learning tasks. For example, the one or more models may include dozens, hundreds, or more models. In some embodiments, the model library may also maintain one or more training data sets for training the one or more models. In some embodiments, the model library may also maintain one or more distribution parameter sets of the one or more training data sets.

[0134] In step 806, the model library may retrieve one or more candidate models corresponding to the task identifier based on the task identifier received in step 805. The one or more candidate models may be those that have been trained for tasks that are the same as or similar to the training task indicated by the task identifier. For example, if the task identifier indicates a task of training a face recognition model, the one or more candidate models may include a set of trained face recognition models maintained in the model library. In some embodiments, the model library may also retrieve one or more training data sets associated with the one or more candidate models retrieved, and each of the one or more candidate models is trained based on a corresponding training data set in the one or more training data sets. In some embodiments, the model library may also retrieve one or more sets of distribution parameters associated with the one or more candidate models (more specifically, the one or more training data sets), and each set of distribution parameters in the one or more sets of distribution parameters indicates the data distribution characteristics of the corresponding training data set.

[0135] In step 807, the model library may send a candidate model response to the NWDAF. The candidate model response may include the one or more candidate models determined in step 806. In some embodiments, the candidate model response may include one or more training data sets (or a portion of the samples) associated with the one or more candidate models. In other embodiments, the candidate model response may include one or more sets of distribution parameters associated with the one or more candidate models (more specifically, the one or more training data sets), without including the training data sets themselves.

[0136] In step 808, the NWDAF may perform model matching. The model matching may be performed based on the list of distribution parameters associated with the client received in step 804 and the one or more training data sets (or one or more sets of distribution parameters of the one or more training data sets) received in step 807. Specifically, for the local set of distribution parameters for a specific client in the list of distribution parameters, the NWDAF may calculate the similarity between the local data set of the specific client and the training data set based on the local set of distribution parameters and the set of distribution parameters of each training data set in the one or more received training data sets. Then, the NWDAF may select the specific training data set in the one or more training data sets that has the highest similarity to the local data set of the specific client. The NWDAF may determine the specific candidate model corresponding to the specific training data set among the one or more candidate models as the matching model for the specific client. The specific candidate model is trained based on the specific training data set. The NWDAF may perform model matching for each client associated with the list of distribution parameters, thereby obtaining the matching model for each client.

[0137] In some embodiments, the NWDAF receives a training data set from a model library. The NWDAF may calculate a set of distribution parameters for the received training data set. The calculation method may be the same as the method by which the client calculates the local distribution parameter set of the local data set. In some embodiments, the NWDAF may determine the distribution type based on a type indicator associated with the distribution parameter set of the local data set, such as Gaussian distribution, gamma distribution, lognormal distribution, etc. Then, the NWDAF may assume that the training data set has this distribution type and calculate one or more distribution parameters corresponding to this distribution type accordingly. In other embodiments, each data set has a consistent distribution type. Therefore, it may not depend on the type indicator.

[0138] In other embodiments, the NWDAF receives a set of distribution parameters of the training data set from the model library. Therefore, the NWDAF does not have to calculate the set of distribution parameters of the training data set.

[0139] In step 809, the one or more obtained matching models may be organized into a model list response to be sent to the server. In some embodiments, each model list response may include one or more matching models associated with a single task identifier. The NWDAF may include the one or more obtained matching models in the model list response in association with the corresponding one or more client identifiers. In other embodiments, each model list response may include one or more matching models associated with multiple task identifiers. The NWDAF may include the one or more obtained matching models in the model list response in association with one or more client identifiers and one or more task identifiers. The one or more client identifiers or one or more task identifiers may be those received by the NWDAF in step 804.

[0140] Depending on the attributes of the AF deployed at the server, step 809 may have different sub-steps. For example, if the AF is untrusted, sub-steps 809a and 809b may be executed. In sub-step 809a, the NWDAF sends the model list response to the NEF. Then, in sub-step 809b, the NEF sends the model list response to the server. If the AF is trusted, there is no need to forward the model list response via the NEF. Therefore, instead of executing sub-steps 809a and 809b, sub-step 809c is executed. In step 809c, the NWDAF sends the model list response directly to the server.

[0141] In step 810, the server may send each matching model in the received model list response to the corresponding client. For example, the server may send the matching models in the model list response to the client corresponding to the client identifier associated with the matching model. The server may determine the matching models for each client based on the client identifier (and optionally the task identifier). In some embodiments, sending the matching model may include sending a set of model parameters of the model. In other embodiments, sending the matching model may include sending the storage location associated with a set of model parameters of the model, such that the client may retrieve the set of model parameters from the storage location.

[0142] After receiving the matching model, the client may apply the matching model to the local dataset in step 811, thereby generating an output data portion.

[0143] When the client is a model user, the output data portion generated by the matching model may be used to take one or more actions, as described previously with respect to the access control device.

[0144] When the client is a model training party participating in federated learning, the output data portion may be used to complete the client's local dataset. The completed local dataset may be used as a training dataset for training the local model. Specifically, in step 812, each client may apply the received matching model to at least a portion of the local dataset (e.g., the input data portion lacking the corresponding output data portion), thereby generating an output data portion. The output data portion may be integrated with the at least a portion of the local dataset to complete the at least a portion. Optionally, the client may send a completion indication to the server after completion. Optionally, the server may send an indication allowing the client to participate in training after receiving the completion indication.

[0145] In step 813, the client may use the completed local dataset as a complete training dataset to train the local model. The local model is associated with the task identifier described previously, for example. That is, the training task identified by the client's task identifier includes training the local model. In the present disclosure, the trained local model may be of various suitable types, including machine learning models that have been developed currently and those to be developed in the future. Each artificial intelligence model may be represented by a set of model parameters characterizing the model. Additionally, the method for training the local model may be various suitable methods. The techniques of the present disclosure do not limit the type of the model and the training method. Example model types include, but are not limited to, models trained by convolutional neural networks (CNNs), deep neural networks (DNNs), recurrent neural networks (RNNs), long short-term memory (LSTMs), generative models, random forests, and their variants and enhanced algorithms.

[0146] After training the local model, at step 814, the client can send the trained local model to the server at step 814.

[0147] Then, the server can aggregate one or more trained local models received from one or more clients at step 815 to generate a global model. The generated global model can be sent to the one or more clients for use or to perform the next training round.

[0148] In process 800, the data completion process of the client may span one or more training rounds of federated learning. In some embodiments, during the data completion process, a particular client participating in the process may skip the one or more training rounds of federated learning. For example, the particular client may not train the local model or upload the local model. Additionally or alternatively, the server may not aggregate the local models from the particular client until a completion indication is received. In other embodiments, the particular client may participate in federated learning based only on a first portion of the local dataset (e.g., the first portion is complete). The incomplete second portion of the local dataset may be used for federated learning after the data completion process.

[0149] Table 1 shows example input information of the NWDAF according to some embodiments of the present disclosure. In these embodiments, the client is implemented as a user equipment (UE) associated with a user.

[0150] Table 1

[0151]

[0152] Table 2 shows example output information of the NWDAF according to some embodiments of the present disclosure.

[0153] Table 2

[0154] Output information Description Model list Matching model for the local dataset of the UE

[0155] It should be understood that the above description is only an exemplary embodiment of process 800. The technology of the present disclosure may include more or fewer steps and may be implemented in a communication environment different from the cellular environment. For example, in some embodiments, the server, NWDAF, and model library may reside in the same device, such that the technology of the present disclosure can be implemented in a simple server-client architecture. Also, for example, in some embodiments, there are intermediate elements not shown between the various elements of process 800 for relaying and forwarding. Still, for example, in some embodiments, the server may be implemented at a network element other than the AF, or the model library may be implemented at a network element other than the ADRF. Alternatively, the operations performed by the NWDAF may be performed by other network elements.

[0156] 5. Exemplary Calculation Methods

[0157] This section describes one or more exemplary calculation methods that can be used to calculate a set of distribution parameters and calculate the similarity between two data sets based on the set of distribution parameters. It should be understood that the data sets and calculation methods used in this section are merely exemplary. In other embodiments, alternative data sets and calculation methods may be used without departing from the scope of the present disclosure.

[0158] According to an embodiment of the present disclosure, the local data set of the client is represented as data set LD.

[0159] The data set LD may include N samples. The data set LD can be represented as follows:

[0160]

[0161] where N represents the number of samples, M represents the number of data points included in each sample, and Q represents the number of feature dimensions of each data point. In expression (1), N samples are listed from left to right. In each sample, M data points are listed from top to bottom. In each data point, Q values of the data point on Q feature dimensions are listed from left to right.

[0162] The local data set can contain various types of data and is accordingly organized. For example, if the data set LD is a radar measurement data set, then N represents the number of measurement samples, M can represent the number of measurement points included in each measurement sample, and Q can represent the number of feature dimensions of each measurement point. If each measurement point represents a coordinate in three-dimensional space, then Q can be equal to 3. Again, for example, if the data set LD is an image set, then N can represent the number of extracted picture samples, M can represent the number of channels of each picture sample, and Q can represent the dimension of the vector obtained by tiling the data of each channel. Tiling indicates connecting the tail of each row of the data matrix to the head of the next row to obtain a vector.

[0163] It should be understood that the above data organization method for the local data set is merely an example. Any other suitable data organization method can be adopted. For example, although the local data set LD is represented as a high-order tensor, it can be represented as a vector or a scalar without departing from the scope of the present disclosure.

[0164] In some embodiments, in order to reduce the processing load, the original local data set can be preprocessed to obtain the local data set LD. For example, the original local data set can be downsampled to reduce the value of N. Additionally or alternatively, the original local data set can be dimensionally reduced to reduce the value of M or Q. Other processes are also possible. Combinations of these processes are also possible.

[0165] The data distribution characteristics can be calculated for each feature dimension of each sample of the local dataset LD. The data distribution characteristics can be represented by a set of distribution parameters including one or more distribution parameters. In the example of this section, it is assumed that each feature dimension of the samples of the local dataset LD follows a Gaussian distribution. Accordingly, the mean (μ) and standard deviation (δ) can be used to characterize the distribution characteristics of the data of the local dataset LD. The mean and standard deviation of the data following a Gaussian distribution can be calculated by known methods. For the aforementioned dataset LD, the calculated set of distribution parameters LX can be represented as follows:

[0166]

[0167] where each pair of mean and standard deviation (μ, δ) characterizes the data distribution characteristics of the corresponding sample in the corresponding feature dimension in the dataset LD.

[0168] Figure 9 shows the correspondence between the local dataset and the local set of distribution parameters according to some embodiments of the present disclosure. As Figure 9 shown, a pair of mean and standard deviation exemplarily characterizes the distribution characteristics of the data of the 2nd sample in the 2nd feature dimension (represented by the bold box) of the local dataset LD. Other pairs of mean and standard deviation can similarly correspond to the distribution characteristics of the data of the corresponding sample in the corresponding feature dimension in the dataset LD.

[0169] It should be understood that the Gaussian distribution used in the present disclosure is merely exemplary. In practical applications, it can be assumed that the dataset LD follows other types of distributions, and the set of distribution parameters can be calculated based on this assumption. Other types of distributions can include, for example, but are not limited to, gamma distribution, lognormal distribution, etc. When the dataset LD follows a gamma distribution, the shape parameter α and the inverse scale parameter β can be used alternatively. When the dataset LD follows a lognormal distribution, the mean μ and the variance δ can be used alternatively. In addition, these distributions can be one-dimensional or multi-dimensional.

[0170] For the training dataset TD in the model library, the distribution parameter set TX of the training dataset can be calculated in the same way. As mentioned above, the training dataset TD retrieved from the model library and the local dataset LD of the client are associated with the same training task because the training dataset TD is retrieved based on the task identifier associated with the client. Therefore, the training dataset TD and the local dataset LD may adopt a similar data organization method. If the data organization methods of the training dataset retrieved from the model library and the local dataset LD are different, the training dataset can be reorganized (e.g., by NWDAF) to obtain the training dataset TD. Accordingly, the distribution parameter set TX of the training dataset TD can have a form similar to the distribution parameter set LX.

[0171] Before calculating the distribution parameter set TX, the training dataset TD can also be preprocessed (e.g., downsampling, dimensionality reduction, etc.). In the example of this section, the training dataset TD contains the same N, M, Q values as the local dataset LD. In other examples, the size of the training dataset TD can be different from that of the local dataset LD. Accordingly, one or more of the N, M, Q values of the training dataset TD can be different from those of the local dataset LD.

[0172] After obtaining the distribution parameter set LX of the local dataset LD and the distribution parameter set TX of the training dataset TD, the similarity between the local dataset LD and the training dataset TD can be measured based on the distribution parameter set LX and the distribution parameter set TX. Figure 10A and Figure 10B illustrates an example process for calculating the similarity between a local dataset and a training dataset according to an embodiment of the present disclosure.

[0173] Figure 10A illustrates an example process for calculating the similarity between a local dataset and a training dataset according to some embodiments of the present disclosure. As shown, each sample of the local dataset LD and the training dataset TD can be placed in the left and right columns respectively. For example, the local dataset LD is placed in the left column, and each sample of the training dataset TD is placed in the right column. Each sample in the left column can be connected to each sample in the right column. A corresponding weight can be assigned to each connection between each sample in the left column and each sample in the right column to establish a bipartite graph. The distance between the local dataset LD and the training dataset TD can be calculated based on this bipartite graph.

[0174] In some embodiments, the KM (Kuhn-Munkres) algorithm can be used to calculate the minimum distribution distance between the local dataset LD and the training dataset TD. The minimum distribution distance is calculated based on the weight of each connection. This minimum distribution distance can represent the similarity between the local dataset LD and the training dataset TD. The larger the minimum distribution distance, the lower the similarity between the local dataset LD and the training dataset TD. In other embodiments, other appropriate algorithms besides the KM algorithm can be used to calculate the distance between the local dataset LD and the training dataset TD, and the calculated distance can be used as an alternative measure of the similarity between the two datasets.

[0175] In some embodiments, the weight assigned to each connection can be calculated based at least in part on the distribution parameter set TX of the distribution parameter set LX. As an example, the weight can be associated with the average sample distance between two samples. In the example of this section, the weight can be equal to the average sample distance. In other examples, the weight can be a function of the average sample distance. In other examples, the weight can also be associated with one or more additional factors in addition to the distribution parameter set TX of the distribution parameter set LX.

[0176] In some embodiments, the average sample distance between two samples can be calculated based at least in part on a set of feature distances d between the respective feature dimensions of the two samples. In the examples of this section, the average sample distance can be expressed as the average of the set of feature distances d. In other examples, the average sample distance between two samples can be measured in other ways. For example, for the hth sample in the local dataset LD and the kth sample in the training dataset TD, the average sample distance between the two samples can be expressed as:

[0177]

[0178] Where h represents the index of the sample in the local dataset LD in the left column, k represents the index of the sample in the training dataset TD in the right column, and q represents the index of the feature dimension (q is less than or equal to Q). hk (q) represents the feature distance between the hth sample in the local dataset LD and the kth sample in the training dataset TD on the qth feature dimension.

[0179] In some embodiments, the characteristic distance d can be calculated based at least in part on the distribution parameter set LX and the distribution parameter set TX. For example, the characteristic distance d hk (q) can be based on the distribution parameters of the h-th sample in the q-th feature dimension in the local dataset LD (∈LX) and the distribution parameter of the kth sample of the training dataset TD in the qth feature dimension (∈TX) is calculated. For example, the feature distance d hk (q) can be calculated as the Euclidean distance between the two-dimensional coordinate points represented by these two distribution parameters.

[0180] Figure 10B Shows an example of calculating the feature distance between two samples of a local dataset and a training dataset according to some embodiments of the present disclosure.

[0181] As shown in the figure, each element in the distribution parameter set LX of the local dataset LD can be placed in a two-dimensional coordinate system with μ as the abscissa and δ as the ordinate. Each element in the distribution parameter set TX of the training dataset TD can also be placed in this two-dimensional coordinate system. Each element can be represented as a point in this two-dimensional coordinate system where n represents the index of the sample, and q represents the index of the feature dimension. For example, n is a positive integer not greater than N, and q is a positive integer not greater than Q. In Figure 10B , the dots can represent the elements (with subscript i) from the distribution parameter set TX, and the squares can represent the elements (with subscript j) from the distribution parameter set LX. The distance d can be represented as the Euclidean distance between two elements (one element from the parameter set TX and the other element from the parameter set LX) in the two-dimensional coordinate system. The distance d can be used to represent the feature distance between two data samples (one sample from the dataset TD and the other sample from the dataset LD) characterized by these two elements in a certain feature dimension. As an example, the shown distance represents the feature distance between the 1st sample of the dataset LD and the 3rd sample of the dataset TD for the q-th feature dimension.

[0182] This distance d can be calculated for each element of the distribution parameter set LX and each element of the distribution parameter set TX. The average feature distance between two samples between the local dataset LD and the training dataset TD can be determined based on the calculated set of distances d, as shown in expression (3). The calculated average feature distance can be used as a weight to calculate the similarity between the local dataset LD and the training dataset TD, as previously described regarding [[ID=CH=38]]Figure 10A has been described.

[0183] As previously described, when there are multiple training datasets TD, a specific training dataset with the highest similarity to the local dataset LD among the multiple training datasets TD can be selected. Then, a specific model that has been trained based on this specific training dataset can be selected as the matching model for the local dataset LD (or the client holding this dataset).

[0184] It should be understood that the calculation methods described above are merely one or more exemplary embodiments of the technology of the present disclosure. One or more steps of the above calculation methods may be modified or replaced without departing from the scope of the present disclosure.

[0185] 6. Application Product Examples

[0186] The technology of the present disclosure can be applied to various products.

[0187] For example, the control device / base station mentioned in the present disclosure can be implemented as any type of base station, such as an eNB, such as a macro eNB and a small eNB. A small eNB can be an eNB that covers a cell smaller than a macro cell, such as a pico eNB, a micro eNB, and a home (femto) eNB. Also for example, it can be implemented as a gNB, such as a macro gNB and a small gNB. A small gNB can be a gNB that covers a cell smaller than a macro cell, such as a pico gNB, a micro gNB, and a home (femto) gNB. Instead, the base station can be implemented as any other type of base station, such as a NodeB and a Base Transceiver Station (BTS). The base station can include: a main body configured to control wireless communication (also referred to as a base station device); and one or more Remote Radio Heads (RRHs) provided in a place different from the main body. In addition, various types of terminals described below can all act as a base station by temporarily or semi-persistently performing base station functions. For example, the terminal device mentioned in the present disclosure can be implemented as a mobile terminal (such as a smart phone, a tablet personal computer (PC), a notebook PC, a portable game terminal, a portable / dongle-type mobile router, and a digital camera device) or a vehicle-mounted terminal (such as a car navigation device) in some embodiments. The terminal device can also be implemented as a terminal that performs machine-to-machine (M2M) communication (also referred to as a machine type communication (MTC) terminal). In addition, the terminal device can be a wireless communication module (such as an integrated circuit module including a single chip) installed on each of the above terminals.

[0188] The application examples according to the present disclosure will be described below with reference to the drawings.

[0189] [Examples of Base Stations]

[0190] It should be understood that the term "base station" in the present disclosure has the full breadth of its ordinary meaning and at least includes a wireless communication station used as part of a wireless communication system or radio system to facilitate communication. Examples of base stations can be, for example, but not limited to the following: A base station can be one or both of a base transceiver station (BTS) and a base station controller (BSC) in a GSM system, can be one or both of a radio network controller (RNC) and a Node B in a WCDMA system, can be an eNB in an LTE and LTE-Advanced system, or can be a corresponding network node in a future communication system (such as a gNB, eLTE eNB, etc. that may appear in a 5G communication system). Some functions of the base stations in the present disclosure can also be implemented as an entity having a control function for communication in D2D, M2M, and V2V communication scenarios, or as an entity that plays a role in spectrum coordination in a cognitive radio communication scenario.

[0191] First example

[0192] Figure 11 FIG. 7 is a block diagram of a first example showing an exemplary configuration of a gNB to which the techniques of the present disclosure can be applied. The gNB 2100 includes a plurality of antennas 2110 and a base station device 2120. The base station device 2120 and each antenna 2110 can be connected to each other via RF cables. In one implementation, the gNB 2100 (or the base station device 2120) here can correspond to the control-side electronic device described above.

[0193] Each of the antennas 2110 includes a single or multiple antenna elements (such as the multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna) and is used for the base station device 2120 to transmit and receive wireless signals. As Figure 11 shown, the gNB 2100 can include a plurality of antennas 2110. For example, the plurality of antennas 2110 can be compatible with multiple frequency bands used by the gNB 2100.

[0194] The base station device 2120 includes a controller 2121, a memory 2122, a network interface 2123, and a wireless communication interface 2125.

[0195] The controller 2121 can be, for example, a CPU or a DSP, and operates various functions of the higher layers of the base station device 2120. For example, the controller 2121 determines the location information of the target terminal device among at least one terminal device based on the location information of at least one terminal device on the terminal side in the radio communication system and the specific location configuration information of at least one terminal device acquired by the radio communication interface 2125. The controller 2121 can have a logical function to execute controls such as radio resource control, radio bearer control, mobility management, access control, and scheduling. The control can be executed in combination with a nearby gNB or a core network node. The memory 2122 includes a RAM and a ROM, and stores programs executed by the controller 2121 and various types of control data (such as a terminal list, transmission power data, and scheduling data).

[0196] The network interface 2123 is a communication interface for connecting the base station device 2120 to the core network 2124. The controller 2121 can communicate with a core network node or another gNB via the network interface 2123. In this case, the gNB 2100 and the core network node or other gNBs can be connected to each other through logical interfaces (such as the S1 interface and the X2 interface). The network interface 2123 can also be a wired communication interface or a radio communication interface for a radio backhaul line. If the network interface 2123 is a radio communication interface, compared with the frequency band used by the radio communication interface 2125, the network interface 2123 can use a higher frequency band for radio communication.

[0197] The radio communication interface 2125 supports any cellular communication scheme (such as Long-Term Evolution (LTE) and LTE-Advanced), and provides a radio connection to terminals located in the cell of the gNB 2100 via the antenna 2110. The radio communication interface 2125 generally can include, for example, a baseband (BB) processor 2126 and an RF circuit 2127. The BB processor 2126 can execute, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and execute various types of signal processing of layers (such as L1, Media Access Control (MAC), Radio Link Control (RLC), and Packet Data Convergence Protocol (PDCP)). Instead of the controller 2121, the BB processor 2126 can have a part or all of the above logical functions. The BB processor 2126 can be a memory storing a communication control program, or a module including a processor configured to execute the program and related circuits. The update program can change the function of the BB processor 2126. The module can be a card or a blade inserted into a slot of the base station device 2120. Alternatively, the module can also be a chip mounted on the card or the blade. At the same time, the RF circuit 2127 can include, for example, mixers, filters, and amplifiers, and transmit and receive radio signals via the antenna 2110. AlthoughFigure 11 An example is shown where an RF circuit 2127 is connected to an antenna 2110. However, the present disclosure is not limited to this illustration, and an RF circuit 2127 can be connected to multiple antennas 2110 simultaneously.

[0198] As Figure 11 shown, the wireless communication interface 2125 can include multiple BB processors 2126. For example, the multiple BB processors 2126 can be compatible with multiple frequency bands used by the gNB 2100. As Figure 11 shown, the wireless communication interface 2125 can include multiple RF circuits 2127. For example, the multiple RF circuits 2127 can be compatible with multiple antenna elements. Although Figure 11 an example is shown where the wireless communication interface 2125 includes multiple BB processors 2126 and multiple RF circuits 2127, the wireless communication interface 2125 can also include a single BB processor 2126 or a single RF circuit 2127.

[0199] Second example

[0200] Figure 12 is a block diagram of a second example showing an exemplary configuration of a gNB to which the techniques of the present disclosure can be applied. The gNB 2200 includes multiple antennas 2210, RRH 2220, and a base station device 2230. The RRH 2220 and each antenna 2210 can be connected to each other via RF cables. The base station device 2230 and the RRH 2220 can be connected to each other via a high-speed line such as an optical fiber cable. In one implementation, the gNB 2200 (or the base station device 2230) here can correspond to the control-side electronic device described above.

[0201] Each of the antennas 2210 includes a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used to transmit and receive wireless signals for the RRH 2220. As Figure 12 shown, the gNB 2200 can include multiple antennas 2210. For example, the multiple antennas 2210 can be compatible with multiple frequency bands used by the gNB 2200.

[0202] The base station device 2230 includes a controller 2231, a memory 2232, a network interface 2233, a wireless communication interface 2234, and a connection interface 2236. The controller 2231, the memory 2232, and the network interface 2233 are the same as the controller 2121, the memory 2122, and the network interface 2123 described with reference to Figure 11 above.

[0203] The wireless communication interface 2234 supports any cellular communication scheme (such as LTE and LTE-Advanced), and provides wireless communication to terminals located in the sector corresponding to the RRH 2220 via the RRH 2220 and the antenna 2210. The wireless communication interface 2234 generally may include, for example, a BB processor 2235. Except that the BB processor 2235 is connected to the RF circuit 2222 of the RRH 2220 via the connection interface 2236, the BB processor 2235 is the same as the BB processor 2126 described with reference to Figure 11 As shown in Figure 12 , the wireless communication interface 2234 may include a plurality of BB processors 2235. For example, the plurality of BB processors 2235 may be compatible with multiple frequency bands used by the gNB 2200. Although Figure 12 shows an example in which the wireless communication interface 2234 includes a plurality of BB processors 2235, the wireless communication interface 2234 may also include a single BB processor 2235.

[0204] The connection interface 2236 is an interface for connecting the base station device 2230 (wireless communication interface 2234) to the RRH 2220. The connection interface 2236 may also be a communication module for communication in the above-mentioned high-speed line for connecting the base station device 2230 (wireless communication interface 2234) to the RRH 2220.

[0205] The RRH 2220 includes a connection interface 2223 and a wireless communication interface 2221.

[0206] The connection interface 2223 is an interface for connecting the RRH 2220 (wireless communication interface 2221) to the base station device 2230. The connection interface 2223 may also be a communication module for communication in the above-mentioned high-speed line.

[0207] The wireless communication interface 2221 transmits and receives wireless signals via the antenna 2210. The wireless communication interface 2221 generally may include, for example, an RF circuit 2222. The RF circuit 2222 may include, for example, a mixer, a filter, and an amplifier, and transmits and receives wireless signals via the antenna 2210. Although Figure 12 shows an example in which one RF circuit 2222 is connected to one antenna 2210, the present disclosure is not limited to this illustration, but one RF circuit 2222 may be connected to multiple antennas 2210 simultaneously.

[0208] As shown in Figure 12 , the wireless communication interface 2221 may include a plurality of RF circuits 2222. For example, the plurality of RF circuits 2222 may support multiple antenna elements. Although Figure 12An example is shown in which the wireless communication interface 2221 includes a plurality of RF circuits 2222, but the wireless communication interface 2221 may also include a single RF circuit 2222.

[0209] [Examples of user equipment / terminal equipment]

[0210] First example

[0211] Figure 13 FIG. is a block diagram showing an example of an exemplary configuration of a communication device 2300 (e.g., a smart phone, a communicator, etc.) to which the techniques of the present disclosure may be applied. The communication device 2300 includes a processor 2301, a memory 2302, a storage device 2303, an external connection interface 2304, a camera device 2306, a sensor 2307, a microphone 2308, an input device 2309, a display device 2310, a speaker 2311, a wireless communication interface 2312, one or more antenna switches 2315, one or more antennas 2316, a bus 2317, a battery 2318, and an auxiliary controller 2319. In one implementation, the communication device 2300 (or the processor 2301) here may correspond to the above-described transmitting device or the terminal-side electronic device.

[0212] The processor 2301 may be, for example, a CPU or a system-on-chip (SoC), and controls the functions of the application layer and other layers of the communication device 2300. The memory 2302 includes a RAM and a ROM, and stores data and programs executed by the processor 2301. The storage device 2303 may include a storage medium such as a semiconductor memory and a hard disk. The external connection interface 2304 is an interface for connecting an external device (such as a memory card and a universal serial bus (USB) device) to the communication device 2300.

[0213] The camera device 2306 includes an image sensor (such as a charge-coupled device (CCD) and a complementary metal oxide semiconductor (CMOS)), and generates a captured image. The sensor 2307 may include a set of sensors such as a measurement sensor, a gyro sensor, a geomagnetic sensor, and an acceleration sensor. The microphone 2308 converts the sound input to the communication device 2300 into an audio signal. The input device 2309 includes, for example, a touch sensor configured to detect a touch on the screen of the display device 2310, a keypad, a keyboard, a button, or a switch, and receives operations or information input from a user. The display device 2310 includes a screen (such as a liquid crystal display (LCD) and an organic light-emitting diode (OLED) display), and displays an output image of the communication device 2300. The speaker 2311 converts the audio signal output from the communication device 2300 into sound.

[0214] The wireless communication interface 2312 supports any cellular communication scheme (such as LTE and LTE-Advanced) and performs wireless communication. The wireless communication interface 2312 generally may include, for example, a BB processor 2313 and an RF circuit 2314. The BB processor 2313 may perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and perform various types of signal processing for wireless communication. At the same time, the RF circuit 2314 may include, for example, mixers, filters, and amplifiers, and transmit and receive wireless signals via an antenna 2316. The wireless communication interface 2312 may be a single chip module on which the BB processor 2313 and the RF circuit 2314 are integrated. As Figure 13 shown, the wireless communication interface 2312 may include a plurality of BB processors 2313 and a plurality of RF circuits 2314. Although Figure 13 illustrates an example in which the wireless communication interface 2312 includes a plurality of BB processors 2313 and a plurality of RF circuits 2314, the wireless communication interface 2312 may also include a single BB processor 2313 or a single RF circuit 2314.

[0215] In addition, in addition to the cellular communication scheme, the wireless communication interface 2312 may support other types of wireless communication schemes, such as short-range wireless communication schemes, near-field communication schemes, and wireless local area network (LAN) schemes. In this case, the wireless communication interface 2312 may include a BB processor 2313 and an RF circuit 2314 for each wireless communication scheme.

[0216] Each of the antenna switches 2315 switches the connection destination of the antenna 2316 among a plurality of circuits included in the wireless communication interface 2312 (for example, circuits for different wireless communication schemes).

[0217] Each of the antennas 2316 includes a single or a plurality of antenna elements (such as a plurality of antenna elements included in a MIMO antenna), and is used for the wireless communication interface 2312 to transmit and receive wireless signals. As Figure 13 shown, the communication device 2300 may include a plurality of antennas 2316. Although Figure 13 illustrates an example in which the communication device 2300 includes a plurality of antennas 2316, the communication device 2300 may also include a single antenna 2316.

[0218] In addition, the communication device 2300 may include an antenna 2316 for each wireless communication scheme. In this case, the antenna switch 2315 may be omitted from the configuration of the communication device 2300.

[0219] The bus 2317 connects the processor 2301, the memory 2302, the storage device 2303, the external connection interface 2304, the imaging device 2306, the sensor 2307, the microphone 2308, the input device 2309, the display device 2310, the speaker 2311, the wireless communication interface 2312, and the auxiliary controller 2319 to each other. The battery 2318 supplies power to Figure 13 each block of the communication device 2300 shown via a feeder line, which is partially shown as a dashed line in the figure. The auxiliary controller 2319 operates the minimum necessary functions of the communication device 2300, for example, in the sleep mode.

[0220] Second example

[0221] Figure 14 is a block diagram showing an example of an exemplary configuration of an in-vehicle navigation device 2400 to which the technology of the present disclosure can be applied. The in-vehicle navigation device 2400 includes a processor 2401, a memory 2402, a Global Positioning System (GPS) module 2404, a sensor 2405, a data interface 2406, a content player 2407, a storage medium interface 2408, an input device 2409, a display device 2510, a speaker 2411, a wireless communication interface 2413, one or more antenna switches 2416, one or more antennas 2417, and a battery 2418. In one implementation, the in-vehicle navigation device 2400 (or the processor 2401) here can correspond to a transmitting device or a terminal-side electronic device.

[0222] The processor 2401 can be, for example, a CPU or an SoC, and controls the navigation function and other functions of the in-vehicle navigation device 2400. The memory 2402 includes a RAM and a ROM, and stores data and programs executed by the processor 2401.

[0223] The GPS module 2404 uses GPS signals received from GPS satellites to measure the position (such as latitude, longitude, and altitude) of the in-vehicle navigation device 2400. The sensor 2405 can include a set of sensors, such as a gyro sensor, a geomagnetic sensor, and an air pressure sensor. The data interface 2406 is connected to, for example, an in-vehicle network 2421 via a terminal (not shown), and acquires data generated by the vehicle (such as vehicle speed data).

[0224] The content player 2407 reproduces content stored in a storage medium (such as a CD and a DVD) inserted into the storage medium interface 2408. The input device 2409 includes, for example, a touch sensor, buttons, or switches configured to detect touches on the screen of the display device 2510, and receives operations or information input from the user. The display device 2510 includes a screen such as an LCD or an OLED display, and displays images of the navigation function or the reproduced content. The speaker 2411 outputs sounds of the navigation function or the reproduced content.

[0225] The wireless communication interface 2413 supports any cellular communication scheme (such as LTE and LTE-Advanced) and performs wireless communication. The wireless communication interface 2413 generally may include, for example, a BB processor 2414 and an RF circuit 2415. The BB processor 2414 may perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and perform various types of signal processing for wireless communication. Meanwhile, the RF circuit 2415 may include, for example, mixers, filters, and amplifiers, and transmit and receive wireless signals via the antenna 2417. The wireless communication interface 2413 may also be a single chip module on which the BB processor 2414 and the RF circuit 2415 are integrated. As Figure 14 shown, the wireless communication interface 2413 may include multiple BB processors 2414 and multiple RF circuits 2415. Although Figure 14 an example in which the wireless communication interface 2413 includes multiple BB processors 2414 and multiple RF circuits 2415 is shown, the wireless communication interface 2413 may also include a single BB processor 2414 or a single RF circuit 2415.

[0226] In addition, in addition to the cellular communication scheme, the wireless communication interface 2413 may support other types of wireless communication schemes, such as short-range wireless communication schemes, near-field communication schemes, and wireless LAN schemes. In this case, for each wireless communication scheme, the wireless communication interface 2413 may include a BB processor 2414 and an RF circuit 2415.

[0227] Each of the antenna switches 2416 switches the connection destination of the antenna 2417 among multiple circuits (such as circuits for different wireless communication schemes) included in the wireless communication interface 2413.

[0228] Each of the antennas 2417 includes a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used for the wireless communication interface 2413 to transmit and receive wireless signals. As Figure 14 shown, the car navigation device 2400 may include multiple antennas 2417. Although Figure 14An example is shown in which the vehicle navigation device 2400 includes a plurality of antennas 2417, but the vehicle navigation device 2400 may also include a single antenna 2417.

[0229] In addition, the vehicle navigation device 2400 may include an antenna 2417 for each wireless communication scheme. In this case, the antenna switch 2416 may be omitted from the configuration of the vehicle navigation device 2400.

[0230] The battery 2418 supplies power to Figure 14 the respective blocks of the illustrated vehicle navigation device 2400 via a feeder line, which is partially shown as a dashed line in the figure. The battery 2418 accumulates power supplied from the vehicle.

[0231] The technology of the present disclosure may also be implemented as an in-vehicle system (or vehicle) 2420 including one or more blocks of the vehicle navigation device 2400, the in-vehicle network 2421, and the vehicle module 2422. The vehicle module 2422 generates vehicle data (such as vehicle speed, engine speed, and fault information) and outputs the generated data to the in-vehicle network 2421.

[0232] The exemplary embodiments of the present disclosure have been described above with reference to the accompanying drawings, but the present disclosure is of course not limited to the above examples. Those skilled in the art can obtain various changes and modifications within the scope of the appended claims, and it should be understood that these changes and modifications will naturally fall within the technical scope of the present disclosure.

[0233] It should be understood that machine-executable instructions in a machine-readable storage medium or program product according to an embodiment of the present disclosure may be configured to perform operations corresponding to the above device and method embodiments. When referring to the above device and method embodiments, embodiments of the machine-readable storage medium or program product are clear to those skilled in the art, and thus will not be described repeatedly. The machine-readable storage medium and program product for carrying or including the above machine-executable instructions also fall within the scope of the present disclosure. Such a storage medium may include, but is not limited to, a floppy disk, an optical disk, a magneto-optical disk, a memory card, a memory stick, and the like.

[0234] In addition, it should be understood that the above-described series of processes and devices may also be implemented by software and / or firmware. In the case of implementation by software and / or firmware, a corresponding program constituting the corresponding software is stored in a storage medium of a related device, and when the program is executed, various functions can be performed.

[0235] For example, a plurality of functions included in one unit in the above embodiments may be implemented by separate devices. Alternatively, a plurality of functions implemented by a plurality of units in the above embodiments may be implemented by separate devices respectively. In addition, one of the above functions may be implemented by a plurality of units. Needless to say, such a configuration is included in the technical scope of the present disclosure.

[0236] In this specification, the steps described in the flowcharts include not only processes executed sequentially in a time series but also processes executed in parallel or individually rather than necessarily in a time series. Further, even in steps processed in a time series, needless to say, the order can be appropriately changed.

[0237] 7. Exemplary Embodiments of the Present Disclosure

[0238] 1. An electronic device, comprising: at least one processing unit; and at least one storage unit including computer program code which, when executed by the at least one processing unit, causes the electronic device to perform the following operations: obtaining a set of local distribution parameters of a second device, the set of local distribution parameters characterizing data distribution characteristics of a local data set of the second device; and sending a trained specific model to the second device, wherein the specific model is determined at least in part based on the set of local distribution parameters.

[0239] 2. The electronic device according to Embodiment 1, wherein the operations further include: sending the set of local distribution parameters together with a client identifier of the second device; and receiving a model list including the specific model.

[0240] 3. The electronic device according to Embodiment 1, wherein the electronic device is configured to be a server implementing a federated learning process, and the second device is configured to be one of a plurality of clients implementing the federated learning process.

[0241] 4. The electronic device according to Embodiment 3, wherein the operations further include: sending the set of local distribution parameters together with a task identifier identifying a specific training task associated with the second device in the federated learning process; and receiving the specific model, wherein the specific model is determined at least in part based on the task identifier.

[0242] 5. The electronic device according to Embodiment 3, wherein the operations further include: receiving a trained local model from the second device, the local model being trained by the second device based on a complemented local data set, wherein the complemented local data set is generated by the specific model based on at least a part of the local data set; and generating a global model at least in part based on aggregation of the local model.

[0243] 6. The electronic device as described in Embodiment 1, wherein the electronic device is configured to implement the application function AF of the cellular network, and the operation further includes: if the AF is not trusted, sending the local distribution parameter set to the network data analysis function NWDAF of the cellular network via the network exposure function NEF of the cellular network; and if the AF is trusted, directly sending the local distribution parameter set to the NWDAF.

[0244] 7. The electronic device as described in Embodiment 1, wherein the specific model is selected from one or more candidate models, and in one or more training data sets used to train the one or more candidate models, the specific training data set used to train the specific model has the highest similarity to the local data set of the second device.

[0245] 8. The electronic device as described in Embodiment 7, wherein the similarity is determined at least in part based on: (1) the local distribution parameter set of the local data set, and (2) the specific distribution parameter set of the specific training data set.

[0246] 9. An electronic device, comprising: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code, wherein when the computer program code is executed by the at least one processing unit, the electronic device performs the following operations: sending a local distribution parameter set to a first device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the electronic device; and receiving a trained specific model from the first device, wherein the specific model is determined at least in part based on the local distribution parameter set.

[0247] 10. The electronic device as described in Embodiment 9, wherein the operation further includes: using the received specific model to generate an output data portion based on at least a part of the local data set.

[0248] 11. The electronic device as described in Embodiment 10, wherein the output data portion includes labels, predictions, and / or decisions generated based on at least a part of the local data set.

[0249] 12. The electronic device as described in Embodiment 10, wherein the operation further includes: using the output data portion to complete the local data set; using the completed local data set as a training data set to train the local model of the electronic device; and sending the trained local model to the first device for aggregation to generate a global model for the federated learning process.

[0250] 13. The electronic device as described in Embodiment 12, wherein the operation further includes: during the process of supplementing the local data set, not participating in the federated learning process.

[0251] 14. The electronic device as described in Embodiment 12, wherein the operation further includes: during the process of supplementing the local data set, not using at least a part of the local data set that has not been supplemented to train the local model.

[0252] 15. The electronic device as described in Embodiment 9, wherein the specific model is selected from one or more candidate models, and in one or more training data sets used to train the one or more candidate models, the specific training data set used to train the specific model has the highest similarity with the local data set of the second device.

[0253] 16. The electronic device as described in Embodiment 15, wherein the similarity is determined at least in part based on: (1) the local distribution parameter set of the local data set, and (2) the specific distribution parameter set of the specific training data set.

[0258] 17. An electronic device, comprising: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code, wherein when the computer program code is executed by the at least one processing unit, the electronic device performs the following operations: receiving one or more local distribution parameter sets, each local distribution parameter set in the one or more local distribution parameter sets characterizing the data distribution characteristics of the local data set of a corresponding client among one or more clients; sending a task identifier to a model library, the task identifier identifying a training task associated with the one or more clients; receiving from the model library one or more trained candidate models corresponding to the task identifier; and sending a model list, the model list including a matching model selected from the one or more candidate models for each client among the one or more clients, the matching model being determined at least in part based on the corresponding local distribution parameter set in the one or more local distribution parameter sets.

[0255] 18. The electronic device as described in Embodiment 17, wherein the operation further includes: for each client among the one or more clients: determining a specific training data set in the one or more training data sets that has the highest similarity with the local data set of the client; and determining the candidate model corresponding to the specific training data set among the one or more candidate models as the matching model for the client.

[0256] 19. The electronic device as described in embodiment 18, wherein the similarity between the local data set of the client and the specific training data set is determined at least in part based on: (1) the local distribution parameter set of the local data set, and (2) the specific distribution parameter set of the specific training data set.

[0257] 20. The electronic device as described in embodiment 19, wherein: the specific distribution parameter set is received from the model library; or the specific distribution parameter set is determined based on the specific training data set in the one or more training data sets received from the model library.

[0258] 21. The electronic device as described in embodiment 17, wherein the electronic device is configured to implement the network data analysis function NWDAF in a cellular network, and the model library resides at the analysis data repository function ADRF in the cellular network.

[0259] 22. The electronic device as described in embodiment 17, wherein the one or more local distribution parameter sets and the task identifier are received from a server for machine learning.

[0260] 23. The electronic device as described in embodiment 22, wherein receiving the one or more distribution parameter sets and the task identifier includes: directly receiving the one or more distribution parameter sets and the task identifier from the server; or receiving the one or more distribution parameter sets and the task identifier from the server via the network exposure function NEF of the cellular network.

[0261] 24. The electronic device as described in embodiment 17, wherein sending the model list includes: sending the matching model of each of the one or more clients together with the client identifier of the client to the server.

[0262] 25. The electronic device as described in embodiment 17, wherein each local distribution parameter set in the one or more local distribution parameter sets includes at least one of the following: the mean and standard deviation for a Gaussian distribution; the shape parameter and inverse scale parameter for a gamma distribution; or the mean and variance for a lognormal distribution.

[0263] 26. A method, comprising the steps of: obtaining a local distribution parameter set of a second device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the second device; and sending a trained specific model to the second device, wherein the specific model is determined at least in part based on the local distribution parameter set.

[0264] 27. A method includes the following steps: sending a local distribution parameter set to a first device, where the local distribution parameter set characterizes the data distribution characteristics of a local data set of the electronic device; and receiving a trained specific model from the first device, where the specific model is determined at least partially based on the local distribution parameter set.

[0265] 28. A method includes the following steps: receiving one or more local distribution parameter sets, where each local distribution parameter set in the one or more local distribution parameter sets characterizes the data distribution characteristics of a local data set of a corresponding client in one or more clients; sending a task identifier to a model library, where the task identifier identifies a training task associated with the one or more clients; receiving from the model library one or more trained candidate models corresponding to the task identifier; and sending a model list, where the model list includes a matching model selected from the one or more candidate models for each client in the one or more clients, and the matching model is determined at least partially based on the corresponding local distribution parameter set in the one or more local distribution parameter sets.

[0266] 29. A computer-readable storage medium storing one or more instructions, where when the one or more instructions are executed by one or more processing circuits of an electronic device, the electronic device is caused to execute the method according to any one of Embodiments 26-28.

[0267] 30. A computer program product includes a computer program that, when executed by a processor, executes the method according to any one of Embodiments 26-28.

[0268] 31. An apparatus including components for executing the method according to any one of Embodiments 26-28.

Claims

1. An electronic device, comprising: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code which, when executed by the at least one processing unit, causes the electronic device to perform the following operations: obtain a local distribution parameter set of a second device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the second device; and send a trained specific model to the second device, wherein the specific model is determined at least in part based on the local distribution parameter set.

2. The electronic device according to claim 1, wherein, The operations further include: sending the local distribution parameter set together with the client identifier of the second device; and receiving a model list, the model list including the specific model.

3. The electronic device according to claim 1, wherein, The electronic device is configured to implement a server for a federated learning process, and the second device is configured to implement one of a plurality of clients for the federated learning process.

4. The electronic device according to claim 3, wherein, The operations further include: sending the local distribution parameter set together with a task identifier, the task identifier identifying a specific training task associated with the second device in the federated learning process; and receiving the specific model, wherein the specific model is determined at least in part based on the task identifier.

5. The electronic device according to claim 3, wherein, The operations further include: receiving a trained local model from the second device, the local model being trained by the second device based on a complemented local data set, wherein the complemented local data set is generated by the specific model based on at least a portion of the local data set; and generating a global model at least in part based on an aggregation of the local models.

6. The electronic device according to claim 1, wherein, The electronic device is configured to implement an application function AF of a cellular network, and the operations further include: if the AF is not trusted, sending the local distribution parameter set to a network data analytics function NWDAF of the cellular network via a network exposure function NEF of the cellular network; and if the AF is trusted, directly sending the local distribution parameter set to the NWDAF.

7. The electronic device according to claim 1, wherein, The specific model is selected from one or more candidate models, and in one or more training data sets used to train the one or more candidate models, the specific training data set used to train the specific model has the highest similarity to the local data set of the second device.

8. The electronic device according to claim 7, wherein, The similarity is determined at least in part based on: (1) the local distribution parameter set of the local data set, and (2) a specific distribution parameter set of the specific training data set.

9. An electronic device, comprising: at least one processing unit; and at least one storage unit, the at least one storage unit including computer program code, wherein the computer program code, when executed by the at least one processing unit, causes the electronic device to perform the following operations: send a local distribution parameter set to a first device, the local distribution parameter set characterizing the data distribution characteristics of the local data set of the electronic device; and Receive a trained specific model from a first device, where the specific model is determined at least in part based on the set of local distribution parameters.

10. The electronic device according to claim 9, wherein, The operations further include: Using the received specific model, generate an output data portion based on at least a part of the local data set.

Images