Electronic device and method for federated learning

Abstract

An electronic device and method for federated learning are provided.SOLUTION: The federated learning process 300 in the central processor S305 determines a set of distributed nodes of the plurality of distributed nodes for generating global model parameters. The correlation between the local training data of the group of distributed nodes meets a specific correlation requirement to obtain a channel resource (for example, a time resource, a frequency resource, and / or a space resource) for sending the local model parameters to the central processing unit. A S310 that generates global model parameters based on the local model parameters of the set of distributed nodes. The local model parameters are generated by the set of distributed nodes based on respective local training data. In the S315, the CPU distributes the generated global model parameters to the plurality of distributed nodes.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] [Priority claim] This application claims priority to a Chinese patent application filed on October 28, 2020, bearing application number 202011173054.X and entitled "Electronic Device and Method Used in Federated Learning," the entire contents of which are incorporated herein by reference.

[0002] The present disclosure relates to electronic devices and methods for use in associative learning. [Background technology]

[0003] Federated learning is an emerging fundamental technology for artificial intelligence. First proposed by Google in 2016, it was originally used to solve the problem of local model updates for Android mobile phone users. Its design goal is to perform efficient machine learning across multiple participants or computing nodes while ensuring information security during big data exchange, protecting the privacy of device and personal data, and ensuring legality and compliance. Machine learning algorithms that can be used in federated learning are not limited to neural networks but also include important algorithms such as random forests. Federated learning is expected to become the foundation for next-generation AI collaborative algorithms and collaborative networks.

[0004] FIG. 1 is a schematic diagram showing a federated learning system in the prior art. As shown in FIG. 1, UE1, UE2, . . . , UE K are distributed nodes that perform federated learning and are implemented by user equipment. AP is a central processing unit that performs federated learning and is implemented by a base station. Each distributed node UE1, UE2, ..., UE K is its own local training data D1, D2, . . . , D K Based on this, an artificial intelligence algorithm (e.g., gradient descent) is used to calculate the local model parameters w t 1. Wt 2,...,w t K The central processing unit AP generates the local model parameters w t 1. W t 2,...,w t K Based on the global model parameters w t+1 (i.e., global aggregation) and each distributed node UE1, UE2,..., UE K to be distributed.

[0005] In this way, the central processing unit AP can train a statistical model from data stored in a large number of distributed node UEs without sacrificing the privacy of each individual distributed node UE. However, federated learning systems face new challenges. Unlike centralized learning in a data center, federated learning is typically performed in a wireless edge network, where channel resources are limited and unreliable. The central processing unit AP typically needs to link a large number of distributed node UEs over a resource-limited spectrum. Therefore, for each global aggregation, only a limited number of distributed node UEs are allowed to transmit their local model parameters through an unreliable channel. Summary of the Invention

[0006] The inventors of the present disclosure have found that when a predetermined number of distributed nodes must be selected to upload local model parameters, if the local training data of each selected distributed node is highly correlated, the models learned by these distributed nodes will be similar, making it impossible to ensure rapid convergence of the global model. Therefore, a technical solution is needed that can improve the efficiency of a federated learning system when channel resources are limited.

[0007] To solve this technical problem, the technical solution proposed in this disclosure schedules the distributed nodes to upload their local model parameters according to the correlation between the local training data of the distributed nodes, thereby promoting the rapid convergence of the global model. The technical solution of the present disclosure can effectively utilize limited channel resources to improve the efficiency of the federated learning system.

[0008] According to one aspect of the present disclosure, there is provided an electronic device for use in federated learning in a central processing unit, the electronic device including a processing circuit, the processing circuit being configured to determine a set of distributed nodes from a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specific correlation requirement, and to generate the global model parameters based on the local model parameters of the set of distributed nodes, wherein the local model parameters are generated based on the local training data of each of the set of distributed nodes.

[0009] According to another aspect of the present disclosure, there is provided an electronic device for use in federated learning at a particular distributed node, the electronic device including a processing circuit configured to determine that the particular distributed node will be used to generate global model parameters, wherein correlations between local training data of the particular distributed node and local training data of other distributed nodes for generating global model parameters satisfy specific correlation requirements, and upload the local model parameters of the particular distributed node to a central processing unit, wherein the local model parameters of the particular distributed node are generated based on its local training data.

[0010] According to another aspect of the present disclosure, there is provided a method for federated learning in a central processing unit, the method including: determining a set of distributed nodes from a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specific correlation requirement; and generating global model parameters based on the local model parameters of the set of distributed nodes, wherein the local model parameters are generated based on the local training data of each of the set of distributed nodes.

[0011] According to another aspect of the present disclosure, there is provided a method for federated learning at a particular distributed node, the method including: determining that the particular distributed node will be used to generate global model parameters, wherein correlations between local training data of the particular distributed node and local training data of other distributed nodes determined to be used to generate the global model parameters meet specific correlation requirements; and uploading the local model parameters of the particular distributed node to a central processing unit, wherein the local model parameters of the particular distributed node are generated based on its local training data.

[0012] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium is provided having stored thereon instructions that, when executed by a processor, cause the processor to perform the method of the present disclosure.

[0013] According to another aspect of the present disclosure, there is provided an apparatus for use in associative learning, including components for performing each step of the method of the present disclosure. [Brief explanation of the drawings]

[0014] A better understanding of the present disclosure can be obtained by considering the following specific description of the embodiments in conjunction with the drawings, in which the same or similar parts are designated by the same or similar reference numerals in the various drawings, and the drawings, together with the following specific description, which are incorporated into and constitute a part of this specification, illustrate, by way of example, embodiments of the present disclosure and serve to illustrate the principles and advantages of the present disclosure.

[0015] [Figure 1] FIG. 1 is a schematic diagram showing a federated learning system in the prior art. [Figure 2] FIG. 2 is a schematic diagram illustrating a federated learning system according to some embodiments of the present disclosure. [Figure 3] FIG. 3 is a flow chart illustrating a federated learning process 300 in a central processing unit according to some embodiments of the present disclosure. [Figure 4] FIG. 4 is a schematic diagram illustrating a distribution of distributed nodes implemented as camera nodes according to some embodiments of the present disclosure. [Figure 5] FIG. 5 is a flow chart illustrating a federated learning process 500 in a distributed node according to some embodiments of the present disclosure. [Figure 6] FIG. 6 is a flowchart illustrating a process for centrally determining distributed nodes for generating global model parameters according to some embodiments of the present disclosure. [Figure 7] FIG. 7 is a schematic diagram illustrating selecting distributed nodes for generating global model parameters by a central processing unit according to some embodiments of the present disclosure. [Figure 8] FIG. 8 is a flowchart illustrating a process for distributedly determining distributed nodes for generating global model parameters according to some embodiments of the present disclosure. [Figure 9] FIG. 9 is a schematic diagram illustrating the correlation of a distributed node with its neighboring nodes' local training data. [Figure 10] Figure 10 is a schematic diagram showing the classification accuracy of the federated learning model obtained using different node selection policies. [Figure 11] FIG. 11 is a block diagram showing an example of a schematic layout of a computing device to which the techniques of the present disclosure can be applied. [Figure 12] Figure 12 is a block diagram showing a first example of a schematic arrangement of a gNB to which the technology of the present disclosure can be applied. [Figure 13] Figure 13 is a block diagram showing a second example of a schematic arrangement of gNBs to which the technology of the present disclosure can be applied. [Figure 14] FIG. 14 is a block diagram illustrating an example of an exemplary smartphone layout to which the techniques of this disclosure can be applied. [Figure 15] FIG. 15 is a block diagram showing an example of a schematic layout of a car navigation device to which the technology of the present disclosure can be applied. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the drawings. In this specification and the drawings, components having substantially the same functions and structures are designated by the same reference numerals, and descriptions of these structural elements will not be repeated.

[0017] The explanation will be given in the following order. 1. System Overview 2. Processing flow 3. Simulation Results 4. Application Examples

[0018] <1. System Overview> First, a federated learning system according to some embodiments of the present disclosure will be described. Fig. 2 is a schematic diagram illustrating a federated learning system 200 according to some embodiments of the present disclosure. As shown in Fig. 2, the federated learning system 200 includes multiple distributed nodes UE1, UE2, ..., UE3, which are implemented by user equipment (UEs). K and a central processing unit AP realized by a base station. Each of the distributed nodes UE1, UE2, . . . , UE K is its own local training data D1, D2, . . . , D KBased on this, an artificial intelligence algorithm (e.g., gradient descent) is used to calculate the local model parameters w t 1. W t 2,...,w t K Generate.

[0019] Unlike the federated learning system in FIG. 1, the federated learning system 200 in FIG. 2 includes multiple distributed nodes UE1, UE2, . . . , UE K It is determined that only one set of distributed nodes among the distributed nodes UE1 and UE2 will be used to generate the global model parameters. The distributed nodes in this set upload their local model parameters to the central processing unit AP, while the distributed nodes not used to generate the global model parameters (e.g., distributed node UE2) do not upload their local model parameters to the central processing unit AP. The central processing unit AP determines the global model parameters w based only on the local model parameters of the distributed nodes in this set. t+1 Generate.

[0020] Similar to the federated learning system in FIG. 1, the central processing unit AP in the federated learning system 200 in FIG. 2 calculates the generated global model parameters w t+1 , UE1, UE2, . . . , UE, including the distributed node that does not upload local model parameters (e.g., distributed node UE2). K Distributed nodes UE1, UE2,..., UE K is the newly generated global model parameter w t+1 When it receives the newly generated global model parameters w t+1 , and then use it to train the next round of local models, or use the newly generated global model parameters w t+1 We apply a federated learning model based on

[0021] In the embodiment shown in FIG. 2, the distributed nodes are implemented by user equipment, and the central processor is implemented by a base station, with wireless communication between them. In some embodiments, the central processor may be implemented by a device separate from the base station. In this case, local model parameters are uploaded by the distributed nodes to the base station and provided by the base station to the central processor, and generated global model parameters are provided by the central processor to the base station and distributed by the base station to the distributed nodes. The communication system consisting of the distributed nodes and the central processor may be a cellular communication system, a machine type communication (MTC), a self-organizing network, or a cognitive radio system (e.g., IEEE P802.19.1a, Spectrum Access System (SAS)), etc.

[0022] Furthermore, in some embodiments of the present disclosure, the distributed nodes and the central processing unit may be implemented by other devices, and wired communication may be performed between them.

[0023] <2. Processing flow> The flow of federated learning processing in the distributed nodes and the central processing unit will be described below. In an embodiment of the present disclosure, the federated learning processing in the distributed nodes and / or the central processing unit may be executed by electronic equipment implemented as the distributed nodes and / or the central processing unit themselves or as components thereof (e.g., chips). The electronic equipment may include a processing circuit. The processing circuit may output signals (digital or analog) to other components in the distributed nodes and / or the central processing unit, or may receive signals (digital or analog) from other components in the distributed nodes and / or the central processing unit. The processing circuit may also control the operation of some or all of the other components in the distributed nodes and / or the central processing unit.

[0024] The processing circuit may be in the form of a general-purpose processor or may be a dedicated processing circuit such as an ASIC. For example, the processing circuit may be constituted by a circuit (hardware) or a central processing unit (e.g., a central processing unit (CPU)). The processing circuit may also be loaded with a program (software) for operating the circuit (hardware) or the central processing unit. The program may be stored in a memory (e.g., located in the distributed node and / or the central processing unit or electronic device), or in an external storage medium connected from the outside, or may be downloaded via a network (e.g., the Internet).

[0025] FIG. 3 is a flow chart illustrating a federated learning process 300 in a central processing unit according to some embodiments of the present disclosure.

[0026] In step S305, the central processing unit determines a set of distributed nodes among the plurality of distributed nodes for generating global model parameters, where the correlations between the local training data of the set of distributed nodes meet certain correlation requirements and can obtain channel resources (e.g., time resources, frequency resources, and / or space resources) for transmitting the local model parameters to the central processing unit.

[0027] There may be correlation between the local training data of multiple distributed nodes. High correlation will result in similar local training results of multiple distributed nodes, making it impossible to ensure that the global training results quickly approach global optimization when channel resources are limited. By selecting distributed nodes with low correlation between their local training data and allowing the distributed nodes with low correlation to obtain channel resources for uploading their local training results (i.e., local model parameters), it is possible to effectively utilize channel resources when channel resources are limited, thereby facilitating the global training results quickly approach global optimization. The correlation between the local training data of distributed nodes can be determined by some of the methods listed below or other methods.

[0028] In some application scenarios, the local training data of a distributed node is derived from sampling of the surrounding environment. For example, a distributed node may be implemented as a sensor node (e.g., a temperature collection node) whose local training data is derived from a sensor installed therein, or as a camera node whose local training data is derived from data collected by a camera installed therein. Some distributed nodes that are close in distance have high correlation of their local training data if the collection times of the local training data are close. Therefore, the correlation between the local training data of multiple distributed nodes may be determined based on the collection times of the local training data. To ensure that the correlation between the local training data of distributed nodes for generating global model parameters is low, a specific correlation requirement may be set such that the time interval between the collection times of the local training data of distributed nodes for generating global model parameters is greater than a predetermined time interval threshold. Also, distributed nodes may be ordered according to the time interval between the collection times of their local training data. A distributed node with a larger time interval has a lower correlation and is therefore preferentially selected as a distributed node for generating global model parameters and uploading its local model parameters. Also, if the data collection time is a time period, to ensure low correlation between the local training data of the distributed nodes for generating the global model parameters, a specific correlation requirement may be set such that the overlap time period between the collection times of the local training data of the distributed nodes for generating the global model parameters is less than a predetermined overlap time threshold.

[0029] Furthermore, when the distance between distributed nodes is short, the correlation of their local training data is high. Therefore, the correlation between the local training data of multiple distributed nodes may be determined based on the distance between the distributed nodes. To ensure low correlation between the local training data of distributed nodes for generating global model parameters, a specific correlation requirement may be set such that the distance between the node locations of distributed nodes for generating global model parameters is greater than a predetermined distance threshold. Also, distributed nodes may be ordered according to the distance between their node locations. Nodes with a greater distance have lower correlation and are therefore preferentially selected as distributed nodes for generating global model parameters and uploading local model parameters. This ensures low correlation between the local training data of distributed nodes selected for generating global model parameters.

[0030] An associated exclusion region may be established for each distributed node such that a set of distributed nodes selected to generate global model parameters are outside of each other's exclusion regions, thereby ensuring low correlation between the local training data of the set of distributed nodes. A specific correlation requirement may be set such that each distributed node among the distributed nodes for generating global model parameters is outside the associated exclusion regions of the remaining distributed nodes among the distributed nodes for generating global model parameters. Once a distributed node is selected as the distributed node for generating global model parameters, distributed nodes within its associated exclusion region are no longer selected as distributed nodes for generating global model parameters.

[0031] The associated exclusion region of each distributed node may be defined by a region in which the distance between each point of the region and the distributed node is smaller than the exclusion distance. For example, the exclusion region of each distributed node may be a circular region centered on the distributed node and having a radius equal to the exclusion distance. Alternatively, the associated exclusion region of each distributed node may be defined by a set of a second predetermined number of distributed nodes that are closest to the distributed node. For example, the exclusion region of each distributed node may be a set of a second predetermined number of closest neighboring nodes that are found by distance from the distributed node, centered on the distributed node. The exclusion distance and the second predetermined number may be predetermined based on one or more of the distribution density of the distributed nodes, the distribution density of the observables, and the maximum correlation margin.

[0032] Furthermore, the relationship between the correlation and distance between the local training data of the distributed nodes may be determined based on the distributed statistical information of the distributed nodes. After determining the relationship between the correlation and distance, the correlation between the training data of the distributed nodes may be determined based on the distance between the distributed nodes. For example, assuming that the node positions of the distributed nodes follow a Poisson point process, the physical locations of the collection targets follow another Poisson point process, data is collected by cameras on the distributed nodes, and the coverage range of the cameras is limited, a function between the correlation between the training data of the distributed nodes and the distance between the distributed nodes can be determined by simulation or theoretical calculation. Furthermore, the relationship between the correlation between the local training data and the distance between the distributed nodes may be provided to the central processing unit or the distributed nodes by a third party.

[0033] When the distributed nodes are implemented as camera nodes, the overlap area of ​​the camera coverage areas may be used based on the distribution positions and angles of the cameras to calculate the relationship between the correlation and distance of the training data collected by the cameras at a given camera density. Figure 4 is a schematic diagram showing the distribution of distributed nodes implemented as camera nodes according to some embodiments of the present disclosure. In Figure 4, dots indicate distributed nodes implemented as camera nodes, and triangular regions indicate the coverage areas (i.e., imaging ranges) of the camera nodes.

[0034] In some application scenarios, the local training data of the distributed nodes are generated by a predetermined generative model. Because the generative models are the same, the generated training data have high correlation. Therefore, the correlation between the local training data of multiple distributed nodes may be determined based on the generative models of the local training data. To ensure that the correlation between the local training data of the distributed nodes for generating the global model parameters is low, a specific correlation requirement may be set so that the generative models of the local training data of the distributed nodes for generating the global model parameters are different.

[0035] In some application scenarios, a distributed node uses an open source platform as a data source, i.e., its local training data originates from the open source platform. Training data originating from the same open source platform have a high correlation. Therefore, the correlation between the local training data of multiple distributed nodes may be determined based on the data source of the local training data. To ensure that the correlation between the local training data of the distributed nodes for generating the global model parameters is low, a specific correlation requirement may be set for different data sources of the local training data of the distributed nodes for generating the global model parameters.

[0036] 3, in step S310, the central processing unit generates global model parameters based on the local model parameters of the set of distributed nodes. The local model parameters are generated by the set of distributed nodes based on their respective local training data. In step S315, the central processing unit distributes the generated global model parameters to the plurality of distributed nodes.

[0037] FIG. 5 is a flowchart illustrating a federated learning process 500 in a distributed node according to some embodiments of the present disclosure. In step S505, a distributed node determines that it will be used to generate global model parameters. Correlations between the distributed node's local training data and the local training data of other distributed nodes to generate the global model parameters satisfy specific correlation requirements, and channel resources (e.g., time resources, frequency resources, and / or space resources) for transmitting the local model parameters to a central processing unit can be obtained. The specific correlation requirements may be the correlation requirements described above with reference to FIG. 3. In step S510, the distributed node transmits its local model parameters to the central processing unit through the obtained channel resources. The local model parameters are generated by the distributed node based on its local training data. In step S515, the distributed node receives the global model parameters from the central processing unit.

[0038] In some embodiments of the present disclosure, determining a distributed node for generating the global model parameters at the central processing unit and the distributed nodes may be implemented centrally based on a selection for the distributed node at the central processing unit. Alternatively, in some embodiments of the present disclosure, determining a distributed node for generating the global model parameters at the central processing unit and the distributed nodes may be implemented distributedly based on contention for channel resources at the distributed nodes.

[0039] 6 is a flowchart illustrating a process 600 for centrally determining distributed nodes for generating global model parameters according to some embodiments of the present disclosure. Process 600 may be implemented by a Radio Resource Control (RRC) process. For simplicity, only two distributed nodes, UE1 and UE2, are shown in FIG. 6. However, it should be understood that process 600 may involve a greater number of distributed nodes.

[0040] In step S605, the distributed node UE1 sends information about the distributed node UE1 to the central processing unit AP, for example in an uplink resource request, so that the central processing unit selects a distributed node for generating global model parameters. The information about the distributed node UE1 may include one or more of data collection information of the distributed node UE1's local training data, network performance parameters, and local training performance parameters.

[0041] The data collection information of the local training data of the distributed node UE1 includes one or more of a collection time of the local training data, a node location, a generative model, and a data source, and can be used to determine a correlation between the local training data of the distributed node UE1 and the local training data of other distributed nodes.

[0042] The network performance parameters of the distributed node UE1 indicate the network performance of the distributed node UE1, and may include, for example, one or more of the channel bandwidth, the degree of interference, and the channel quality of the distributed node UE1.

[0043] The local training performance parameter of the distributed node UE1 indicates the performance of the distributed node UE1 to locally perform model training, and may include, for example, the magnitude of the update gradient or the level of the update gradient of the objective function of the model training.

[0044] In step S610, the distributed node UE2 sends information about the distributed node UE2 to the central processing unit AP, for example in an uplink resource request, so that the central processing unit selects a distributed node for generating global model parameters. The information about the distributed node UE2 may include one or more of data collection information of the distributed node UE2's local training data, network performance parameters, and local training performance parameters.

[0045] The data collection information of the local training data of the distributed node UE2 includes one or more of a collection time of the local training data, a node location, a generative model, and a data source. The data collection information of the local training data of the distributed node UE2 can be used to determine a correlation between the local training data of the distributed node UE2 and the local training data of other distributed nodes.

[0046] The network performance parameters of the distributed node UE2 indicate the network performance of the distributed node UE2 and may include, for example, one or more of the channel bandwidth, the degree of interference, and the channel quality of the distributed node UE2.

[0047] The local training performance parameter of the distributed node UE2 indicates the performance of the distributed node UE2 to locally perform model training, and may include, for example, the magnitude of the update gradient or the level of the update gradient of the objective function of the model training.

[0048] In step S615, the central processing unit AP selects a set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on information about the distributed nodes from the plurality of distributed nodes, such that correlations between local training data of the selected set of distributed nodes meet a specific correlation requirement. The number of selected distributed nodes may be a first predetermined number. The central processing unit AP can allocate channel resources (e.g., time resources, frequency resources, and / or space resources) to the selected set of distributed nodes for transmitting the local model parameters.

[0049] The central processing unit AP may also select a set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on the network performance parameters and / or local training performance parameters of the distributed nodes, so that the network performance parameters and / or local training performance parameters of the selected set of distributed nodes meet specific network performance requirements and / or local training performance requirements. By selecting distributed nodes for generating global model parameters based on the network performance parameters and / or local training performance parameters of the distributed nodes, the central processing unit can select distributed nodes with good network performance and / or high local training performance to generate global model parameters, thereby improving the efficiency of global model training. For example, some distributed nodes may be unable to upload local model parameters in a timely manner due to poor network performance or unable to generate local model parameters in a timely manner due to poor local training performance, which affects the efficiency of global model training. For example, a distributed node that meets a specific correlation requirement may be located at a cell edge and therefore have poor channel quality. Therefore, when selecting a distributed node for generating global model parameters, the central processing unit AP may take into account the maximum coverage range of the base station to prevent a distributed node located outside the maximum coverage range of the base station from being selected.

[0050] 7 is a schematic diagram illustrating the selection of distributed nodes for generating global model parameters by a central processing unit according to some embodiments of the present disclosure. In FIG. 7, black circles indicate distributed nodes selected for generating global model parameters, white circles indicate distributed nodes not selected for generating global model parameters, dashed circles indicate exclusion regions, and R indicates the radius of the exclusion region.

[0051] As shown in FIG. 7 , the central processing unit selects K distributed nodes for generating global model parameters over K cycles, selecting one distributed node per cycle. After selecting k distributed nodes (k is a non-negative integer less than K), in the k+1 cycle, the central processing unit selects the k+1 distributed nodes from regions other than the associated exclusive regions of the selected k distributed nodes. In some embodiments, the central processing unit further considers the network performance and / or local training performance of the distributed nodes when selecting the distributed nodes. In this case, in the k+1 cycle, the central processing unit selects the k+1 distributed nodes from regions other than the associated exclusive regions of the selected k distributed nodes whose network performance parameters and / or local training performance parameters satisfy specific network performance requirements and / or local training performance requirements.

[0052] Returning to FIG. 6, in step S620, the central processing unit AP sends an instruction to upload the local model parameters to the distributed node selected to generate the global model parameters. The central processing unit AP may notify the distributed node selected to generate the global model parameters of the allocated channel resources. For example, the instruction to send the local model parameters may include an indication of the allocated channel resources. For example, in FIG. 6, the central processing unit AP sends an instruction to upload the local model parameters to the distributed node UE1.

[0053] In step S625, the distributed node selected to generate the global model parameters determines that it will be used to generate the global model parameters based on the instruction to upload the local model parameters received from the central processing unit. For example, in Figure 6, the distributed node UE1 determines that it will be used to generate the global model parameters based on the instruction to upload the local model parameters received from the central processing unit.

[0054] 8 is a flowchart illustrating a process 880 for distributedly determining distributed nodes for generating global model parameters according to some embodiments of the present disclosure. Process 880 may be implemented by a Radio Resource Control (RRC) process. For simplicity, only one distributed node, UE1, is shown in FIG. 8. However, it should be understood that process 880 may involve a greater number of distributed nodes.

[0055] In step S882, the distributed node UE1 determines, based on its local training data and the data collection information of the local training data of the other distributed nodes determined to be used in generating the global model parameters, whether the correlation between its local training data and the local training data of the other distributed nodes determined to be used in generating the global model parameters satisfies a specific correlation requirement. The data collection information of the local training data of the distributed nodes determined to be used in generating the global model parameters may be broadcast by the distributed node itself, or may be sent to the central processing unit AP and broadcast by the central processing unit AP.

[0056] In some embodiments, the data collection information may include a collection time of the local training data, and the specific correlation requirement may be set so that the time interval between the collection time of the local training data of the distributed node UE1 and the local training data of other distributed nodes determined to be used in generating the global model parameters is greater than a predetermined time interval threshold.

[0057] In some embodiments, the data collection information may include the node location of the distributed node, and the specific correlation requirement may be set such that the distance between the distributed node UE1 and the other distributed nodes determined to be used in generating the global model parameters is greater than a predetermined distance threshold. The specific correlation requirement may also be set such that the distributed node UE1 is located outside the associated exclusion regions of the other distributed nodes determined to be used in generating the global model parameters. This specific distance threshold or exclusion region may be preset, broadcast by the distributed nodes or central processing unit AP determined to be used in generating the global model parameters, or determined by the distributed node UE1 based on current network conditions.

[0058] If it is determined in step S882 that the specific correlation condition is met, the process 880 proceeds to step S884, in which the distributed node UE1 competitively acquires channel resources for transmitting the local model parameters to the central processing unit AP, for example, by carrier sensing, and determines whether the acquisition of the channel resources is successful.

[0059] In some embodiments, the distributed node UE1 may adjust parameters for acquiring channel resources (e.g., energy detection threshold, back-off window size) based on one or both of its network performance parameters and local training performance parameters. For example, if the network performance and / or local training performance is good (e.g., higher than a corresponding performance threshold), the distributed node UE1 may increase the probability of successfully acquiring channel resources by adjusting parameters for acquiring channel resources (e.g., increasing its energy detection threshold or decreasing the back-off window size). Thus, a distributed node with good network performance and / or good local training performance may be more likely to upload higher local model parameters.

[0060] In some embodiments, the central processing unit AP broadcasts an instruction to stop uploading after a first predetermined number of distributed nodes determined to be used in generating the global model parameters have been reached, and upon receiving the instruction to stop uploading from the central processing unit AP, the distributed node UE1 stops acquiring channel resources for uploading local model parameters to the central processing unit.

[0061] In some embodiments, the distributed node UE1 may itself determine the number of distributed nodes determined to be used in generating the global model parameters. For example, the distributed node UE1 may increment this number by one each time it receives a broadcast from a distributed node determined to be used in generating the global model parameters. If this number is equal to or greater than a first predetermined number, the distributed node UE1 may stop acquiring channel resources for transmitting local model parameters to the central processing unit AP. The first predetermined number may be preset in the distributed node UE1 or may be broadcast by the central processing unit AP.

[0062] If the acquisition of the channel resources is successful in step S884, the process 880 proceeds to step S886. In step S886, the distributed node UE1 determines that it will be used to generate global model parameters. In step S888, the distributed node UE1 transmits its local model parameters to the central processing unit AP through the acquired channel resources. In step S890, in response to receiving the local model parameters of the distributed node UE1, the central processing unit AP determines the distributed node UE1 as the distributed node for generating the global model parameters.

[0063] Although not shown in Fig. 8, after determining that the distributed node UE1 will be used to generate the global model parameters, the distributed node UE1 may broadcast the data collection information of its local training data, or may transmit the data collection information to the central processing unit AP and have the central processing unit AP broadcast it, thereby determining the distributed node for generating the next global model parameters in a distributed manner.

[0064] The technical solution of the embodiments of the present disclosure may be applied to an intelligent driving scenario. In the intelligent driving scenario, a vehicle uses an artificial intelligence model to determine whether pedestrians are crossing the street at an intersection and the status of the traffic lights, and then determines a braking policy. Each vehicle's local training data includes intersection situation data collected by a vehicle camera and a braking label provided by a human (e.g., a driver or an intersection observer) or a monitoring processing device. The braking label indicates whether braking is necessary in the current intersection situation and may be provided by a human observing the intersection situation, or may be obtained by a monitoring processing device inputting intersection situation data captured by a camera at the traffic lights into a trained artificial intelligence model. During the local training process, each vehicle may use an algorithm to update its local model using the inverse of the difference between the braking decision provided by the local model based on the intersection situation data in some intersection situations and the braking label provided by a human or monitoring processing device as an objective function.

[0065] In some embodiments of the present disclosure, each vehicle, as a distributed node, can upload its local training data to a central processing unit to perform federated learning. Because there may be correlation between the local training data of different vehicles at the same intersection, a vehicle with low correlation can be selected based on the collection time of the vehicle's local training data and / or the vehicle's location, and upload its local model parameters to generate global model parameters. Furthermore, a vehicle with good network performance and / or good local training performance can be selected based on the vehicle's network performance parameters and / or local training performance parameters, and upload its local model parameters to generate global model parameters. The central processing unit distributes the generated global model parameters to each vehicle. Each vehicle processes intersection situation data collected by its vehicle camera according to a model based on the global model parameters to determine whether braking is required.

[0066] <3. Simulation Results> When the local training data of distributed nodes comes from sampling of the surrounding environment, the correlation between the local training data of multiple distributed nodes is related to the distance between them: the greater the distance between distributed nodes, the lower the correlation between their local training data.

[0067] FIG. 9 is a schematic diagram illustrating the correlation of local training data of a distributed node and its neighboring nodes. In FIG. 9, the abscissa represents the correlation coefficient between neighboring nodes, and the ordinate represents the cumulative distribution function of the correlation between the local training data of a given distributed node and its neighboring nodes. Curve 991 represents the correlation between a given distributed node and its nearest neighboring node, curve 992 represents the correlation between a given distributed node and its second-nearest neighboring node, curve 993 represents the correlation between a given distributed node and its third-nearest neighboring node, curve 994 represents the correlation between a given distributed node and its fourth-nearest neighboring node, and curve 995 represents the correlation between a given distributed node and its fifth-nearest neighboring node. As shown in FIG. 9, as the distance between neighboring nodes increases, the correlation decreases.

[0068] FIG. 10 is a schematic diagram showing the classification accuracy of federated learning models obtained using different node selection policies. In FIG. 10, the abscissa represents the number of iterations of the federated learning model, and the ordinate represents the classification accuracy of the generated federated learning model. Curve 1001 represents the classification accuracy of the federated learning model when generated using local training data from a single randomly selected node. Curve 1002 represents the classification accuracy of the federated learning model when generated using local training data from a first predetermined number of randomly selected nodes. Curve 1003 represents the classification accuracy of the federated learning model when generated using the technology disclosed herein by selecting local training data from the first predetermined number of nodes. Curve 1004 represents the classification accuracy of the federated learning model when generated using local training data from all nodes. As shown in FIG. 10, the classification accuracy of the federated learning model when generated using the technology disclosed herein by selecting local training data from the first predetermined number of nodes is close to the classification accuracy of the federated learning model when generated using local training data from all nodes. Therefore, the technique of the present disclosure can efficiently generate a federated learning model without significantly reducing the classification accuracy of the federated learning model.

[0069] <4. Application Examples> The techniques of this disclosure may be applied to a variety of products, for example, base stations and user equipment may be implemented as various types of computing devices.

[0070] A base station may also be realized as any type of evolved Node B (eNB), gNB, or TRP (Transmit Receive Point), such as a macro eNB / gNB or a small eNB / gNB. A small eNB / gNB is an eNB / gNB that covers a cell smaller than a macro cell, such as a pico eNB / gNB, a micro eNB / gNB, or a home (femto) eNB / gNB. Alternatively, a base station may be realized as any other type of base station, such as a Node B or a base transceiver station (BTS). A base station may include a main body (also referred to as a base station device) configured to control wireless communications and one or more remote radio heads (RRHs) located at different locations from the main body. Various types of terminals, described below, can operate as a base station by temporarily or semi-permanently performing base station functions.

[0071] The user equipment may also be realized as a mobile terminal (e.g., a smartphone, a tablet personal computer (PC), a notebook PC, a portable game console, a portable / dongle mobile router, and a digital imaging device) or an in-vehicle terminal (e.g., a car navigation device). The user equipment may also be realized as a terminal that performs machine-to-machine (M2M) communication (also called a machine-type communication (MTC) terminal). The user equipment may also be a wireless communication module (e.g., an integrated circuit module including a single chip) mounted on each of the above terminals.

[0072] [3-1. Application examples for computing devices] 11 is a block diagram showing an example of a schematic arrangement of a computing device 700 to which the technology of this disclosure can be applied. The computing device 700 includes a processor 701, a memory 702, a storage device 703, a network interface 704, and a bus 706.

[0073] The processor 701 may be, for example, a central processing unit (CPU) or a digital signal processor (DSP), and controls the functions of the server 700. The memory 702 includes a random access memory (RAM) and a read-only memory (ROM), and stores data and programs executed by the processor 701. The storage device 703 may include, for example, a storage medium such as a semiconductor memory and a hard disk.

[0074] The network interface 704 is a wired communication interface for connecting the server 700 to a wired communication network 705. The wired communication network 705 may be a core network such as an evolved packet core (EPC) or a packet data network (PDN) such as the Internet.

[0075] The bus 706 interconnects the processor 701, the memory 702, the storage device 703, and the network interface 704. The bus 706 may include two or more buses having different speeds (e.g., a high-speed bus and a low-speed bus).

[0076] [3-2. Application examples for base stations] (First application example) 12 is a block diagram showing a first example of a schematic arrangement of a gNB to which the technology of the present disclosure can be applied. The gNB 800 includes one or more antennas 810 and base station equipment 820. The base station equipment 820 and each antenna 810 can be connected to each other via an RF cable.

[0077] Each of the antennas 810 includes one or more antenna elements (e.g., multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna) and is used by the base station equipment 820 to transmit and receive radio signals. As shown in Figure 12, the gNB 800 may include multiple antennas 810. For example, the multiple antennas 810 may be compatible with multiple frequency bands used by the gNB 800. Although Figure 12 shows an example in which the gNB 800 includes multiple antennas 810, the gNB 800 may also include a single antenna 810.

[0078] The base station equipment 820 includes a controller 821 , a memory 822 , a network interface 823 and a wireless communication interface 825 .

[0079] The controller 821 may be, for example, a CPU or DSP, and may operate various functions of higher layers of the base station equipment 820. For example, the controller 821 may generate data packets based on data in signals processed by the wireless communication interface 825 and transmit the generated packets via the network interface 823. The controller 821 may bundle data from multiple baseband processors to generate bundled packets and transmit the generated bundled packets. The controller 821 may have logic functions to perform the following controls, such as radio resource control, radio bearer control, mobility management, admission control, and scheduling. This control may be performed in conjunction with nearby gNBs or core network nodes. The memory 822 includes RAM and ROM, and stores programs executed by the controller 821 and various types of control data (e.g., terminal lists, transmission power data, and scheduling data).

[0080] The network interface 823 is a communication interface for connecting the base station equipment 820 to the core network 824. The controller 821 can communicate with a core network node or another gNB via the network interface 823. In this case, the gNB 800 and the core network node or another gNB can be connected to each other by a logical interface (e.g., an S1 interface and an X2 interface). The network interface 823 may be a wired communication interface or a wireless communication interface used for a wireless backhaul line. If the network interface 823 is a wireless communication interface, the network interface 823 can use a higher frequency band for wireless communication compared to the frequency band used by the wireless communication interface 825.

[0081] The wireless communication interface 825 supports any cellular communication method (e.g., Long Term Evolution (LTE) and LTE-Advanced) and provides wireless connectivity to terminals in a cell located in the gNB 800 via the antenna 810. The wireless communication interface 825 typically includes, for example, a baseband (BB) processor 826 and an RF circuit 827. The BB processor 826 can perform, for example, encoding / decoding, modulation / demodulation, multiplexing / demultiplexing, and various types of signal processing (e.g., L1, media access control (MAC), radio link control (RLC), and packet data aggregation protocol (PDCP)). Instead of the controller 821, the BB processor 826 may have some or all of the above-mentioned logical functions. The BB processor 826 may be a memory in which a communication control program is stored, or may be a module including a processor and related circuits configured to execute the program. Program updates can change the functionality of the BB processor 826. This module may be a card or board inserted into a slot in the base station equipment 820. Alternatively, this module may be a chip mounted on a card or board. At the same time, RF circuitry 827 may include, for example, mixers, filters, and amplifiers to transmit and receive radio signals via antenna 810.

[0082] As shown in Figure 12, the wireless communication interface 825 may include multiple BB processors 826. For example, the multiple BB processors 826 may be compatible with multiple frequency bands used by the gNB 800. As shown in Figure 12, the wireless communication interface 825 may include multiple RF circuits 827. For example, the multiple RF circuits 827 may be compatible with multiple antenna elements. Although Figure 12 shows an example in which the wireless communication interface 825 includes multiple BB processors 826 and multiple RF circuits 827, the wireless communication interface 825 may include a single BB processor 826 or a single RF circuit 827.

[0083] (Second application example) 13 is a block diagram showing a second example of a schematic arrangement of a gNB to which the technology of the present disclosure can be applied. The gNB 830 includes one or more antennas 840, a base station device 850, and an RRH 860. The RRH 860 and each antenna 840 can be connected to each other via an RF cable. The base station device 850 and the RRH 860 can be connected to each other via a high-speed line such as an optical fiber cable.

[0084] Each of the antennas 840 includes a single or multiple antenna elements (e.g., multiple antenna elements included in a MIMO antenna) and is used for transmitting and receiving radio signals for the RRH 860. As shown in FIG. 13, the gNB 830 may include multiple antennas 840. For example, the multiple antennas 840 may be compatible with multiple frequency bands used by the gNB 830. Although FIG. 13 shows an example in which the gNB 830 includes multiple antennas 840, the gNB 830 may also include a single antenna 840.

[0085] The base station device 850 includes a controller 851, a memory 852, a network interface 853, a wireless communication interface 855, and a connection interface 857. The controller 851, the memory 852, and the network interface 853 are similar to the controller 821, the memory 822, and the network interface 823 described with reference to FIG.

[0086] The wireless communication interface 855 supports any cellular communication method (e.g., LTE and LTE-Advanced) and provides wireless communication to terminals located in a sector corresponding to the RRH 860 via the RRH 860 and the antenna 840. The wireless communication interface 855 may typically include, for example, a BB processor 856. The BB processor 856 is similar to the BB processor 826 described with reference to FIG. 12 except that the BB processor 856 is connected to the RF circuit 864 of the RRH 860 via a connection interface 857. As shown in FIG. 13, the wireless communication interface 855 may include multiple BB processors 856. For example, the multiple BB processors 856 may be compatible with multiple frequency bands used by the gNB 830. Although FIG. 13 illustrates an example in which the wireless communication interface 855 includes multiple BB processors 856, the wireless communication interface 855 may also include a single BB processor 856.

[0087] The connection interface 857 is an interface for connecting the base station device 850 (wireless communication interface 855) to the RRH 860. The connection interface 857 may be a communication module for communication in the above-mentioned high-speed line that connects the base station device 850 (wireless communication interface 855) to the RRH 860.

[0088] The RRH 860 includes a connection interface 861 and a wireless communication interface 863 .

[0089] The connection interface 861 is an interface for connecting the RRH 860 (wireless communication interface 863) to the base station device 850. The connection interface 861 may be a communication module for communication in the above-mentioned high-speed line.

[0090] The wireless communication interface 863 transmits and receives wireless signals via the antenna 840. The wireless communication interface 863 may typically include, for example, an RF circuit 864. The RF circuit 864 may include, for example, a mixer, a filter, and an amplifier, and may transmit and receive wireless signals via the antenna 840. As shown in FIG. 13 , the wireless communication interface 863 may include multiple RF circuits 864. For example, the multiple RF circuits 864 may support multiple antenna elements. Although FIG. 13 shows an example in which the wireless communication interface 863 includes multiple RF circuits 864, the wireless communication interface 863 may also include a single RF circuit 864.

[0091] [3-3. Application examples for terminal devices] (First application example) 14 is a block diagram showing an example of an exemplary configuration of a smartphone 900 to which the technology of this disclosure can be applied. The smartphone 900 includes a processor 901, a memory 902, a storage device 903, an external connection interface 904, an image capture device 906, a sensor 907, a microphone 908, an input device 909, a display device 910, a speaker 911, a wireless communication interface 912, one or more antenna switches 915, one or more antennas 916, a bus 917, a battery 918, and an auxiliary controller 919.

[0092] The processor 901 is, for example, a CPU or a system-on-chip (SoC) and can control the functions of the application layer and other layers of the smartphone 900. The memory 902 includes RAM and ROM and stores data and programs executed by the processor 901. The storage device 903 may include storage media such as semiconductor memory and a hard disk. The external connection interface 904 is an interface for connecting external devices (e.g., memory cards and universal serial bus (USB) devices) to the smartphone 900.

[0093] The imaging device 906 includes an image sensor (e.g., a charge-coupled device (CCD) and a complementary metal-oxide semiconductor (CMOS)) and generates a captured image. The sensor 907 may include a set of sensors, such as a measurement sensor, a gyro sensor, a geomagnetic sensor, and an acceleration sensor. The microphone 908 converts sounds input to the smartphone 900 into audio signals. The input device 909 includes, for example, a touch sensor, a keypad, a keyboard, buttons, or switches arranged to detect touches on the screen of the display device 910 and receives operations or information input from a user. The display device 910 includes a screen (e.g., a liquid crystal display (LCD) and an organic light-emitting diode (OLED) display) and displays an output image of the smartphone 900. The speaker 911 converts audio signals output from the smartphone 900 into sound.

[0094] The wireless communication interface 912 supports any cellular communication system (e.g., LTE and LTE-Advanced) and performs wireless communication. The wireless communication interface 912 typically includes, for example, a BB processor 913 and an RF circuit 914. The BB processor 913 can perform, for example, encoding / decoding, modulation / demodulation, multiplexing / demultiplexing, and various types of signal processing for wireless communication. At the same time, the RF circuit 914 can include, for example, a mixer, a filter, and an amplifier, and can transmit and receive wireless signals via an antenna 916. The wireless communication interface 912 may be a single chip module on which the BB processor 913 and the RF circuit 914 are integrated. As shown in FIG. 14, the wireless communication interface 912 may include multiple BB processors 913 and multiple RF circuits 914. Although FIG. 14 illustrates an example in which the wireless communication interface 912 includes multiple BB processors 913 and multiple RF circuits 914, the wireless communication interface 912 may include a single BB processor 913 or a single RF circuit 914.

[0095] In addition to the cellular communication system, the wireless communication interface 912 may support other types of wireless communication systems, such as a short-range wireless communication system, a proximity communication system, a wireless local network (LAN) system, etc. In this case, the wireless communication interface 912 may include a BB processor 913 and an RF circuit 914 for each wireless communication system.

[0096] Each of the antenna switches 915 switches the connection of the antenna 916 between a plurality of circuits (for example, circuits used for different wireless communication methods) included in the wireless communication interface 912.

[0097] Each of the antennas 916 includes a single or multiple antenna elements (e.g., multiple antenna elements included in a MIMO antenna) and is used to transmit and receive radio signals over the wireless communication interface 912. As shown in Fig. 14, the smartphone 900 may include multiple antennas 916. Although Fig. 14 shows an example in which the smartphone 900 includes multiple antennas 916, the smartphone 900 may also include a single antenna 916.

[0098] The smartphone 900 may include an antenna 916 for each wireless communication method. In this case, the antenna switch 915 may be omitted from the arrangement of the smartphone 900.

[0099] The bus 917 interconnects the processor 901, memory 902, storage device 903, external connection interface 904, image capture device 906, sensor 907, microphone 908, input device 909, display device 910, speaker 911, wireless communication interface 912, and auxiliary controller 919. A battery 918 provides power to each block of the smartphone 900 shown in Fig. 14 via power supply lines, which are partially indicated by dotted lines in the drawing. The auxiliary controller 919 operates the minimum necessary functions of the smartphone 900, for example, in sleep mode.

[0100] (Second application example) 15 is a block diagram showing an example of a schematic layout of a car navigation device 920 to which the technology of the present disclosure can be applied. The car navigation device 920 includes a processor 921, a memory 922, a global positioning system (GPS) module 924, a sensor 925, a data interface 926, a content player 927, a storage medium interface 928, an input device 929, a display device 930, a speaker 931, a wireless communication interface 933, one or more antenna switches 936, one or more antennas 937, and a battery 938.

[0101] The processor 921 is, for example, a CPU or an SoC, and can control the navigation function and other functions of the car navigation device 920. The memory 922 includes RAM and ROM, and stores data and programs executed by the processor 921.

[0102] The GPS module 924 measures the position (e.g., latitude, longitude, and altitude) of the car navigation device 920 using GPS signals received from GPS satellites. The sensor 925 may include a set of sensors, such as a gyro sensor, a geomagnetic sensor, and an air pressure sensor. The data interface 926 connects to, for example, an in-vehicle network 941 via a terminal (not shown) to acquire data generated by the vehicle (e.g., vehicle speed data).

[0103] The content player 927 plays content stored on a storage medium (e.g., a CD or DVD), which is inserted into a storage medium interface 928. The input device 929 includes, for example, a touch sensor, button, or switch arranged to detect a touch on the screen of the display device 930, and receives operations or information input from a user. The display device 930 includes, for example, an LCD or OLED display screen, and displays images of the navigation function or played content. The speaker 931 outputs sounds of the navigation function or played content.

[0104] The wireless communication interface 933 can support any cellular communication system (e.g., LTE and LTE-Advanced) and perform wireless communication. The wireless communication interface 933 typically includes, for example, a BB processor 934 and an RF circuit 935. The BB processor 934 can perform, for example, encoding / decoding, modulation / demodulation, multiplexing / demultiplexing, and various types of signal processing for wireless communication. At the same time, the RF circuit 935 can include, for example, a mixer, a filter, and an amplifier, and can transmit and receive wireless signals via an antenna 937. The wireless communication interface 933 may be a single chip module on which the BB processor 934 and the RF circuit 935 are integrated. As shown in FIG. 15, the wireless communication interface 933 may include multiple BB processors 934 and multiple RF circuits 935. Although FIG. 15 illustrates an example in which the wireless communication interface 933 includes multiple BB processors 934 and multiple RF circuits 935, the wireless communication interface 933 may include a single BB processor 934 or a single RF circuit 935.

[0105] In addition to the cellular communication system, the wireless communication interface 933 can support other types of wireless communication systems, such as a short-range wireless communication system, a proximity communication system, and a wireless LAN system. In this case, the wireless communication interface 933 may include a BB processor 934 and an RF circuit 935 for each wireless communication system.

[0106] Each of the antenna switches 936 switches the connection destination of the antenna 937 between a plurality of circuits (for example, circuits used for different wireless communication methods) included in the wireless communication interface 933 .

[0107] Each of the antennas 937 includes a single antenna element or multiple antenna elements (for example, multiple antenna elements included in a MIMO antenna) and is used for transmitting and receiving radio signals via the wireless communication interface 933. As shown in FIG. 15, the car navigation device 920 may include multiple antennas 937. Although FIG. 15 shows an example in which the car navigation device 920 includes multiple antennas 937, the car navigation device 920 may include a single antenna 937.

[0108] The car navigation device 920 may include an antenna 937 for each wireless communication method. In this case, the antenna switch 936 may be omitted from the arrangement of the car navigation device 920.

[0109] 15 via power supply lines, which are partially indicated by dotted lines in the drawing. The battery 938 stores power provided by the vehicle.

[0110] The technology of the present disclosure may be realized as an in-vehicle system (or vehicle) 940 including one or more blocks of a car navigation device 920, an in-vehicle network 941, and a vehicle module 942. The vehicle module 942 generates vehicle data (e.g., vehicle speed, engine speed, and failure information) and outputs the generated data to the in-vehicle network 941.

[0111] The various example blocks and components described in connection with this disclosure may be implemented or performed by a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA or other programmable logic device, discrete-gate or transistor logic circuitry, discrete hardware components, or any combination thereof, designed to perform the functions described herein. A general-purpose processor may be a microprocessor. Alternatively, the processor may be any conventional processor, controller, microcontroller, and / or state machine. A processor may also be implemented as a combination of computing devices. For example, a combination of a DSP and a microprocessor, a combination of multiple microprocessors, a combination of one or more microprocessors in conjunction with a DSP core, and / or any other such configuration.

[0112] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored in or transmitted as one or more instructions or code on a non-transitory computer-readable storage medium. Other examples and implementations are within the scope and spirit of this disclosure and the appended claims. For example, considering the nature of software, the functions described above may be implemented in software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Features that implement the functions may also be physically located in various locations, including distributed implementations where parts of the functions are implemented in different physical locations.

[0113] Furthermore, disclosure of components within or separate from other components is exemplary and includes incorporating all, most, and / or some of the elements as part of one or more single or separate structures, potentially allowing various other architectures to be implemented to achieve the same functionality.

[0114] Non-transitory computer-readable media can be any available non-transitory media that can be accessed by a general purpose or special purpose computer, including, but not limited to, RAM, ROM, EEPROM, flash memory, CD-ROM, DVD or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store instructions or data structured form of desired program code components, a general purpose or special purpose computer, or a general purpose or special purpose processor.

[0115] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be apparent to those skilled in the art. The general principles defined herein may be applied to other variations without departing from the scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed.

[0116] The embodiments of the present disclosure further include the following items. 1. An electronic device for use in associative learning in a central processing unit, comprising a processing circuit, the processing circuit comprising: determining a set of distributed nodes among a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specific correlation requirement; and generating global model parameters based on local model parameters of the set of distributed nodes, the local model parameters being arranged such that the set of distributed nodes are generated based on their respective local training data. 2. The specific correlation requirement is: Item 10. The electronic device of item 1, wherein the time interval between collection times of local training data of the set of distributed nodes is greater than a predetermined time interval threshold. 3. The specific correlation requirement is: Item 1, wherein the distance between node locations of the set of distributed nodes is greater than a predetermined distance threshold. 4. The specific correlation requirement is: Item 1. The electronic device of item 1, wherein each distributed node of the set of distributed nodes is located outside the associated exclusion regions of the remaining distributed nodes of the set of distributed nodes, and the associated exclusion region of each distributed node is defined by an area within which each point is closer to the distributed node than the exclusion distance, or is defined by a set of a second predetermined number of distributed nodes that are closest to the distributed node. 5. The specific correlation requirement is: Item 10. The electronic device of item 1, wherein the generative models of the local training data of the set of distributed nodes are different. 6. The specific correlation requirement is: Item 10. The electronic device of item 1, wherein the data sources of the local training data for the set of distributed nodes are different. 7. Determining the set of distributed nodes from the plurality of distributed nodes for generating global model parameters includes: obtaining data collection information of local training data of the plurality of distributed nodes, the data collection information being usable to determine correlations between local training data of the plurality of distributed nodes; and selecting the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on data collection information of local training data of the plurality of distributed nodes. 8. The electronic device described in item 7, wherein the data collection information includes one or more of the collection time of the local training data, node location, generative model, and data source. 9. The processing circuitry further comprises: obtaining network performance parameters for the plurality of distributed nodes; 8. The electronic device of claim 7, wherein the electronic device is configured to select the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on network performance parameters of the plurality of distributed nodes, wherein the network performance parameters of the set of distributed nodes satisfy specific network performance requirements. 10. The processing circuitry further comprises: obtaining local training performance parameters of the plurality of distributed nodes; Item 8. The electronic device of item 7, configured to select the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on local training performance parameters of the plurality of distributed nodes, wherein the local training performance parameters of the set of distributed nodes meet specific local training performance requirements. 11. Determining the set of distributed nodes from the plurality of distributed nodes for generating global model parameters includes: receiving local model parameters from a distributed node of the plurality of distributed nodes; and after receiving local model parameters of a first predetermined number of distributed nodes, determining the first predetermined number of distributed nodes as the set of distributed nodes among the plurality of distributed nodes for generating global model parameters. 12. An electronic device for use in associative learning, comprising a processing circuit, the processing circuit comprising: determining that a particular distributed node will be used to generate global model parameters, where correlations between the local training data of the particular distributed node and the local training data of other distributed nodes for generating the global model parameters meet specific correlation requirements; transmitting local model parameters of the particular distributed node to a central processing unit, the local model parameters being configured such that the particular distributed node is generated based on its local training data. 13. Determining that the particular distributed node will be used to generate global model parameters comprises: transmitting data collection information of the local training data of the particular distributed node to the central processing unit, wherein the data collection information can be used to determine correlations between the local training data of the particular distributed node and local training data of other distributed nodes determined to be used in generating global model parameters; and determining, based on an instruction from the central processing unit to upload local model parameters, that the particular distributed node will be used to generate global model parameters. 14. Determining that the particular distributed node will be used to generate global model parameters comprises: determining, based on data collection information of local training data of the particular distributed node and other distributed nodes determined to be used in generating global model parameters, that a correlation between the local training data of the particular distributed node and the local training data of the other distributed nodes determined to be used in generating global model parameters satisfies a specific correlation requirement; obtaining channel resources for transmitting local model parameters to the central processing unit; and determining, after successful acquisition of channel resources, that the particular distributed node will be used to generate global model parameters. 15. The data collection information includes node locations of distributed nodes; and Item 15. The electronic device of item 14, wherein the specific correlation request includes that the specific distributed node be located outside the associated exclusion region of the other distributed nodes determined to be used in generating the global model parameters, and the associated exclusion region of a distributed node is defined by an area within which the distance between each point and the distributed node is less than an exclusion distance, or is defined by a set of a second predetermined number of distributed nodes that are closest to the distributed node. 16. The processing circuitry further comprises: Item 15. The electronic device of item 14, configured to adjust parameters for acquiring channel resources based on one or both of network performance parameters and local training performance parameters of the particular distributed node. 17. The processing circuitry further comprises: Item 15. The electronic device of item 14, configured to stop acquiring channel resources for transmitting local model parameters to the central processing unit upon receiving an instruction to stop uploading from the central processing unit. 18. The processing circuit further comprises: determining the number of other distributed nodes determined to be used in generating the global model parameters; and Item 15. The electronic device of item 14, configured to stop acquiring channel resources for transmitting local model parameters to the central processing unit if the number is greater than or equal to a first predetermined number. 19. A method for associative learning in a central processing unit, comprising: determining a set of distributed nodes among a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specific correlation requirement; generating global model parameters based on local model parameters of the set of distributed nodes, the local model parameters being generated based on local training data for each of the set of distributed nodes. 20. A method for federated learning in a particular distributed node, comprising: determining that the particular distributed node will be used to generate global model parameters, wherein correlations between the local training data of the particular distributed node and the local training data of the other distributed nodes determined to be used to generate the global model parameters meet specific correlation requirements; sending local model parameters of the particular distributed node to a central processing unit, the local model parameters being generated by the particular distributed node based on its local training data. 21. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method described in item 19 or 20. 22. An apparatus for use in associative learning, comprising components for performing each step of the method described in item 19 or 20.

Claims

1. 1. An electronic device for use in associative learning in a central processing unit, comprising: a processing circuit; determining a set of distributed nodes among a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specified correlation requirement; and generating global model parameters based on local model parameters of the set of distributed nodes, the local model parameters being arranged such that the set of distributed nodes are generated based on their respective local training data.

2. The specific correlation requirement may be: The electronic device of claim 1 , further comprising: a time interval between collection times of local training data for the set of distributed nodes being greater than a predetermined time interval threshold.

3. The specific correlation requirement may be: The electronic device of claim 1 , wherein a distance between node locations of the set of distributed nodes is greater than a predetermined distance threshold.

4. The specific correlation requirement may be:

2. The electronic device of claim 1, wherein each distributed node of the set of distributed nodes is located outside the associated exclusion regions of the remaining distributed nodes of the set of distributed nodes, the associated exclusion region of each distributed node being defined by a region within which each point is closer to the distributed node than an exclusion distance, or by a set of a second predetermined number of distributed nodes that are closest to the distributed node.

5. The specific correlation requirement may be: The electronic device of claim 1 , wherein the generative models of the local training data of the set of distributed nodes are different.

6. The specific correlation requirement may be: The electronic device of claim 1 , wherein the data sources of the local training data for the set of distributed nodes are different.

7. Determining the set of distributed nodes from the plurality of distributed nodes for generating global model parameters comprises: obtaining data collection information of local training data of the plurality of distributed nodes, the data collection information being usable to determine correlations between local training data of the plurality of distributed nodes; and selecting the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on data collection information of local training data of the plurality of distributed nodes.

8. The electronic device of claim 7 , wherein the data collection information includes one or more of a collection time, a node location, a generative model, and a data source of the local training data.

9. The processing circuitry further comprises: obtaining network performance parameters for the plurality of distributed nodes; 8. The electronic device of claim 7, configured to select the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on network performance parameters of the plurality of distributed nodes, wherein the network performance parameters of the set of distributed nodes meet specific network performance requirements.

10. The processing circuitry further comprises: obtaining local training performance parameters of the plurality of distributed nodes; 8. The electronic device of claim 7, configured to select the set of distributed nodes from the plurality of distributed nodes for generating global model parameters based on local training performance parameters of the plurality of distributed nodes, wherein the local training performance parameters of the set of distributed nodes meet specific local training performance requirements.

11. Determining the set of distributed nodes from the plurality of distributed nodes for generating global model parameters comprises: receiving local model parameters from a distributed node of the plurality of distributed nodes; and after receiving the local model parameters of a first predetermined number of distributed nodes, determining the first predetermined number of distributed nodes as the set of distributed nodes of the plurality of distributed nodes for generating global model parameters.

12. 1. An electronic device for use in associative learning, comprising: a processing circuit; determining that a particular distributed node will be used to generate global model parameters, where correlations between the local training data of the particular distributed node and the local training data of other distributed nodes for generating the global model parameters meet specific correlation requirements; transmitting local model parameters of the particular distributed node to a central processing unit, the local model parameters being configured such that the particular distributed node is generated based on its local training data.

13. Determining that the particular distributed node will be used to generate global model parameters includes: transmitting data collection information of the local training data of the particular distributed node to the central processing unit, wherein the data collection information can be used to determine correlations between the local training data of the particular distributed node and local training data of other distributed nodes determined to be used in generating global model parameters; and determining, based on an instruction from the central processing unit to upload local model parameters, that the particular distributed node will be used to generate global model parameters.

14. Determining that the particular distributed node will be used to generate global model parameters includes: determining, based on data collection information of local training data of the particular distributed node and other distributed nodes determined to be used in generating global model parameters, that a correlation between the local training data of the particular distributed node and the local training data of the other distributed nodes determined to be used in generating global model parameters satisfies a specific correlation requirement; obtaining channel resources for transmitting local model parameters to the central processing unit; and determining, after successful acquisition of channel resources, that the particular distributed node will be used to generate global model parameters.

15. The data collection information includes node locations of distributed nodes; and 15. The electronic device of claim 14, wherein the specific correlation requirement includes the specific distributed node being located outside the associated exclusion regions of the other distributed nodes determined to be used in generating global model parameters, the associated exclusion region of a distributed node being defined by a region within which each point is closer to the distributed node than an exclusion distance, or by a set of a second predetermined number of distributed nodes that are closest to the distributed node.

16. The processing circuitry further comprises:

15. The electronic device of claim 14, configured to adjust parameters for acquiring channel resources based on one or both of network performance parameters and local training performance parameters of the particular distributed node.

17. The processing circuitry further comprises:

15. The electronic device of claim 14, arranged to stop acquiring channel resources for transmitting local model parameters to the central processing unit upon receiving an instruction to stop uploading from the central processing unit.

18. The processing circuitry further comprises: determining the number of other distributed nodes determined to be used in generating the global model parameters; and 15. The electronic device of claim 14, arranged to stop acquiring channel resources for transmitting local model parameters to the central processing unit if the number is greater than or equal to a first predetermined number.

19. A method for associative learning in a central processing unit, comprising: determining a set of distributed nodes among a plurality of distributed nodes for generating global model parameters, wherein correlations between local training data of the set of distributed nodes satisfy a specific correlation requirement; generating global model parameters based on local model parameters of the set of distributed nodes, the local model parameters being generated based on local training data for each of the set of distributed nodes.

20. A method for federated learning in a particular distributed node, comprising: determining that the particular distributed node will be used to generate global model parameters, wherein correlations between the local training data of the particular distributed node and the local training data of the other distributed nodes determined to be used to generate the global model parameters meet specified correlation requirements; sending local model parameters of the particular distributed node to a central processing unit, the local model parameters being generated by the particular distributed node based on its local training data.

21. A non-transitory computer readable storage medium storing instructions which, when executed by a processor, cause the processor to perform the method of claim 19 or 20.

22. An apparatus for use in associative learning, comprising components for performing the steps of the method according to claim 19 or 20.

Images