Apparatus, method and system for machine learning

Abstract

Apparatuses, methods, and systems for machine learning are disclosed. In some embodiments, an electronic device for a first UE is disclosed. The electronic device includes processing circuitry. The processing circuitry is configured to: obtain a first artificial intelligence model trained by a first UE; receiving a second artificial intelligence model trained by a second UE; determining a model difference between the second artificial intelligence model and the first artificial intelligence model; and sending the model difference to a federated learning server.

Description

Technical Field

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

[0002] Artificial Intelligence (AI) technology is increasingly being used. Artificial Intelligence models can be trained through machine learning. Improved devices, methods, and systems for machine learning are desired. Summary of the invention

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

[0004] One aspect of the present disclosure relates to an electronic device for a first UE. The electronic device includes a processing circuit. The processing circuit is configured to: obtain a first artificial intelligence model trained by the first UE; receive a second artificial intelligence model trained by a second UE; determine a model difference between the second artificial intelligence model and the first artificial intelligence model; and send the model difference to a federated learning server.

[0005] Another aspect of the present disclosure relates to an electronic device for a network side. The electronic device includes a processing circuit. The processing circuit is configured to: receive a first artificial intelligence model from a first UE, the first artificial intelligence model is trained by the first UE in a federated learning process; receive a model difference from a second UE, the model difference represents a difference between the first artificial intelligence model and a second artificial intelligence model trained by the second UE in the federated learning process; and generate a global artificial intelligence model based on aggregating at least the first artificial intelligence model and the model difference.

[0006] One aspect of the present disclosure relates to a method performed by a first UE. The method includes: obtaining a first artificial intelligence model trained by the first UE; receiving a second artificial intelligence model trained by a second UE; determining a model difference between the second artificial intelligence model and the first artificial intelligence model; and sending the model difference to a federated learning server.

[0007] Another aspect of the present disclosure relates to a method performed by a device on a network side. The method includes: receiving a first artificial intelligence model from a first UE, the first artificial intelligence model being trained by the first UE in a federated learning process; receiving a model difference from a second UE, the model difference representing a difference between the first artificial intelligence model and a second artificial intelligence model trained by the second UE in a federated learning process; and generating a global artificial intelligence model based on aggregating at least the first artificial intelligence model and the model difference.

[0008] Another aspect of the present disclosure relates to a computer-readable storage medium storing one or more instructions, which, when executed by one or more processing circuits of an electronic device, causes the electronic device to perform any method as described in the present disclosure.

[0009] Another aspect of the present disclosure relates to a computer program product, comprising a computer program, which implements any method as described in the present disclosure when executed by a processor.

[0010] Another aspect of the present disclosure relates to an apparatus comprising means for performing any method as described in the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The above and other purposes and advantages of the present disclosure are further described below in conjunction with specific embodiments and with reference to the accompanying drawings. In the accompanying drawings, the same or corresponding technical features or components will be represented by the same or corresponding reference numerals.

[0012] Figure 1 A schematic diagram of an exemplary system for federated learning is shown.

[0013] Figure 1A A schematic diagram of an exemplary system for federated learning including some embodiments of the present disclosure is shown.

[0014] Figure 2 An exemplary block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0015] Figure 3 An exemplary block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0016] Figure 4 A flowchart of a method for machine learning according to some embodiments of the present disclosure is shown.

[0017] Figure 5 A flowchart of a method for machine learning according to some embodiments of the present disclosure is shown.

[0018] Figure 6 A communication process for machine learning according to some embodiments of the present disclosure is shown.

[0019] Figure 7 Another communication process for machine learning according to some embodiments of the present disclosure is shown.

[0020] Figure 8 A schematic diagram of determining model differences according to some embodiments of the present disclosure is shown.

[0021] Fig. 9A schematic diagram of resource allocation according to some embodiments of the present disclosure is shown.

[0022] Fig.10 A block diagram of a cellular system is shown in which techniques according to some embodiments of the present disclosure may be implemented.

[0023] Fig.11 An exemplary process for screening member UEs participating in federated learning according to some embodiments of the present disclosure is shown.

[0024] Fig.12 An exemplary process of federated learning according to some embodiments of the present disclosure is described.

[0025] Although the embodiments described in the present disclosure may be susceptible to various modifications and alternative forms, specific embodiments thereof are shown as examples in the drawings and described in detail herein. It should be understood that the drawings and detailed description thereof are not intended to limit the embodiments to the specific forms disclosed, but on the contrary, the purpose is to cover all modifications, equivalents and alternatives within the spirit and scope of the claims. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present disclosure will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of the embodiments are described in the specification. However, it should be understood that many implementation-specific settings must be made in the process of implementing the embodiments in order to achieve the specific goals of the developer, for example, to meet those restrictions related to the device and the service, and these restrictions may vary depending on the implementation. In addition, it should be understood that although the development work may be very complex and time-consuming, it is only a routine task for those skilled in the art who benefit from the content of this disclosure.

[0027] It should also be noted here that in order to avoid obscuring the present disclosure due to unnecessary details, only the processing steps and / or equipment structures closely related to at least the scheme according to the present disclosure are shown in the accompanying drawings, while other details that are not very relevant to the present disclosure are omitted.

[0028] 1. Exemplary System

[0029] The application of artificial intelligence technology is becoming more and more extensive. Artificial intelligence models can be trained through machine learning. Federated Learning (FL) is a distributed learning technology. In a system using federated learning, each of the multiple training devices uses its local data as training data to train the corresponding local model. Each training device can upload the trained corresponding local model to the server. The server aggregates multiple local models received from multiple training devices to obtain a global model. In the process of federated learning, each training device uploads a local model without having to exchange training data with the server or other training devices. Therefore, federated learning can train artificial intelligence models using the training capabilities and training data of a large number of users while ensuring data security and user privacy. This helps to improve training efficiency and improve the performance of the resulting artificial intelligence model.

[0030] Figure 1 A schematic diagram of an exemplary system 100 for federated learning is shown. System 100 may include a server 110. Server 110 may act as a server for federated learning. System 100 may also include a plurality of UEs 120-1 to 120-5. The plurality of UEs may be uniformly represented using 120. Each UE 120 may be communicatively coupled to server 110. Each UE 120 may act as a training device for federated learning.

[0031] During federated learning, the training of the artificial intelligence model can be an iterative process including multiple training rounds. In one training round, each UE 120 can perform training using the local data of the UE to obtain a corresponding trained local model 121. The local data of different UEs 120 may be different, so the obtained local models 121 may be different. For example, UE 120-1, 120-2, 120-3, 120-4, 120-5 can obtain trained local models 121-1, 121-2, 121-3, 121-4, 121-5, respectively. Each UE 120 can upload the trained local model 121 to the server 110 via an uplink between the UE and the server 110. The server 110 can aggregate multiple local models 121 received from each UE 120 to generate a global model 111.

[0032] Then, the global model 111 may be distributed to each UE 120 via a downlink between the server 110 and each UE 120. Based on the received global model 111, each UE 120 may update the local model 121 of the UE.

[0033] In the next training round, each UE 120 may continue to train the updated local model 121 of the UE, thereby obtaining a trained local model 121. Once again, each UE 120 may upload the trained local model 121 obtained in the current training round to the server 110. The server 110 may aggregate multiple local models 121 received from various UEs 120 to generate a global model 111.

[0034] The global model 111 may be distributed again to each UE 120 for subsequent training rounds. This process may be repeated until an end condition is reached. The performance of the global model 111 may be improved by iteratively going through multiple training rounds.

[0035] As artificial intelligence technology develops, artificial intelligence models become larger and more complex. The inventors of the present application recognize that systems using federated learning face various challenges.

[0036] For example, uploading complex AI models will require a much larger amount of data to be transferred. Figure 1 In the system 100, the local model 121 obtained by each UE 120 may become very large. In some cases, the local model 121 may need to be represented by a large number of model parameters. In order to upload the local model 121 to the server 110, the UE 120 needs to upload a large amount of data containing these model parameters. In most communication systems (e.g., wireless communication systems), the uplink for uploading data is limited (e.g., having limited bandwidth, etc.). Uploading a local model may generate a significant load on the uplink. This may cause undesirable congestion, delay and / or data loss. For another example, due to the different characteristics of each UE participating in federated learning (e.g., location, communication conditions, local data held, etc.), each UE may perform different training tasks. In some scenarios, during federated learning for the same global image recognition model, some UEs may use face image data to train local models, while other UEs may use car image data to train local models. Direct aggregation of multiple local models trained for different tasks may reduce the training effect. In addition, there are one or more other problems to be solved.

[0037] The inventors of the present application recognize that there may be model redundancy between local models trained by different UEs. By exchanging local models between UEs and removing model redundancy before uploading the model, it is possible to reduce the occupancy of uplink resources. The inventors of the present application also recognize that multiple UEs between which local models are to be exchanged can be selected to maximize the removed model redundancy. In addition, the inventors of the present application also recognize that by providing high-quality communication resources for multiple UEs that generate similar local models, it is possible to improve the training efficiency of the entire system. The present disclosure provides improved techniques for machine learning including these and other aspects.

[0038] Figure 1A A schematic diagram of an exemplary system 100A for federated learning including some embodiments of the present disclosure is shown. Similar to system 100, system 100A may include a server 110 and multiple UEs 120. Server 110 may act as a federated learning server. Multiple UEs 120 may act as training devices for federated learning.

[0039] Similar to system 100, in one training round of federated learning, UEs 120-1, 120-2, 120-3, 120-4, and 120-5 can obtain trained local models 121-1, 121-2, 121-3, 121-4, and 121-5 respectively through local model training. Unlike system 100, not all of the multiple UEs 120-1 to 120-5 in system 100A upload corresponding local models. Instead, only a portion of the UEs 120-1 to 120-5 are configured to upload their corresponding local models 121 to server 110. Each of the remaining UEs can be configured to upload a model difference between the UE and a corresponding UE in the portion of UEs. The model difference can represent a non-redundant portion between two local models. In other words, each of the remaining UEs does not upload a model redundancy between the local model of the UE and the corresponding UE.

[0040] By way of example and not limitation, Figure 1AIn the illustrated embodiment, UE 120-1 and UE 120-3 may be configured to upload respective local models 121-1 and 121-3 to server 110. UE 120-2 may be configured to upload model differences between respective local model 121-2 and local model 121-1 of UE 120-1 to server 110. UE 120-4 may be configured to upload model differences between respective local model 121-4 and local model 121-3 of UE 120-3 to server 110. UE 120-5 may be configured to upload model differences between respective local model 121-5 and local model 121-3 of UE 120-3 to server 110. Since specific UEs (e.g., UE 120-2, 120-4, 120-5) are configured to upload model differences instead of local models to server 110, the amount of data uploaded by these UEs may be significantly reduced. Accordingly, the load of the uplink associated with these UEs may be significantly reduced.

[0041] According to some embodiments of the present disclosure, multiple UEs 120 participating in federated learning may be grouped. A specific UE in each group may be configured to upload its corresponding local model, while other UEs in the group may be configured to upload the model difference between the UE and the specific UE.

[0042] Multiple UEs 120 participating in federated learning can be grouped according to one or more grouping criteria. In some embodiments, the one or more grouping criteria may include the spatial location of the UE. For example, multiple UEs that are spatially close to each other can be divided into the same group. Specifically, if the distance between two UEs is less than a first predetermined distance threshold, the two UEs can be divided into the same group. For another example, multiple UEs that are close enough to a specific geographic location can be divided into the same group. Specifically, if the distance between the UE and a specific geographic location is less than a second predetermined distance threshold, the UE can be divided into a group associated with the specific geographic location. There can be multiple specific geographic locations, and each specific geographic location can be any spatial location of interest.

[0043] In an alternative embodiment, one or more grouping criteria may include a trust relationship between multiple UEs. In this case, multiple UEs that have established or are capable of establishing a trust relationship may be divided into the same group. As an example, the establishment of a trust relationship may include the establishment of a specific channel. For example, if a specific channel has been established between two UEs, it can be considered that there is a trust relationship between the two UEs. In some scenarios, the specific channel may be a sidelink between the two UEs. In some scenarios, the specific channel may be a Wi-Fi link. In other scenarios, the specific channel may be any other type of link through which two UEs can communicate with each other. Additionally, the establishment of a trust relationship may also include a discovery process or an authentication process between UEs.

[0044] In an optional embodiment, one or more grouping criteria may include one or more attributes associated with a specific channel between multiple UEs. The one or more attributes may include, but are not limited to, the quality of a specific channel. Specifically, if the quality of a specific channel between two UEs is higher than a predetermined threshold, the two UEs may be divided into the same group. Any suitable parameter may be used to measure the quality of a specific channel. As an example, the parameters for measuring the quality of a specific channel may include one or more of a reference signal received power (RSRP), a reference signal received quality (RSRQ), a quality of service (QoS), or a signal to interference plus noise ratio (SINR) associated with the specific channel.

[0045] In a preferred embodiment, one or more grouping criteria may include a model distance between local models trained by UEs. Model distance may be used to measure the degree of difference between local models trained by two UEs. Small model distance corresponds to small model difference. In other words, small model distance corresponds to high model similarity. Preferably, multiple UEs having a model distance below a predetermined model distance threshold may be divided into the same group. In this case, since the multiple local models of the multiple UEs in each group have a sufficiently small model distance, the model difference uploaded will be sufficiently small, thereby further reducing the amount of uploaded data.

[0046] Those skilled in the art will appreciate that the one or more grouping criteria described above for grouping multiple UEs 120 participating in federated learning are exemplary and non-restrictive. The one or more grouping criteria may be applied individually or in combination. In some embodiments, only a grouping criterion based on a trust relationship may be applied. For example, only the side link connection between multiple UEs 120 may be considered without considering the model distance. In other embodiments, a grouping criterion based on model distance may be used in combination with a grouping criterion based on spatial location or a grouping criterion based on a trust relationship.

[0047] According to an embodiment of the present disclosure, the grouping of UEs may be determined by one or more entities associated with federated learning. In some embodiments, an autonomous grouping method may be implemented, in which multiple UEs participating in federated learning are autonomously grouped. The UE may report the grouping results to a federated learning server connected to these network side devices through one or more network side devices (e.g., Wi-Fi access points, cellular base stations, or core network nodes). In some embodiments, a centralized grouping method may be implemented. In some scenarios, the centralized grouping method may be performed by a federated learning server. For example, when a grouping criterion based on model distance is used, grouping may be performed by a federated learning server. This is because the federated learning server is able to obtain the local model trained by each UE and calculate the model distance. The federated learning server may be located at the application function (AF) of the cellular network, so grouping may be performed by the AF. In some scenarios, the centralized grouping method may be performed by a network side device (e.g., Wi-Fi access point, cellular base station, core network node). For example, in the case of a grouping criterion based on a trust relationship, grouping may be performed by a base station, a core network node, or an access point. Specifically, a base station (e.g., gNB) can learn about the sidelink connections (and additionally, the quality of the sidelinks) between the UEs and perform grouping based on this link information. In some scenarios, the federated learning server can collaborate with network-side devices (access points, base stations, or core network nodes, etc.) to perform grouping. This allows the model information collected by the federated learning server to be used in combination with the link information collected by the network-side devices to potentially achieve improved grouping. In some embodiments, the one or more grouping criteria can be performed as part of the process of screening UEs participating in federated learning, as described later with respect to Fig.11 Further described.

[0048] By way of example and not limitation, Figure 1A The illustrated embodiment applies a grouping criterion based on model distance. In this example, UEs 120-1 to 120-5 may be divided into two groups 130A and 130B, where group 130A includes UEs 120-1 and 120-2, and group 130B includes UEs 120-3, 120-4, and 120-5.

[0049] In the process of training an artificial intelligence model, multiple UEs performing the same or similar training tasks may obtain similar local models (i.e., having a small model distance). For example, if two UEs use face image data to train an image recognition model, it can be considered that the training tasks of the two UEs are associated with face recognition. In this case, the two local models obtained are likely to be similar. If two UEs use face image data and car image data to train an image recognition model, respectively, it can be considered that the training tasks of the two UEs are associated with face recognition and car recognition, respectively. In this case, the model difference between the two local models obtained may be large. Therefore, it is expected that multiple UEs with the same or similar training tasks will be divided into the same group. If the server 110 already knows the training tasks of each UE, the server 110 can pre-group each UE based on the training tasks. However, in most cases, the server 110 is not aware of the training tasks / training data of each UE. For example, the server 110 may not know whether the local data of the UE is face image data or car image data. Therefore, each UE can be grouped based on a comparison between the local models trained by each UE (e.g., the local models obtained in a certain training round). The model distance between each local model can be determined by comparing the models. Based on the comparison, multiple UEs with small model distances can be divided into the same group. Correspondingly, multiple UEs with large model distances can be divided into different groups. In some embodiments, a centralized grouping method can be adopted, in which the server 110 determines the grouping of multiple UEs based on the model distance of the local models of the multiple UEs. In some embodiments, an autonomous grouping method can be adopted, in which the UE itself determines whether the two UEs belong to the same group based on the model distance between the local model of the UE and another UE. As an example and not a limitation, in Figure 1A In the illustrated embodiment, UEs 120-1 and 120-2 may be associated with task 131A, and thus these UEs are grouped into a first group 130A. UEs 120-1, 120-2, 120-3 may be associated with a task 131B different from task 131A, and thus these UEs are grouped into a second group 130B.

[0050] In order to calculate the model difference, a model exchange may be performed between multiple UEs in a group. Group-specific resource allocation may improve the communication efficiency of each UE in each group and / or reduce communication interference with different groups. This facilitates model exchange between UEs in a group. As an example and not limitation, in Figure 1AIn the illustrated embodiment, a first resource set 1400A specific to the group may be allocated to the group 130A, and a second resource set 1400B specific to the group may be allocated to the group 130B. Interference between the first resource set 1400A and the second resource set 1400B may be minimized. The UEs in each group may be configured to communicate with the UEs in the group using the resource set specific to the group.

[0051] Group-specific resources may be determined by a network-side device. In the example of a Wi-Fi system, group-specific resources may be determined by an access point. In the example of a cellular system, an entity that determines the grouping (e.g., a federated learning server residing at an AF) may communicate with a policy control function (PCF) to inform a base station (e.g., gNB) of the grouping result through an authentication management function (AMF). The base station may allocate group-specific resources based on the received grouping result for use by UEs associated with the group. Alternatively, when the grouping of UEs is determined by the base station, the base station may allocate group-specific resources based on the grouping result determined by itself. The base station may include configurations associated with such resource allocations in appropriate signaling to notify one or more UEs in the corresponding group. Preferably, the signaling used may include RRC signaling.

[0052] It should be understood that Figure 1A The embodiments shown in the embodiment are merely exemplary. Each of one or more aspects of the embodiment may be implemented separately or in combination. It should also be understood that although Figure 1A Although described with respect to federated learning, the techniques of the present disclosure may also be applied in other distributed machine learning systems. In some alternative embodiments, system 100A may include more or fewer training devices (e.g., UE 120). In some alternative embodiments, groups of UEs may also be associated with Figure 1A The embodiments shown are different. Additional details regarding the technology of the present disclosure are further described below.

[0053] 2. Exemplary Equipment

[0054] Figure 2 An exemplary block diagram of an electronic device 200 according to some embodiments of the present disclosure is shown. The electronic device 200 may be implemented on the user side or client side of the system. Therefore, the electronic device 200 may be referred to as a user side device. The electronic device 200 may be used to implement or control a UE (e.g., a UE) for performing training in a distributed training system. Figure 1A120 in the embodiment of the present invention). The UE may be implemented as various types of clients. In some embodiments, the UE may be implemented as a client in a server-client architecture. In some embodiments, the UE may be implemented as a user terminal in a cellular network. In some embodiments, the UE may be implemented as a user terminal in a Wi-Fi system. The electronic device 200 may be used to perform one or more operations related to the UE described herein. Specifically, the electronic device 200 may be implemented as the UE itself, as a part of the UE, or as a control device for controlling the UE. For example, the electronic device 200 may be implemented as a chip for controlling the UE. In some embodiments of the present invention, the electronic device 200 is implemented as the UE itself, which is merely for the convenience of description and is not intended to be limiting.

[0055] According to some embodiments of the present disclosure, the electronic device 200 may include a communication unit 210 , a storage unit 220 , and a processing circuit 230 .

[0056] The communication unit 210 of the electronic device 200 may be used to receive or send wired transmissions or radio transmissions. The communication unit 210 may be used to establish and maintain one or more communication links. Each communication link may carry an associated transmission. The one or more communication links may include a communication link between the electronic device 200 and a network side device (e.g., a wireless communication link). Figure 3 The electronic device 300 described includes but is not limited to a base station and / or a federated learning server) and a communication link between the electronic device 200 and the network side device, such as an uplink or a downlink between the electronic device 200 and a network side device. The one or more communication links may also include a communication link between the electronic device 200 and an electronic device associated with other UEs (for example, another electronic device 200 associated with other UEs), such as a side link (for example, a side-link) between multiple UEs. In some embodiments of the present disclosure, the communication unit 210 may perform functions such as up-conversion and digital-to-analog conversion on the transmitted signal and / or perform functions such as down-conversion and analog-to-digital conversion on the received signal. Various technologies may be used to implement the communication unit 210. For example, the communication unit 210 may be implemented as a communication interface component such as an antenna device, a radio frequency circuit, and a partial baseband processing circuit. In Figure 2 , the communication unit 210 is drawn with dashed lines because the communication unit 210 may alternatively be located within the processing circuit 230 or outside the electronic device 200 .

[0057] The storage unit 220 of the electronic device 200 may store information generated by the processing circuit 230, information received from other devices through the communication unit 210 or information to be sent to other devices, programs, machine codes and data for the operation of the electronic device 200, etc. According to some embodiments of the present disclosure, the storage unit 220 may store training data of the electronic device 200, a local model obtained by the electronic device 200, a determined model difference, a global model received by the electronic device 200, resource configuration data and / or other associated data. The storage unit 220 may be a volatile memory and / or a non-volatile memory. For example, the storage unit 220 may include, but is not limited to, a random access memory (RAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), a read-only memory (ROM) and a flash memory. The storage unit 220 is drawn with a dotted line because it may be alternatively located within the processing circuit 230 or outside the electronic device 200.

[0058] The processing circuit 230 of the electronic device 200 may be configured to perform one or more operations, thereby providing various functions of the electronic device 200. The processing circuit 230 may perform corresponding operations by executing one or more executable instructions stored in the storage unit 220. The processing circuit 230 may be used to implement one or more steps performed by the UE in various methods according to some embodiments of the present disclosure.

[0059] According to some embodiments of the present disclosure, the processing circuit 230 may be configured to perform the method for machine learning described herein. Specifically, the processing circuit 230 may be configured to obtain a first artificial intelligence model trained by the electronic device 200, and receive a second artificial intelligence model trained by another electronic device different from the electronic device 200. The processing circuit 230 may be further configured to determine a model difference between the second artificial intelligence model and the first artificial intelligence model. Then, the processing circuit 230 may be configured to send the determined model difference to the federated learning server.

[0060] According to some embodiments of the present disclosure, the processing circuit 230 may include a model training unit 231 and a model comparison unit 232. The model training unit 231 may be configured to train a first artificial intelligence model. The model training unit 231 is drawn with a dotted line because it may be located outside the processing circuit 230 or the electronic device 200 instead. In other words, in some embodiments, model training may not necessarily be a built-in function of the processing circuit 230, but may be implemented by other parts of the electronic device 200 instead. For example, model training may be performed by a peripheral device associated with the electronic device 200 (such as a PC, a neural network computing stick), etc. The model comparison unit 232 may receive a first artificial intelligence model trained by the peripheral device. The processing circuit 230 may receive a second artificial intelligence model through the communication unit 210. The model comparison unit 232 may be configured to determine a model difference between the received second artificial intelligence model and the first artificial intelligence model. The processing circuit 230 may send the determined model difference to the federated learning server through the communication unit 210. Details of the method performed by the processing circuit 230 will be further described below.

[0061] It should be understood that Figure 2 The processing circuit 230 in FIG. 2 is merely exemplary. The processing circuit 230 may include one or more additional units for performing the techniques of the present disclosure.

[0062] Figure 3 An exemplary block diagram of an electronic device 300 according to some embodiments of the present disclosure is shown. The electronic device 300 can be implemented on the network side or the server side of the system. Therefore, the electronic device 300 can be referred to as a network side device, a control side device, or a server side device. In the present disclosure, the network side device, the control side device, or the server side device can collectively refer to one or more devices that are distinguished from the user side device (e.g., UE). Depending on the functions and operations performed, the one or more devices may include a wireless access point, a base station, a core network node, or a federated learning server. For example, in the context of performing actions (e.g., resource allocation) on the radio access network (RAN) side, the one or more devices may include a base station. For another example, in the context of model training, aggregation, and distribution, the one or more devices may include a federated learning server. When discussing an overall solution that includes model training and resource allocation, the one or more devices may include both a base station and a federated learning server as a whole relative to the user side device. In this case, the one or more devices may also include one or more intermediate devices (e.g., core network nodes) between the base station and the federated learning server. The federated learning server may be communicatively coupled to the base station. For example, the federated learning server may reside on the AF and communicate with the base station through one or more interfaces (e.g., Fig.10The user side device may be communicatively coupled to the base station and communicatively coupled to the federated learning server via the base station. In the present disclosure, the link from the user side device to the base station or the federated learning server is referred to as an uplink, and the link from the base station or the federated learning server to the user side device is referred to as a downlink.

[0063] In some embodiments, the electronic device 300 can be used to implement a server (eg, Figure 1A In some embodiments, the electronic device 300 may be connected to the server and communicate with the UE (e.g., Figure 1A In some embodiments, the electronic device 300 may be implemented as a base station (BS). For example, the electronic device 300 may be implemented as the BS itself, as a part of the BS, or as a control device for controlling the BS. The electronic device 300 may be implemented as a chip for controlling the BS. In some embodiments, the electronic device 300 may be implemented as an access point (AP) in a Wi-Fi system. In some embodiments of the present invention, the electronic device 300 is implemented as the BS itself, which is merely for the convenience of description and is not intended to be limiting. In other embodiments, the electronic device 300 may be implemented as other network-side devices, control-side devices, or server-side devices associated with the BS.

[0064] According to some embodiments of the present disclosure, the electronic device 300 may include a communication unit 310 , a storage unit 320 , and a processing circuit 330 .

[0065] The communication unit 310 of the electronic device 300 can be used to receive or send wired transmissions or radio transmissions. The communication unit 310 can be used to establish and maintain one or more communication links. Each communication link can carry an associated transmission. For example, the one or more communication links can be a communication link between the electronic device 300 and the UE (e.g., the electronic device 200), including an uplink and / or a downlink. The communication unit 310 can perform functions such as up-conversion, digital-to-analog conversion on the transmitted signal, and / or perform functions such as down-conversion, analog-to-digital conversion on the received signal. Various technologies can be used to implement the communication unit 310. For example, the communication unit 310 can be implemented as a communication interface component such as an antenna device, a radio frequency circuit, and a partial baseband processing circuit. In Figure 1 , the communication unit 310 is drawn with dashed lines because it may alternatively be located within the processing circuit 330 or outside the electronic device 300 .

[0066] The storage unit 320 of the electronic device 300 may store information generated by the processing circuit 330, information received from other devices through the communication unit 310 or information to be sent to other devices, programs, machine codes and data used for the operation of the electronic device 300, etc. According to some embodiments of the present disclosure, the storage unit 320 may store local models, model differences, aggregated global models, resource configuration data and / or other associated data uploaded by the UE. The storage unit 320 may be a volatile memory and / or a non-volatile memory. For example, the storage unit 320 may include, but is not limited to, a random access memory (RAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), a read-only memory (ROM), and a flash memory. The storage unit 320 is drawn with a dotted line because it may be alternatively located within the processing circuit 330 or outside the electronic device 300.

[0067] The processing circuit 330 of the electronic device 300 may be configured to perform one or more operations, thereby providing various functions of the electronic device 300. For example, the processing circuit 330 may perform corresponding operations by executing one or more executable instructions stored in the storage unit 320. The processing circuit 330 may be used to implement one or more steps performed by a network-side device or a server-side device in various methods according to some embodiments of the present disclosure.

[0068] According to some embodiments of the present disclosure, the processing circuit 330 may perform the method for machine learning described herein. Specifically, the processing circuit 330 may be configured to receive a first artificial intelligence model from a first UE. The first artificial intelligence model is trained by the first UE in a federated learning process. The processing circuit 330 may be configured to receive a model difference from a second UE. The model difference represents the difference between the first artificial intelligence model and the second artificial intelligence model trained by the second UE in the federated learning process. The processing circuit 330 may also be configured to generate a global artificial intelligence model based on aggregating at least the first artificial intelligence model and the model difference.

[0069] According to some embodiments of the present disclosure, the processing circuit 330 may include a model aggregation unit 331. The processing circuit 330 may receive a first artificial intelligence model from the first UE and a model difference from the second UE through the communication unit 310. The model aggregation unit 331 may aggregate the received first artificial intelligence model and the model difference to generate a global artificial intelligence model. Details of the method performed by the processing circuit 330 will be further described below.

[0070] It should be understood that Figure 3 The processing circuit 330 in FIG. 3 is merely exemplary. The processing circuit 330 may include one or more additional units for executing the method of the present disclosure.

[0071] It should be noted that the various units described above are exemplary and / or preferred modules for implementing the processing described in the present disclosure. These modules can be hardware units (such as central processing units, field programmable gate arrays, digital signal processors or application-specific integrated circuits, etc.) and / or software modules (such as computer-readable programs). The modules for implementing the various steps described below are not described in detail above. However, as long as there are steps for performing a certain processing, there can be corresponding modules or units (implemented by hardware and / or software) for implementing the same processing. The technical solutions defined by all combinations of the steps described below and the units corresponding to these steps are included in the disclosure of the present disclosure, as long as the technical solutions they constitute are complete and applicable.

[0072] In addition, the device composed of various units can be incorporated into a hardware device such as a computer as a functional module. In addition to these functional modules, the electronic device can of course have other hardware or software components.

[0073] 3. Exemplary Methods

[0074] Figure 4 A flow chart of a method 4000 for machine learning according to some embodiments of the present disclosure is shown. The method 4000 may be performed on a UE side. For example, the method 4000 may be performed by a first UE. The first UE may be a plurality of UEs (e.g., Figure 1A The first UE may be implemented by the aforementioned electronic device 200. Accordingly, the method 4000 may be executed by the processing circuit 230 of the electronic device 200.

[0075] Method 4000 may start at step 4100. In step 4100, the first UE may be configured to obtain a first artificial intelligence model trained by the first UE. The first artificial intelligence model may be a local model trained by the first UE using local data of the UE during a federated learning process. For example, the first artificial intelligence model may be Figure 1A More specifically, in the case where the first UE is UE 120-1, the first artificial intelligence model may be the local model 121-1.

[0076] The technology of the present disclosure can be applied to various types of artificial intelligence models, including models that have been developed and models that may be developed in the future. Exemplary artificial intelligence models include, but are not limited to, models trained by convolutional neural networks (CNNs), deep neural networks (DNNs), recursive neural networks (RNNs), long short-term memories (LSTMs), generative models, random forests, and their deformation and enhancement algorithms. Each artificial intelligence model can be represented by a set of model parameters that characterize the model. For example, for a CNN model, the set of model parameters that characterizes the model may include, but are not limited to, one or more parameters that describe the number of layers of the neural network, the number and distribution of neurons in each layer, the connection relationship between neurons, the weight of each neuron connection, etc. For other types of models, other types of model parameter sets can be used to characterize them. In the present disclosure, each UE can use various appropriate techniques to train its corresponding local model, including training techniques that have been developed and training techniques that may be developed in the future. During the training process, the UE can use local data as training data to tune one or more parameters in the model parameter set that characterizes the local model. Therefore, the trained local model can be represented by a tuned model parameter set.

[0077] Method 4000 may continue to step 4200. In step 4200, the first UE may be configured to receive a second artificial intelligence model trained by a second UE. The second UE may be another UE different from the first UE that participates in federated learning. The second artificial intelligence model may be a local model that the second UE trains using local data of the UE during the process of federated learning. Receiving the second artificial intelligence model may include receiving a set of model parameters characterizing the model. In some embodiments, the first UE may receive the second artificial intelligence model via a side link between the first UE and the second UE. In some embodiments, the second UE may be a UE that completes the current training round earlier than the first UE among multiple UEs participating in federated learning. Preferably, the second UE may be the UE that completes the current training round the fastest among the multiple UEs. This makes it possible for the first UE to complete the comparison of model parameters, the determination of model differences, and the reporting of model differences within the time limit of the current training round.

[0078] Method 4000 may continue to step 4300. In step 4300, the first UE may be configured to determine a model difference between the second artificial intelligence model and the first artificial intelligence model. The model difference may represent a difference between the two models. According to some embodiments of the present disclosure, the first UE may determine the model difference between the two models based on a first model parameter set representing the first artificial intelligence model and a second model parameter set representing the second artificial intelligence model. As an example and not limitation, Figure 8A schematic diagram of determining model differences according to some embodiments of the present disclosure is shown.

[0079] According to some embodiments of the present disclosure, various metrics can be used to characterize model differences. Exemplary metrics may include one or more of gradients, weights, divergences, distances, eigenvalues, and the like. In some embodiments, the value of each metric characterizing the model difference may be the difference between the values ​​of two corresponding parameters of the first artificial intelligence model and the second artificial intelligence model. For example, in the case where the corresponding parameter is the gradient of the weight of the model, the value characterizing the model difference between the two models may include the difference in the gradient associated with the two models. In some embodiments, the value of each metric may be a distance metric or a similarity metric (such as Euclidean distance, KL divergence, normalized value, etc.) between the two parameters. In some embodiments, the features of model differences can be used to represent model differences to further reduce the amount of data. The features of model differences can be represented by compression or enhancement of model differences, including but not limited to distribution, eigenvalues, pruning of model differences, model difference top-k weights, weights of model differences above a certain threshold, certain layers of model differences, and the like.

[0080] The method 4000 may continue to step 4400. In step 4400, the first UE may be configured to send the model difference determined in step 4300 to the federated learning server. The federated learning server may be, for example, Figure 1A The federated learning server may obtain a corresponding global model based on aggregating the model difference received from the first UE and the second intelligent model received from the second UE.

[0081] In method 4000, the first UE sends the model difference instead of sending the local model trained by the UE (e.g., the first artificial intelligence model). In this way, it is possible to reduce the amount of uploaded data performed by the first UE, thereby reducing the load of the uplink of the first UE.

[0082] According to some embodiments of the present disclosure, method 4000 may include one or more optional or preferred steps.

[0083] In some optional embodiments, the execution of method 4000 may be triggered in response to a specific link quality condition. For example, method 4000 may be executed when the uplink of one or more UEs participating in federated learning is limited. Specifically, when the uplink quality of the first UE is lower than the predetermined link quality condition, the execution of method 4000 may be triggered so that the first UE can send the model difference to the federated learning server. This can achieve the purpose of reducing the load of the uplink of the first UE. When the uplink quality of the first UE is not lower than the predetermined link quality condition, the first UE can directly transmit the first artificial intelligence model itself without calculating and transmitting the model difference. The predetermined link quality condition may be a predetermined suitable condition, including but not limited to a predetermined link bandwidth and the number of available resources. As an example and not a limitation, the predetermined link quality condition may be configured by a server-side device (e.g., a federated learning server) to each UE. For example, in a cellular communication system, the federated learning server may send the configured predetermined link quality condition to each UE through the signaling of the base station. For another example, in a Wi-Fi system, the federated learning server may send the configured predetermined link quality condition to each UE through the AP. In some embodiments, each UE may monitor its own uplink quality. In some embodiments, a network side device (e.g., a base station or an AP) may monitor the uplink quality of each UE. When the uplink quality of one or more UEs is lower than a predetermined link quality condition, the network side device may trigger the execution of method 4000.

[0084] In some optional embodiments, the execution of method 4000 may be triggered based on the expected model transmission time of the first UE. If the expected model transmission time is too long (e.g., exceeds a certain threshold), the first UE may be configured to transmit model differences to save model transmission time. If the expected model transmission time is not too long (e.g., does not exceed the certain threshold), the first UE may be configured to transmit the first artificial intelligence model itself. The expected model transmission time may be evaluated based on the amount of data to be transmitted and the transmission rate of the uplink. As an example and not limitation, the sum of the model training time and the expected model transmission time may be compared with a delay tolerance threshold associated with the model training. If the sum exceeds the delay tolerance threshold, the first UE may be configured to transmit the model difference. Otherwise, the first UE may be configured to transmit the first artificial intelligence model itself. The delay tolerance threshold may be set by the federated learning server. The delay tolerance threshold may be associated with each training round.

[0085] In some optional embodiments, the first UE may send the model difference only when the model difference is small enough. The model distance may be a numerical measure of the model difference. The greater the difference between the two models, the greater the value of the model distance between the two models. When the model distance corresponding to the model difference is less than the predetermined model distance threshold, the amount of data required to send the model difference may be small enough. Therefore, the first UE may send the model difference to the federated learning server when the model distance corresponding to the model difference is less than the predetermined model distance threshold. When the model distance corresponding to the model difference is not less than the predetermined model distance threshold, the benefit of uploading the model difference may be limited. In this case, the first UE may optionally consider a third UE different from the second UE to evaluate the model difference between the first UE and the third UE. The predetermined model distance threshold may be a predetermined appropriate value. As an example and not limitation, the predetermined model distance threshold may be configured by a server-side device (e.g., a federated learning server) to each UE. For example, the federated learning server may send the configured predetermined model distance threshold to each UE through a base station or an AP. In some embodiments, the base station may include the predetermined model distance threshold in various types of signaling sent to the UE, including but not limited to RRC signaling, MAC CE signaling, DCI signaling, and the like.

[0086] Model distance can be calculated based on model differences in various ways. In some embodiments, model differences are represented as a set of differences between one or more parameters corresponding to the first artificial intelligence model and the second artificial intelligence model. Therefore, the model distance can be a weighted average of the differences between the one or more parameters. Alternatively or additionally, the model distance can be a weighted average of a distance metric or a similarity metric (such as Euclidean distance, KL divergence, normalized value, etc.) between the one or more parameters. In other embodiments, any other method can also be used to calculate the model distance.

[0087] In some embodiments, method 4000 may be performed when the first UE and the second UE belong to the same specific group of federated learning. One or more grouping criteria as described above may be adopted, such as a grouping criterion based on the location of the UE, a grouping criterion based on the trust relationship of the UE, or a grouping criterion based on the model distance. Preferably, the individual UEs participating in federated learning may be grouped so that multiple UEs with small model distances are divided into the same group. Therefore, if the first UE and the second UE are divided into the same group, the model distance between the first artificial intelligence model and the second artificial intelligence model is small enough. In this case, method 4000 may be performed, wherein the first UE may perform a model exchange with the second UE and send the model difference to the federated learning server.

[0088] According to an embodiment of the present disclosure, various suitable methods may be used to group the various UEs participating in federated learning. In some embodiments, a centralized grouping method may be used, wherein the grouping is determined by the federated learning server based on the model distance. In other embodiments, an autonomous grouping method may be used, wherein the grouping is determined by each UE (e.g., the first UE or the second UE) based on the model distance, rather than by the federated learning server.

[0089] In some embodiments, when the first UE and the second UE belong to the same group, specific communication resources may be allocated to the first UE and / or the second UE based on the group. These communication resources may be group-specific resources. Group-specific resources may not be used by UEs that do not belong to the group. Group-specific resources may improve the communication efficiency of each UE within each group and / or reduce communication interference with different groups. In some embodiments, resources specific to the group may be allocated to a specific group including the first UE and the second UE, so that the first UE may be configured to communicate with the second UE using the resources. For example, resources in a first frequency band may be allocated to a specific group including the first UE and the second UE, while resources in a second frequency band different from the first frequency band may be allocated to another group different from the specific group. As an example and not a limitation, Fig. 9 A schematic diagram of resource allocation according to some embodiments of the present disclosure is shown.

[0090] In some embodiments, the group-specific resources may be side link resources allocated for the side link between the UEs within the group. For example, in response to the first UE and the second UE belonging to the same group, the first UE and the second UE may receive resource configuration information from a network side device, which may indicate a side link resource specific to the group for use in side link communication between the first UE and the second UE. In some embodiments, these side link resources may form a resource pool specific to the group. In other words, a resource pool of side link resources may be allocated based on a training task of federated learning, so that the same side link resource pool is allocated to multiple UEs (these UEs are located in the same group of federated learning) with the same or similar training tasks. In some embodiments, a side link resource pool may be allocated under mode II associated with a side link resource.

[0091] In some embodiments, each UE participating in federated learning can autonomously form group-specific resources. Specifically, in response to determining that the first UE and the second UE belong to the same group, the first UE can maintain the initial communication channel between the second UE. In response to determining that the first UE and the third UE do not belong to the same group, the first UE or the third UE can deactivate the initial communication channel between them. In subsequent communications, the first UE and the second UE can continue to use the communication resources (e.g., frequency bands) associated with the initial communication channel between them. The third UE can deactivate the communication resources associated with the initial communication channel between the first UE and the third UE, and instead use another communication resource (e.g., an orthogonal communication resource) different from the communication resource. This resource adjustment process can be iterated with iterations of training rounds. In this way, group-specific resources (where each UE uses resources associated with the group to which the UE belongs) can be autonomously formed, thereby reducing the explicit resource configuration performed by the network-side device.

[0092] In some embodiments, method 4000 may additionally include filtering the model differences before sending the model differences to the federated learning server to further reduce the amount of data required to transmit the model differences. For example, one or more values ​​in the model differences may be filtered. As previously described, one or more values ​​in the model differences may characterize the difference or distance between the corresponding parameters of the first artificial intelligence model and the second artificial intelligence model. Each of the one or more values ​​representing the model differences may be compared with a corresponding predetermined transmission threshold. The predetermined transmission threshold may be a non-zero value. If a specific value in the one or more values ​​is greater than the corresponding predetermined transmission threshold, the specific value may be included in the model differences for transmission. If the specific value is not greater than the corresponding predetermined transmission threshold, the specific value may not be included in the model differences. This is because minor aspects of the model differences can be ignored without significantly affecting the performance of the model aggregation. Optionally, a small number of bits (e.g., several 1-bit flags) may be set in the model differences to indicate that certain values ​​are omitted. In this way, the number of values ​​contained in the model differences and the data space they occupy can be reduced, thereby compressing the size of the model differences.

[0093] In some optional embodiments, method 4000 may also include receiving, by the first UE, a global artificial intelligence model from a federated learning server. The global artificial intelligence model may be generated by the federated learning server by at least aggregating a model difference received from the first UE and a second artificial intelligence model received from the second UE.

[0094] In some optional embodiments, method 4000 may also include updating a first artificial intelligence model local to the first UE based on the received global artificial intelligence model. Additionally, the updated first artificial intelligence model may be used by the first UE for the next training round of federated learning. Similarly, the second UE may be configured to update its local second artificial intelligence model based on the received global artificial intelligence model. The updated second artificial intelligence model may be used by the second UE for the next training round of federated learning.

[0095] It should be understood that the above description is merely an exemplary embodiment of the method 4000. Details of various embodiments of the method 4000 as well as additional or alternative embodiments will be further described below.

[0096] Figure 5 A flow chart of a method 5000 for machine learning according to some embodiments of the present disclosure is shown. The method 5000 may be performed on the network side or the server side. For example, the method 5000 may be performed by a server 110 for federated learning, or by a network side device (e.g., a base station) associated with the server 110, or by both of them in collaboration. The device for performing the method 5000 may be implemented by the aforementioned electronic device 300. Accordingly, the method 5000 may be performed by the processing circuit 330 of the electronic device 300.

[0097] Method 5000 may start from step 5100. In step 5100, the network side device may be configured to receive a first artificial intelligence model for a first UE. The first artificial intelligence model may be a local model trained by the first UE in a federated learning process. Receiving the first artificial intelligence model may include receiving a set of model parameters representing the first artificial intelligence model.

[0098] Method 5000 may continue to step 5200. In step 5200, the network-side device may be configured to receive a model difference from the second UE. The model difference may represent a difference between a first artificial intelligence model trained by the first UE and a second artificial intelligence model trained by the second UE during the federated learning process. As described above, the model difference may be characterized using various suitable metrics.

[0099] Optionally, the model difference is received from the second UE when one or more conditions are met. In some embodiments, the model difference is received when the uplink quality of the second UE is less than a predetermined link quality condition. In some embodiments, the model difference is received when the model distance corresponding to the model difference is less than a predetermined model distance threshold. In some embodiments, the model difference is received when the first UE and the second UE belong to the same specific group.

[0100] Method 5000 may continue to step 5300. In step 5300, the network-side device may generate a global artificial intelligence model based on aggregating at least the first artificial intelligence model received from the first UE and the model difference received from the second UE.

[0101] In some embodiments, the network side device may restore the second artificial intelligence model based on the first artificial intelligence model and the model difference between the first artificial intelligence model and the second artificial intelligence model. Then, the network side device may aggregate the first artificial intelligence model and the restored second artificial intelligence model to obtain a global artificial intelligence model. In other embodiments, the network side device may directly aggregate the first artificial intelligence model and the model difference. In addition to the first artificial intelligence model and the associated model difference from the first UE and the second UE, the network side device may additionally aggregate one or more additional artificial intelligence models and associated additional model differences from other UEs participating in federated learning. Various suitable aggregation techniques may be used. For example, the model parameter sets of multiple models (or model differences) may be combined, pruned, weighted summed, and the like to obtain a model parameter set that can characterize a global artificial intelligence model.

[0102] In some embodiments, a specific value in a model difference is received when the specific value is greater than a predetermined transmission threshold. For certain parameters whose values ​​are less than the predetermined transmission threshold, the model difference may not include the corresponding values ​​of these parameters. In this case, the network side device can be configured not to consider the values ​​of these parameters in the model difference when restoring the model or aggregating the model, or to use default values ​​(e.g., zero) as the values ​​of these parameters in the model difference. In some embodiments, the network side device can identify whether the model difference contains the value of the parameter corresponding to the flag bit by identifying the flag bit in the received model difference.

[0103] Additionally, method 5000 may include one or more optional or preferred steps. In some embodiments, method 5000 may include sending the obtained global artificial intelligence model to the first UE and the second UE. In some embodiments, sending the global artificial intelligence model includes sending a set of model parameters characterizing the model. In some embodiments, sending the global artificial intelligence model includes sending an increment of the model relative to the global artificial intelligence model obtained in the previous training round. For example, the increment can be characterized by the gradient of the model parameters of the global artificial intelligence model. The first UE or the second UE may update the first artificial intelligence model and the second artificial intelligence model local to the UE based on the received global artificial intelligence model. The updated first artificial intelligence model and the second artificial intelligence model may be used by the corresponding UE for the next training round of federated learning.

[0104] In some optional embodiments, method 5000 may include grouping the UEs participating in federated learning. The grouping may be determined based on one or more grouping criteria as described above. Preferably, the grouping may be determined based on the model distance between the UEs participating in federated learning. In some embodiments, the grouping may be determined by a network-side device based on the model distance. In other embodiments, the grouping may be determined autonomously by each UE (e.g., a first UE or a second UE) based on the model distance. The network-side device may aggregate the first artificial intelligence model from the first UE and the model difference from the second UE when the first UE and the second UE belong to the same specific group.

[0105] In some optional embodiments, the network side device may be configured to allocate resources specific to each group. In some embodiments, these resources include group-specific side link resources for communication between different UEs in the group. For example, these side link resources may include a group-specific resource pool. The resource pool may be allocated based on the training task of the federated learning process, so different groups with different training tasks may be associated with different resource pools.

[0106] It should be understood that what is described above is merely an exemplary embodiment of method 5000. Details of various embodiments of method 5000 as well as additional or alternative embodiments will be further described below.

[0107] 4. Example process

[0108] Figure 6 The communication process 6000 for machine learning according to some embodiments of the present disclosure is shown. The communication process 6000 adopts a centralized grouping method. In the centralized grouping method, the grouping of each UE participating in federated learning is determined by a network side device or a server side device. For simplicity, Figure 6 Three UEs 120-1, 120-2, 120-3 participating in federated learning are shown. It can be understood that in other embodiments, there may be more UEs acting as training devices without limitation.

[0109] In step 6010, each UE 120-1, 120-2, 120-3 may be configured to perform model training using corresponding local data. Through model training, each UE may obtain a corresponding local model, such as local models 121-1, 121-2, 121-3 (not shown). Local models 121-1, 121-2, 121-3 may be obtained by UE 120-1, 120-2, 120-3 in the same training round of federated learning.

[0110] In step 6020, each UE 120-1, 120-2, 120-3 may be configured to send the trained local models 121-1, 121-2, 121-3 to the network side device (or server side device) 110. Specifically, each UE 120-1, 120-2, 120-3 may be configured to send a model parameter set representing the trained local models 121-1, 121-2, 121-3 to the network side device 110.

[0111] In step 6030, the network side device 110 may be configured to compare the received local models 121-1, 121-2, 121-3, and group the UEs 120-1, 120-2, 120-3 based on the comparison. Specifically, the network side device 110 may be configured to determine the model differences between the local models 121-1, 121-2, 121-3. The network side device 110 may compare the model distance corresponding to the model difference with a predetermined model distance threshold. When the model distance corresponding to the model difference between the two local models is less than the predetermined model distance threshold, the network side device 110 may determine that the two UEs associated with the two local models belong to the same group. When the model distance corresponding to the model difference between the two local models is not less than the predetermined model distance threshold, the network side device 110 may determine that the two UEs associated with the two local models belong to different groups. Figure 6 In the illustrated embodiment, UE 120-1 and UE 120-2 may be classified into a first group, and UE 120-3 may be classified into a second group different from the first group. It should be understood that such grouping results are merely examples and not limitations.

[0112] In some embodiments, the predetermined model distance threshold may be dynamically changed. For example, the predetermined model distance threshold may change with the training rounds of federated learning. The predetermined model distance threshold may be dynamically set by the federated learning server. If the network side device 110 performing the grouping is a federated learning server, the predetermined model distance threshold may be directly applied. If the network side device 110 performing the grouping is a device other than the federated learning server (e.g., a base station or a core network node), the federated learning server may transmit the set predetermined model distance threshold to the device for use in grouping.

[0113] In step 6040, the network side device 110 may be configured to send group information to the UEs 120-1, 120-2, and 120-3. The group information sent to each UE may indicate the corresponding group to which the UE belongs. Additionally, the network side device 110 may be configured to send resource configuration information to the UEs 120-1, 120-2, and 120-3. The resource configuration information may be based on the determined group. Preferably, the resource configuration information may indicate resources specific to the group for each group.

[0114] In step 6050, each UE 120-1, 120-2, 120-3 may perform resource configuration based on the received group information and / or resource configuration information. For example, each UE may identify specific resources allocated to the group to which the UE belongs and use the specific resources for communication. Specifically, UE 120-1 and UE 120-2 may communicate using a first resource set specific to a first group, while UE 120-3 may communicate using a second resource set specific to a second group. The first resource set and the second resource set may be different, for example, associated with different frequency bands and / or different time slots. In some embodiments, the first resource set and the second resource set are sets of side link resources. Side link resources may be used for side link communications between UEs, including transmitting local models and / or model differences between UEs.

[0115] In step 6060, the network side device 110 may perform model aggregation to generate a global model. The model aggregation in this step may be based on the local models 121-1, 121-2, 121-3 from each UE. The network side device 110 may send the obtained global model to each UE 120-1, 120-2, 120-3 in step 6070. It should be understood that in alternative embodiments, steps 6060-6070 may also be performed before steps 6040-6050, or in parallel with steps 6040-6050.

[0116] In step 6080, each UE 120-1, 120-2, 120-3 may be configured to perform model training for the next training round. Each UE may be configured to update the local model of the UE based on the global model received in step 6070. Each UE may be configured to perform the next round of model training on the updated local model. Accordingly, each UE 120-1, 120-2, 120-3 may each obtain a further trained local model 121-1, 121-2, 121-3.

[0117] After grouping, not every UE 120-1, 120-2, 120-3 sends the trained local model to the network device 110. A specific UE in each group is configured to send the local model trained by the UE to the network device 110, and the remaining UEs in the group are configured to send the model difference between the UE and the specific UE. The UE in each group configured to send the local model to the network device 110 can be referred to as the main UE in the group, and the other UEs can be referred to as auxiliary UEs of the group.

[0118] Various suitable criteria may be used to select the master UE of each group from one or more UEs of the group. In some embodiments, the UE that completes the current training round fastest in a training round among the multiple UEs of each group may be selected as the master UE of the group. Because the master UE needs to send the trained local model to other UEs in addition, selecting the UE that completes the current training round fastest can reduce the overall time of the system. In this case, the UE that sends the trained local model to the network side device 110 earliest in each group may be determined as the master UE of the group. For example, UE 120-2 in the first group may complete the model training of step 6010 faster than UE 120-1. Therefore, UE 120-2 may be determined as the master UE of the first group. The master UE of each group may be specified in the group information of step 6040.

[0119] In some embodiments, the UE with the best uplink quality among the multiple UEs in each group can be selected as the master UE of the group. Because the master UE needs to send the trained local model to the network device 110 via its uplink, the uplink requirement is relatively high. In some embodiments, the network device 110 can allocate more uplink resources to the selected master UE in the resource configuration information sent in step 6040.

[0120] exist Figure 6 In the illustrated embodiment, UE 120-2 may be selected as the master UE in the first group. UE 120-2 is configured to send the local model 121-2 trained by the UE to the network device 110 in step 6091. In the second group, UE 120-3 may be the master UE of the group, which is configured to send the local model 121-3 trained by the UE to the network device 110 in step 6093. It should be understood that the selection of the master UE shown is merely exemplary and is not intended to be limiting.

[0121] The remaining UEs 120-1 in the first group may be auxiliary UEs of the group. UE 120-1 may be configured to send model differences to the network side device 110 instead of the local model 121-1. Specifically, UE 120-1 may be configured to receive the local model 121-2 trained by UE 120-2 from UE 120-2 in the same group in step 6092. In some embodiments, the local model may be transmitted between UE 120-1 and 120-2 using side link resources specific to the first group. Then, UE 120-1 may be configured to determine the model difference between the two models based on a comparison of the received local model 121-2 with the local model 121-1 trained by the UE in step 6100. UE 120-1 may be configured to send the determined model difference to the network side device 110 in step 6110. It should be understood that steps 6091, steps 6092-6110, and step 6093 may be performed in any suitable order. For example, step 6091, steps 6092-6110, and step 6093 may be executed in parallel.

[0122] In some embodiments, UE 120-1 may be configured to determine the uplink quality between the UE and the network side device 110, and send the determined model difference based on the uplink quality being lower than a predetermined link quality condition. In some cases, if the uplink quality is so poor (e.g., lower than a second predetermined link quality condition) that the transmission of the model difference cannot even be supported, UE 120-1 may be configured to upload the model difference using other links. As an example and not a limitation, UE 120-1 may transmit the determined model difference to another UE by means of a side link with the other UE. The other UE may act as a relay to upload the model difference to the network side device 110. In this case, the identification information of UE 120-1 may be transmitted in association with the model difference, so that one or more devices receiving the model difference may identify the source of the model difference based on the identification information. In some embodiments, another UE acting as a relay may be in the same group as UE 120-1, because the quality of the side link between multiple UEs in the same group may be better. For example, the other UE may be the main UE in the group, such as UE 120-2.

[0123] In some embodiments, UE 120-1 filters the model difference before sending the model difference to the network side device 110. Specifically, UE 120-1 may compare each of one or more values ​​in the model difference with a corresponding predetermined transmission threshold. If a specific value in the model difference is greater than the corresponding predetermined transmission threshold, the specific value may be included in the model difference for transmission. If the specific value in the model difference is not greater than the corresponding predetermined transmission threshold, the specific value may not be included in the model difference. Optionally, UE 120-1 may set a small number of bits (e.g., several 1-bit flags) in the model difference to indicate that certain values ​​are omitted. In this way, UE 120-1 can further compress the size of the model difference.

[0124] In step 6120, the network side device 110 may perform model aggregation to generate a global model. The model aggregation of this step may be based on the local model 121-2 received in step 6091, the local model 121-3 received in step 6093, and the model difference received in step 6110. In some embodiments, the network side device 110 may restore the local model 121-1 trained by the UE 120-1 based on the local model 121-2 received in step 6091 and the model difference received in step 6110, and then aggregate the local models 121-1, 121-2, and 121-3 to obtain a global model. In some embodiments, the network side device 110 may not need to restore the local model 121-1, but directly aggregate the local model 121-2 received in step 6091, the local model 121-3 received in step 6093, and the model difference received in step 6110.

[0125] In step 6130 , the network-side device 110 may send the global model obtained in step 6120 to each UE 120 - 1 , 120 - 2 , and 120 - 3 .

[0126] Then, the communication process 6000 can repeatedly perform steps 6080-6120 until the end condition of the federated learning is met. In some embodiments, the end condition can include reaching a specified number of training rounds. In some embodiments, the end condition can include the resulting global model reaching a predetermined performance condition.

[0127] It should be understood that what is described above is merely an exemplary embodiment of process 6000. In some embodiments, one or more steps in process 6000 are optional and not required. In some embodiments, process 6000 may include one or more optional steps not shown. In some embodiments, the members of the group may change. For example, if in step 6092, UE 120-1 does not receive a local model from UE 120-2 within a specified time threshold, the grouping between UE 120-1 and UE 120-2 may be released. In this case, UE 120-1 may send its trained local model 121-1 to network side device 110 instead of the model difference.

[0128] Figure 7 Another communication process 7000 for machine learning according to some embodiments of the present disclosure is shown. The communication process 7000 adopts an autonomous grouping method. In the autonomous grouping method, the groups of these UEs are autonomously determined by each UE participating in federated learning. For simplicity, Figure 7 Three UEs 120-1, 120-2, 120-3 participating in federated learning are shown. It can be understood that in other embodiments, there may be more UEs acting as training devices without limitation.

[0129] In step 7010, each UE may be configured to perform model training using corresponding local data. Through model training, each UE may obtain a corresponding local model. For example, UE 120-1, 120-2, 120-3 may obtain local models 121-1, 121-2, 121-3 (not shown), respectively. Local models 121-1, 121-2, 121-3 may be obtained by UE 120-1, 120-2, 120-3 in the same training round of federated learning.

[0130] In step 7020, a collaboration request may be sent by a specific UE to one or more other UEs. The collaboration request may indicate that the specific UE expects to perform federated learning in collaboration with a suitable UE in the one or more other UEs. The collaboration request may be issued when one or more specific conditions are met. In some embodiments, when the specific UE quickly completes the current training round, the specific UE may issue a collaboration request. In some embodiments, when the uplink quality of the specific UE is lower than a predetermined link quality condition, the specific UE may issue a collaboration request. The destination of the collaboration request may be other UEs located near the specific UE (e.g., within a certain distance threshold range). For example, the UE may broadcast a message on a physical side link discovery channel (PSDCH) to discover other nearby UEs participating in federated learning. Specifically, if the UE participates in federated learning, the UE may send a collaboration request on the PSDCH to find other terminals participating in federated learning. After the UEs discover each other, a PC5 link may be established between the UEs. The UEs may exchange each other's RRC configurations. In the RRC configuration, the UEs may interact with resource pools configured specifically for artificial intelligence models. At least a portion of the communication resources in the resource pool may be used by the UE to exchange model parameters of the local model on a physical sidelink shared channel (PSSCH), as described below. Figure 7 In the example of FIG. 1 , UE 120 - 3 may issue a cooperation request to the other two UEs 120 - 1 and 120 - 2 .

[0131] In response to receiving the collaboration request from UE 120-3, UEs 120-1 and 120-2 may be configured to send the trained local models 121-1 and 121-2, respectively, to UE 120-3 in step 7030. Sending the local model may include sending a set of model parameters characterizing the local model. In some embodiments, the corresponding local model may be sent via a side link between UEs 120-1 and 120-2 and UE 120-3.

[0132] In step 7040, UE 120-3 may be configured to compare the received local models 121-1 and 121-2 with the local model 121-3 trained by UE 120-3, respectively. UE 120-3 may determine whether the UE and UE 120-1, 120-2 belong to the same group based on the comparison. Specifically, UE 120-3 may be configured to determine the model difference between local models 121-1, 121-2 and local model 121-3. UE 120-3 may compare the model distance corresponding to the model difference with a predetermined model distance threshold. When the model distance corresponding to the model difference between the two local models is less than the predetermined model distance threshold, UE 120-3 may determine that the two UEs associated with the two local models belong to the same group. When the model distance corresponding to the model difference between the two local models is not less than the predetermined model distance threshold, UE 120-3 may determine that the two UEs associated with the two local models belong to different groups. Figure 7 In the illustrated embodiment, UE 120-3 may determine that UE 120-3 and UE 120-2 may belong to the same group, and accordingly send a cooperation confirmation message to UE 120-1 in step 7051. UE 120-3 may also determine that UE 120-3 and UE 120-1 do not belong to the same group, and accordingly send a cooperation rejection message to UE 120-1 in step 7052. It should be understood that such grouping results are merely examples and not limitations.

[0133] In some embodiments, the predetermined model distance threshold may be dynamically changed. For example, the predetermined model distance threshold may change with the training rounds of federated learning. The predetermined model distance threshold may be dynamically set by the federated learning server. The set predetermined model distance threshold may be transmitted from the federated learning server to the associated one or more UEs via the base station.

[0134] After grouping, each UE may upload a model difference or a local model based on the grouping result. Specifically, UE 120-3 may be configured to send a model difference. The model difference may be a model difference between local models 121-3 and 121-2 determined in step 7040. In some implementations, UE 120-3 may upload the model difference via an uplink between the UE and the network side device 110.

[0135] In some embodiments, if the quality of the uplink between UE 120-3 and the network side device 110 is so poor (e.g., lower than the second predetermined link quality condition) that it cannot even support the transmission of the model difference, UE 120-3 may be configured to use other links to upload the model difference. As an example and not a limitation, UE 120-3 may transmit the determined model difference to another UE by means of a side link with the other UE. The other UE may act as a relay to upload the model difference to the network side device 110. In this case, the identification information of UE 120-3 may be transmitted in association with the model difference, so that one or more devices receiving the model difference may identify the source of the model difference based on the identification information. In some embodiments, another UE acting as a relay may be in the same group as UE 120-3, because the quality of the side link between multiple UEs in the same group may be better. For example, the other UE may be the main UE in the group, such as UE 120-2.

[0136] In some embodiments, UE 120-3 filters the model difference before sending the model difference to the network side device 110. Specifically, UE 120-3 may compare each of one or more values ​​in the model difference with a corresponding predetermined transmission threshold. If a specific value in the model difference is greater than the corresponding predetermined transmission threshold, the specific value may be included in the model difference for transmission. If the specific value in the model difference is not greater than the corresponding predetermined transmission threshold, the specific value may not be included in the model difference. Optionally, UE 120-3 may set a small number of bits (e.g., several 1-bit flags) in the model difference to indicate that certain values ​​are omitted. In this way, UE 120-3 can further compress the size of the model difference.

[0137] In the case where UE 120-3 sends a model difference, UE 120-2 belonging to the same group as UE 120-3 may be configured to send its trained local model 121-2 in step 7062. UE 120-1 not belonging to the same group as UE 120-3 may be configured to send its trained local model 121-1 in step 7063. It should be understood that steps 7061-7063 may be performed in any suitable order. For example, steps 7061-7063 may be performed in parallel.

[0138] In step 7070, the network side device 110 may perform model aggregation to generate a global model. The network side device 110 may perform model aggregation to generate a global model. The model aggregation of this step may be based on the model difference received in step 7061, the local model 121-2 received in step 7062, and the local model 121-1 received in step 7063. In some embodiments, the network side device 110 may restore the local model 121-3 trained by the UE 120-3 based on the model difference received in step 7061 and the local model 121-2 received in step 7062, and then aggregate the local models 121-1, 121-2, and 121-3 to obtain a global model. In some embodiments, the network side device 110 may not need to restore the local model 121-1, but directly aggregate the model difference received in step 7061, the local model 121-2 received in step 7062, and the local model 121-1 received in step 7063.

[0139] In step 7080 , the network-side device 110 may send the global model obtained in step 7070 to each UE 120 - 1 , 120 - 2 , and 120 - 3 .

[0140] In step 7090, each UE 120-1, 120-2, 120-3 may perform resource configuration based on the grouping determined in step 7040, so that each group has resources specific to the grouping. For example, UE 120-3 and UE 120-2 may communicate using a first resource set specific to a first grouping, while UE 120-1 may communicate using a second resource set specific to a second grouping. The first resource set and the second resource set may be different, for example, associated with different frequency bands and / or different time slots. In some embodiments, the first resource set and the second resource set are sets of sidelink resources. The sidelink resources may be used for sidelink communications between UEs, including transmitting local models between UEs.

[0141] In some embodiments, the UE may perform resource configuration in response to a sent / received collaboration confirmation message or collaboration rejection message. For example, in response to determining that UE 120-3 and 120-2 belong to the same group, the initial communication channel between UE 120-2 and 120-3 may be maintained. Maintaining the initial communication channel may include fixedly using communication resources associated with the initial communication channel. The initial communication channel is, for example, the channel used by UE 120-3 to send a collaboration request to UE 120-2 in step 7020, or the channel used by UE 120-3 to receive a local model from UE 120-2 in step 7030. Specifically, in subsequent communications (e.g., inter-UE communications), UE 120-2 and 120-3 may continue to use the channel for communication. In response to determining that UE 120-3 and 120-1 do not belong to the same group, UE 120-1 may deactivate the initial communication channel between UE 120-1 and 120-3. The initial communication channel is, for example, the channel used by UE 120-3 to send a collaboration request to UE 120-1 in step 7020, or the channel used by UE 120-3 to receive a local model from UE 120-1 in step 7030. Specifically, in subsequent communications (e.g., UE-to-UE communications), UE 120-1 may avoid using this channel for communication. Alternatively, UE 120-1 uses another channel for UE-to-UE communications. The other channel may be orthogonal to the initial communication channel between UE 120-1 and 120-3. For example, the other channel may be located in a different frequency band and / or time slot from the initial communication channel. In this way, multiple UEs belonging to the same group can be autonomously associated with the same communication resource, while multiple UEs not belonging to the same group can be autonomously associated with different communication resources.

[0142] In some embodiments, the UE may perform resource configuration based on the resource configuration information from the network device 110. This may be similar to step 6050 of process 6000. Specifically, the UE 120-3 may report the group determined in step 7040 to the network device 110. The network device 110 may determine the resource configuration information of each group based on the received group information, and send the determined resource configuration information to each UE.

[0143] It should be understood that step 7090 and steps 7061-7080 may be performed in any suitable order. For example, step 7090 and steps 7061-7080 may be performed in parallel.

[0144] In step 7100, each UE 120-1, 120-2, 120-3 may be configured to perform model training for the next training round. Each UE may be configured to update the local model of the UE based on the global model received in step 7080. Each UE may be configured to perform the next round of model training on the updated local model. Accordingly, each UE 120-1, 120-2, 120-3 may each obtain a further trained local model 121-1, 121-2, 121-3.

[0145] Next, not every UE 120-1, 120-2, 120-3 sends the trained local model to the network device 110. A specific UE in each group is configured to send the local model trained by the UE to the network device 110, and the remaining UEs in the group are configured to send the model difference between the UE and the specific UE. The UE in each group configured to send the local model to the network device 110 can be referred to as the main UE in the group, and the other UEs can be referred to as auxiliary UEs of the group.

[0146] exist Figure 7 In the illustrated embodiment, UE 120-2 may be selected as the master UE in the first group. UE 120-2 is configured to send the local model 121-2 trained by the UE to the network device 110 in step 7112. In the second group, UE 120-1 may be the master UE of the group, which is configured to send the local model 121-1 trained by the UE to the network device 110 in step 7111. It should be understood that the selection of the master UE shown is merely exemplary and is not intended to be limiting.

[0147] The remaining UEs 120-3 in the first group may be auxiliary UEs of the group. The UEs 120-3 may be configured to send model differences to the network side device 110 instead of the local model 121-3.

[0148] Specifically, UE 120-3 may be configured to receive the local model 121-2 trained by UE 120-2 from UE 120-2 in the same group in step 7113. In some embodiments, the local model may be transmitted between UE 120-3 and 120-2 using a side link resource specific to the first group. For example, UE 120-2 may broadcast the local model 121-2 using a side link resource specific to the first group. Because UEs in different groups communicate on different sets of resources, the local model 121-2 broadcast by UE 120-2 can be received by UE 120-3 in the same group, but is unlikely to be received by UE 120-1 in a different group. Similarly, UE 120-3 is unlikely to receive the local model 121-1 broadcast by UE 120-1. In some embodiments, the transmission of the local model may also be in a multicast or unicast manner.

[0149] In step 7120 , UE 120 - 3 may be configured to determine the model difference between the received local model 121 - 2 and the local model 121 - 3 trained by the UE based on the comparison of the two models. UE 120 - 3 may be configured to send the determined model difference to network side device 110 in step 7130 .

[0150] It should be understood that step 7111, step 7112, and steps 7113-7130 may be performed in any suitable order. For example, step 7111, step 7112, and steps 7113-7130 may be performed in parallel.

[0151] In step 7140, the network side device 110 may perform model aggregation to generate a global model. The network side device 110 may perform model aggregation to generate a global model. The model aggregation of this step may be based on the model difference received in step 7130, the local model 121-2 received in step 7112, and the local model 121-1 received in step 7111.

[0152] The network-side device 110 may send the obtained global model to each UE 120-1, 120-2, 120-3 in step 7140. Then, the communication process 7000 may repeatedly perform steps 7100-7140 until the end condition of the federated learning is met.

[0153] It should be understood that the above description is merely an exemplary embodiment of process 7000. In some embodiments, one or more steps in process 7000 are optional and not required. In some embodiments, process 7000 may include one or more optional steps not shown.

[0154] In some embodiments, for each training round of federated learning, the master UE of each group may be different. In some embodiments, the UE that completes the current training round fastest in a training round among multiple UEs in each group may be selected as the master UE of the group. In some embodiments, the UE with the best uplink quality among multiple UEs in each group may be selected as the master UE of the group.

[0155] In some embodiments, the members of the group may change. For example, if in step 7113, UE 120-3 does not receive the local model from UE 120-2 within a specified time threshold, the group between UE 120-3 and UE 120-2 may be released. In this case, UE 120-3 may send its trained local model 121-3 to network side device 110 instead of the model difference.

[0156] In some alternative embodiments, there may be a grouping process different from the communication process 7000. For example, instead of sending a cooperation request in step 7020, UE 120-3 sends the local model 121-3 trained by the UE to UE 120-1 and UE 120-2. In this case, UE 120-1 and UE 120-2 may each determine whether they belong to the same group as UE 120-3 based on model comparison. For example, UE 120-1 may compare the local model 121-1 trained by the UE with the local model 121-3 received from UE 120-3. In response to the model distance between the local model 121-1 and the local model 121-3 being not less than a predetermined distance threshold, UE 120-1 may send a cooperation rejection message to UE 120-3. Similarly, UE 120-2 may compare the local model 121-2 trained by the UE with the local model 121-3 received from UE 120-3. In response to the model distance between local model 121-2 and local model 121-3 being less than a predetermined distance threshold, UE 120-2 may send a cooperation confirmation message to UE 120-3.

[0157] Figure 8 Schematic diagram 8000 of determining model differences according to some embodiments of the present disclosure is shown. Figure 8In the example, the first artificial intelligence model 8100 is represented as a plurality of nodes, which are organized into a plurality of layers, and there are one or more edges between the nodes, each edge being associated with a corresponding weight coefficient w. For example, the weight coefficient w may represent the weight of the edge in the first artificial intelligence model 1210-N. The second artificial intelligence model 1210-M may also be represented similarly, but with a different weight coefficient w'. The model difference 8100 between the first artificial intelligence model 1210-N and the second artificial intelligence model 1210-M may be characterized using the difference dw between the weight coefficients w' and w (i.e., dw=w'-w).

[0158] In some embodiments, the UE may send the model difference 8100 to the federated learning server by sending a determined set of dw. In other embodiments, the model difference 8100 may be sent to the federated learning server by sending a representation of the set of dw. (1) the distribution or interleaved sequence of dw; or (2) the eigenvalue of the matrix composed of dw may be used as a representation of dw. This may further reduce the amount of data transmitted. As an example and not a limitation, the distribution of dw may be obtained by the following method: all weight coefficients of the model difference 8100 are represented in a coordinate system in a natural order (index, weight); the represented weight coefficients are rearranged through an interleaved sequence (this adjusts the order of the weight coefficients, making curve fitting easier); after obtaining the fitting curve, one or more distributions are used to characterize the fitting curve. Then, the parameters of the one or more distributions and the interleaved sequence used (if any) may be sent to the federated learning server. By sending the parameters of the one or more distributions and the interleaved sequence used instead of all weight coefficients of the model difference 8100, the communication overhead may be further reduced. In some embodiments, the federated learning server generates a model difference based on the received distribution, performs a deinterleaving operation according to the received interleaving sequence, and then superimposes it on the second artificial intelligence model 1210-M based on the model difference, thereby obtaining the restored first artificial intelligence model 1210-N.

[0159] It should be understood that Figure 8 What is shown is only an exemplary embodiment. In other embodiments, the model difference 8100 may be calculated and represented in other ways. For example, Figure 8 The process uses all the weight coefficients of the model differences to characterize the model differences, while in alternative embodiments the model differences may also be characterized based on only a portion of the weight coefficients. The partial weight coefficients may include, for example, partial weight coefficients obtained after pruning the model, or top-k weight coefficients of the model, or weight coefficients of the model above a certain threshold, or weight coefficients of certain layers of the model, etc. Any other appropriate way may be used to characterize the model differences.

[0160] Fig. 9 A schematic diagram 9000 of resource allocation according to some embodiments of the present disclosure is shown. The schematic diagram is applicable to a Wi-Fi system, in which the frequency band of the Wi-Fi is divided into different frequency bands to be allocated to different channels. In the Wi-Fi system, the bandwidth of the frequency bands may be different. Fig. 9 In the figure, the horizontal axis may represent the frequency, and the vertical axis may represent the bandwidth of each frequency band. When a bandwidth of 160 MHz is used, two non-overlapping frequency bands B1 and B2 may be divided. When a bandwidth of 80 MHz is used, four non-overlapping frequency bands BB1, BB2, BB3, and BB4 may be divided. When a bandwidth of 40 MHz is used, eight non-overlapping frequency bands BBB1, BBB2, BBB3, BBB4, BBB5, BBB6, BBB7, and BBB8 may be divided. When a bandwidth of 20 MHz is used, 16 non-overlapping frequency bands may be divided. In some embodiments, in a Wi-Fi system, a specific UE may broadcast a set of model parameters of a local model of the specific UE on a Wi-Fi channel allocated to the specific UE after completing local model training. After monitoring the set of model parameters, another UE may decide whether to change the Wi-Fi channel used by the other UE based on the model distance between the local model trained by the UE and the received local model. Specifically, if the model distance is less than a predetermined model distance threshold, the other UE may use the same Wi-Fi channel as the specific UE. If the model distance is not less than the predetermined model distance threshold, the other UE may use a different Wi-Fi channel than the specific UE. In this way, multiple UEs with similar local models may be divided into the same Wi-Fi channel, and these UEs may be regarded as one group. Different resource sets of different groups may be associated with different non-overlapping frequency bands in the above frequency bands. In one example, Figure 1A The first resource set 1400A may be associated with frequency band B1. UEs 120-1 and 120-2 in group 130A may communicate on frequency band B1. The second resource set 1400B may be associated with a selected frequency band from B2, BB3-BB4, BBB5-BBB8, or any other available frequency band that does not overlap with B1. UEs 120-3, 120-4, and 120-5 in group 130B may communicate on the selected frequency band. In another example, Figure 1A The first resource set 1400A may be associated with frequency band BB1, and the second resource set 1400B may be associated with any of B2, BB2, BB3-BB4, BBB5-BBB8 or any other available frequency band that does not overlap with BB1. It should be understood that these examples are illustrative. In other embodiments, other methods may be used based on Fig. 9The first resource set 1400A and the second resource set 1400B are allocated by dividing the frequency band. In addition, federated learning can have more than two resource sets, which are distributed in Fig. 9 On non-overlapping frequency bands in the various frequency bands shown.

[0161] Fig.10 A block diagram of a cellular system 1000 in which techniques according to some embodiments of the present disclosure may be implemented is shown. System 1000 may include a service-based architecture. The architecture may include one or more network element nodes, such as a network exposure function (NEF), a network storage function (NRF), a unified data management function (UDM), a policy control function (PCF), a network slice selection function (NSSF), an authentication service function (AUSF), an authentication management function (AMF), and a session management function (SMF). In system 1000, various types of UEs (e.g., Figure 1A In some embodiments, the federated learning server (e.g., Figure 1A The server 110 in the example may be deployed at the application function (AF). Each UE may access the data network (DN) through the gNB and the user plane function (UPF) and link to the AF. Each UE may send a local model and / or model difference to the federated learning server through the link. The federated learning server may group each UE based on the model distance. In some embodiments, the federated learning server may send the grouping result via the PCF. Then, the management policy of the UE side link, such as the resource configuration of the UE of each group, may be determined based on the grouping result. The management policy of the UE side link based on the grouping result may be sent to the gNB through the AMF. The gNB may change the use of the side link resource pool in mode II (e.g., autonomous mode) associated with the side link resource so that a side link resource specific to the group is allocated to each group. In other embodiments, the control mechanism of mode I (e.g., in centralized mode) may be used to directly provide side link resources between multiple UEs belonging to the same group. For example, the base station may divide the side link resource pool into multiple subsets, each of which may be allocated to a corresponding group. Typically, the side link resource pool may be allocated based on a geographic location area. According to an embodiment of the present disclosure, the side link resource pool allocated based on the geographic location area can be subdivided into one or more group-specific side link resource subsets according to the UE group. The base station can include the resource allocation result in appropriate signaling (e.g., RRC signaling) to notify one or more UEs in the corresponding group. In an alternative embodiment, the federated learning server can be located at other locations in the system 1000. Accordingly, the management policy of the UE side link can be issued by any other suitable means.

[0162] Fig.11 An exemplary process 1100 for selecting member UEs participating in federated learning according to some embodiments of the present disclosure is shown. In some scenarios, the process 1100 can be performed in a cellular system (e.g., Fig.10 is implemented in a cellular system 1000).

[0163] Process 1100 may start from step 1101. In this step, the AF may be configured to send a request to the NEF. The request may be a subscription request for subscribing the NEF to a member selection assistance functionality. For example, the AF may send the subscription request by sending a Nnef_UEMemberSelectionAssistance_subscribe request to the NEF. In other examples, the AF may also send the subscription request in other ways. The request may include a list of target UEs and one or more UE member screening criteria. Optionally, the request may also include one or more time windows.

[0164] According to some embodiments of the present disclosure, one or more UE member screening criteria included in the request of step 1101 may include one or more of the optional example criteria shown in Table 1. These optional example criteria include, but are not limited to, QoS, access type and / or RAT type of PDU session, end-to-end data volume transmission time, UE current location, UE historical location, UE direction, etc. Preferably, a specific optional example standard may include a standard associated with a UE side link. The UE side link standard may include which UEs can establish or have established a side link. Optionally, the UE side link standard may also include the quality of the side link. In some embodiments, the UE side link standard may be used to screen member UEs selected to participate in federated learning. In some embodiments, the UE side link standard may also be used to group member UEs. It should be understood that the optional example standards shown in Table 1 are merely examples and not limitations. In other embodiments, other optional example standards may be used.

[0165] Table 1 Example screening criteria

[0166]

[0167] Process 1100 may continue to step 1102. In this step, the NEF may be configured to verify the request from the AF and map one or more UE member screening criteria contained in the request to corresponding service operations. Specifically, the NEF may verify the authorization of the AF request, identify the information that needs to be collected, and perform corresponding service operations based on the UE member screening criteria provided by the AF. Table 1 also shows example service operations corresponding to optional example criteria. It should be understood that the service operations shown are merely examples and not limitations. In other embodiments, each optional example criterion may correspond to a different service operation.

[0168] Process 1100 may continue to step 1103. In this step, the NEF may be configured to interact with one or more 5G core network functions (5GC NFs) to collect the required information. These interactions may be performed through corresponding service operations. The one or more 5GC NFs with which the NEF interacts may be determined based at least in part on one or more UE member screening criteria indicated in the request of step 1101. The one or more 5GC network functions may span cellular systems (such as Fig.10 The collected information may include those identified by the NEF in step 1102. The interaction between the NEF and one or more 5GC NFs may be based at least in part on one or more UE member screening criteria indicated in the request of step 1101.

[0169] Process 1100 may continue to step 1104. In this step, the NEF may be configured to integrate the collected information and derive a list of candidate UEs. For example, based at least in part on the information collected from other 5GC NFs in step 1103, the NEF may integrate all information collected from other 5GC NFs to derive a list of candidate UEs that meet the UE member screening criteria indicated in the AF request of step 1101. The list of candidate UEs may include one or more candidate UEs selected to participate in federated learning. The one or more candidate UEs may be selected from a list of target UEs. For example, based on the collected information, UEs (if any) in the list of target UEs that do not meet the UE member screening criteria may be excluded and thus not included in the list of candidate UEs. In some embodiments, the list of candidate UEs may also be grouped based on the collected information.

[0170] Process 1100 may continue to step 1105. In this step, the NEF may be configured to notify the AF of the list of candidate UEs. As an example, the NEF may notify the AF of the list of candidate UEs by sending Nnef_UEMemberSelectionAssistance_Notifyrequest to the AF. In other examples, the NEF may also notify the AF of the list of candidate UEs in other ways. Preferably, optional additional information may be sent to the AF together with the list of candidate UEs for one or more purposes. In some embodiments, the additional information may include side link information. The side link information may indicate which UEs in the list of candidate UEs have established or can establish side links. Optionally, the side link information may also indicate the quality of the side link between the UEs. The AF may group the UEs based at least in part on the side link information, as described above. In some embodiments, the additional information may include UE location information. The UE location information may indicate the location of each UE in the list of candidate UEs, the distance between each UE, or the distance of each UE from a specific geographical location. The AF may group the UEs based at least in part on the UE location information, as described above. Additionally or alternatively, the sidelink information may also include one or more additional information that can be used to facilitate the AF to group the UE. These additional information may be used in conjunction with other information when grouping the UE. In some embodiments, the grouping may be performed by a federated learning server residing on or connected to the AF. In an alternative embodiment, the grouping is not performed by the AF, but by the NEF. In this case, the NEF may group the candidate UEs based on the collected information (e.g., sidelink information) in step 1104. The result of the grouping may be sent to the AF as part of a list of candidate UEs or as additional information in step 1105. For example, the NEF may send a list of one or more candidate UEs to the AF, each list corresponding to a UE group.

[0171] Fig.12 An exemplary process 1200 for federated learning according to some embodiments of the present disclosure is described. The process 1200 may be implemented in a cellular system. Fig.12 , a network side device NW is shown, which can be implemented as the network side device 110 described above. In addition, four UEs (UE1-UE4) are shown, each of which can be implemented as the UE 120 described above. Those skilled in the art will appreciate that other embodiments may include a greater or lesser number of UEs. In some embodiments, the federated learning server may reside on or be connected to the NW. In other words, the NW may interface with the UE that performs model training on behalf of the federated learning server.

[0172] Process 1200 may start from step 1201. In this step, the NW may be configured to select a training UE for performing model training. In an optional embodiment, step 1201 may include sub-steps 1201-1 and 1201-2. In sub-step 1201-1, UE1-UE4 may be configured to report UE resources to the NW. In sub-step 1201-2, the NW may be configured to select a training UE to be used for performing model training from candidate UE1-UE4. The selection may be based at least in part on the reported UE resources. In this example, UE1-UE4 are all selected as training UEs.

[0173] In step 1202, the NW may be configured to distribute a global model to selected training UEs (e.g., UE1-UE4). The global model may be broadcast or multicast to the selected training UEs via a base station of a cellular network (e.g., a 5G network). If the current training round is the first training round of federated learning (or the initial training round), the distributed global model may be an initial global model, as described in step 1202. If the current training round occurs after the first training round of federated learning, the distributed global model may be an aggregated global model, as described in step 1206. The aggregated global model may be obtained based on the aggregation of the training results of each UE.

[0174] In response to receiving the global model distributed by the NW, each training UE (eg, UE1-UE4) may train the global model (or a portion of the global model) using the local data of the UE in step 1203. Through model training, each training UE may obtain a local model associated with the UE.

[0175] In step 1204, training result reporting may be performed. Specifically, each training UE (e.g., UE1-UE4) may send the model parameters of the trained local model to the NW in sub-step 1204-1. In some embodiments, the model parameters may be sent in the form of gradients. The gradient may represent the difference between the trained local model and the model before training. For example, for one or more model parameters describing weights, each training UE (e.g., UE1-UE4) may send the gradient of the weights to the NW, which helps to reduce the amount of data sent. In step 1204-2, the NW may aggregate the training results received in step 1204-1 to generate an aggregated global model.

[0176] Then, the next training round of federated learning can be performed. Process 1200 can continue to step 1205. In this step, the NW can be configured to select a training UE for performing model training. Similar to sub-step 1201-1, the UE can be configured to report UE resources to the NW in sub-step 1205-1. In the example shown, UE1, UE2, and UE4 reported UE resources to the NW, while UE3 did not report (this may be because UE3 decided to exit federated learning). In sub-step 1205-2, as an example, the NW selects UE1, UE2, and UE4 as training UEs for performing model training.

[0177] According to some embodiments of the present disclosure, the NW may additionally group the selected training UEs. For example, based on the training results (e.g., local models) reported by each training UE in step 1204-1, the NW may determine the model distance between the local models of each training UE. The NW may at least apply the grouping criteria based on model distance as described above. For example, the NW may divide multiple UEs whose model distance is less than a predetermined model distance threshold into the same group. As an alternative or additional embodiment, the NW may also apply one or more other grouping criteria.

[0178] In step 1206, the NW may be configured to distribute the global model to the selected training UEs (e.g., UE1, UE2, UE4). Since the current training round occurs after the first training round of federated learning, the distributed global model may be an aggregated global model, such as the aggregated global model generated in sub-step 1204-2.

[0179] According to some embodiments of the present disclosure, in addition to sending the aggregated global model, the NW may also send the grouping result of the UE to the selected training UE. In the example shown, the grouping result may indicate that UE2 and UE4 are divided into the same group, while UE1 is divided into another group. In other embodiments, other grouping results are possible without limitation.

[0180] In step 1207, each selected training UE may perform model training. For the received aggregated global model, each training UE may further train the global model (or a portion of the global model) using the local data of the UE. Through model training, each training UE may obtain a local model that is potentially superior to the corresponding local model trained in step 1203.

[0181] In step 1208, training result reporting may be performed. Since the training UEs (e.g., UE1, UE2, and UE4) participating in the federated learning have been grouped, the operation in step 1208 is different from the operation in step 1204. Specifically, only a specific UE in each group is configured to report a local model to the NW, while other UEs in the group are configured to report the model difference between the UE and the specific UE to the NW.

[0182] exist Fig.12 In the example of , for the group consisting of UE2 and UE4, UE2 may be selected as a specific UE of the group. For example, UE2 may be selected as a specific UE of the group based on the fact that UE2 completes the model training of step 1207 faster than UE4. Accordingly, UE2 may be configured to send the local model trained by the UE in step 1207 to the NW in sub-step 1208-1. UE2 may also be configured to report the local model trained by UE2 in step 1207 to other UEs (e.g., UE4) in the same group in sub-step 1208-2. In some embodiments, the local model sent by UE2 may be represented by a gradient of a model parameter, such as a gradient of a weight. In some embodiments, only significant gradients (e.g., values ​​exceeding a specified threshold) may be sent, while insignificant gradients may be ignored. In sub-step 1208-3, UE4 may be configured to send a model difference to the NW. The model difference describes the difference between the local model trained by UE4 in step 1207 and the local model trained by UE2 in step 1207. The model difference may be determined by UE4 based on the model received in sub-step 1208-2 and the local model of UE4. In some embodiments, the model difference sent by UE4 to NW may be a difference in the gradient of the weight.

[0183] For the group to which UE1 belongs, since the group only includes a single UE1, UE1 may be a specific UE of the group. Accordingly, UE1 may be configured to send the local model trained by the UE in step 1207 to the NW in sub-step 1208-1.

[0184] In sub-step 1208-4, the NW may be configured to perform model aggregation based at least in part on the local models and model differences received from each UE, thereby obtaining an aggregated global model. For example, the NW may aggregate the local models (e.g., gradients of weights) reported by UE1 and UE2 and the model differences (e.g., gradient differences) reported by UE4. The aggregated global model may be distributed to the selected one or more training UEs in the next training round (not shown).

[0185] Process 1200 may repeatedly perform one or more training rounds until an end condition is reached. In some embodiments, the end condition may include reaching a specified number of training rounds. In some embodiments, the end condition may include the resulting global model reaching a predetermined performance condition. In response to reaching the end condition, the NW may provide the latest aggregated global model to one or more users using the model.

[0186] It should be understood that the above description is merely an exemplary embodiment of process 1200. In some embodiments, one or more steps in process 1200 are optional and not required. For example, when the change in one or more conditions associated with the UE is not large, the step of selecting a training UE (e.g., step 1205) may be omitted for one or more training rounds. In some embodiments, process 1200 may include one or more optional steps not shown. For example, the step of selecting a training UE may include selecting a training UE. Fig.11 One or more of the various steps described. In some embodiments, the global model sent by the NW to the UE may be a sparse model. When performing model training, the UE may perform sparse-to-dense model training to obtain a dense local model. In this case, the NW may perform compression aggregation based on the collected dense local models to obtain a sparse global model as an aggregated global model.

[0187] The present disclosure provides devices, methods and systems for machine learning. According to some aspects of the technology of the present disclosure, by exchanging local models between UEs and removing model redundancy before uploading the model, it is possible to reduce the occupancy of uplink resources. According to some aspects of the technology of the present disclosure, multiple UEs to be exchanged for local models can be selected to maximize model redundancy. According to some aspects of the technology of the present disclosure, high-quality communication resources can be provided for multiple UEs that generate similar local models, thereby improving the training efficiency of the entire system. Various aspects of the present disclosure may be implemented individually or in combination without limitation. The present disclosure provides improved technologies for machine learning including these and other aspects.

[0188] 5. Application Examples

[0189] The technology of the present disclosure can be applied to various products.

[0190] For example, the control side electronic device according to the embodiment of the present disclosure can be implemented as various control devices / base stations or included in various control devices / base stations. For example, the transmitting device and the terminal device according to the embodiment of the present disclosure can be implemented as various terminal devices or included in various terminal devices.

[0191] For example, the control device / base station mentioned in the present disclosure may be implemented as any type of base station, such as an eNB, such as a macro eNB and a small eNB. A small eNB may be an eNB that covers a cell smaller than a macro cell, such as a pico eNB, a micro eNB, and a home (femto) eNB. For another example, it may be implemented as a gNB, such as a macro gNB and a small gNB. A small gNB may be a gNB that covers a cell smaller than a macro cell, such as a pico gNB, a micro gNB, and a home (femto) gNB. Alternatively, the base station may be implemented as any other type of base station, such as an IAB (Integrated Access Backhaul). The 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 (RRH) arranged at a different place from the main body. In addition, the various types of terminals described below may all work as base stations by temporarily or semi-persistently performing base station functions. For example, the terminal device mentioned in the present disclosure may be implemented as a mobile terminal (such as a smart phone, a tablet personal computer (PC), a notebook PC, a portable game terminal, a portable / dongle-type mobile router, and a digital camera) or a vehicle-mounted terminal (such as an OBU) in some embodiments. The terminal device may also be implemented as a terminal for performing machine-to-machine (M2M) communication (also referred to as a machine type communication (MTC) terminal). In addition, the terminal device may be a wireless communication module (such as an integrated circuit module including a single chip) installed on each of the above-mentioned terminals.

[0192] The exemplary embodiments of the present disclosure are described above with reference to the accompanying drawings, but the present disclosure is certainly not limited to the above examples. Those skilled in the art may obtain various changes and modifications within the scope of the appended claims, and it should be understood that these changes and modifications will naturally fall within the technical scope of the present disclosure.

[0193] It should be understood that the machine executable instructions in the machine readable storage medium or program product according to the embodiment of the present disclosure can be configured to perform operations corresponding to the above-mentioned device and method embodiments. When referring to the above-mentioned device and method embodiments, the embodiments of the machine readable storage medium or program product are clear to those skilled in the art, so they are not repeatedly described. The machine readable storage medium and program product for carrying or including the above-mentioned machine executable instructions also fall within the scope of the present disclosure. Such storage media may include, but are not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, and the like.

[0194] In addition, it should be understood that the above series of processes and devices can also be implemented by software and / or firmware. In the case of being implemented by software and / or firmware, the storage medium of the relevant device stores the corresponding program constituting the corresponding software, and when the program is executed, various functions can be performed.

[0195] For example, a plurality of functions included in one unit in the above embodiments may be implemented by separate devices. Alternatively, a plurality of functions implemented by a plurality of units in the above embodiments may be implemented by separate devices, respectively. In addition, one of the above functions may be implemented by a plurality of units. Needless to say, such a configuration is included in the technical scope of the present disclosure.

[0196] In this specification, the steps described in the flowchart include not only processing performed in order in time series, but also processing performed in parallel or individually rather than necessarily in time series. In addition, even in the steps processed in time series, it goes without saying that the order can be appropriately changed.

[0197] 6. Exemplary embodiments of the present disclosure

[0198] 1. An electronic device for a first UE, wherein the electronic device comprises: a processing circuit, wherein the processing circuit is configured to: obtain a first artificial intelligence model trained by the first UE; receive a second artificial intelligence model trained by a second UE; determine a model difference between the second artificial intelligence model and the first artificial intelligence model; and send the model difference to a federated learning server.

[0199] 2. The electronic device as described in Example 1, wherein the model difference is sent to the federated learning server when the uplink quality of the first UE is lower than a predetermined link quality condition.

[0200] 3. The electronic device as described in Example 1, wherein the model difference is sent to the federated learning server when the model distance corresponding to the model difference is less than a predetermined model distance threshold.

[0201] 4. The electronic device as described in Example 3, wherein the first UE and the second UE belong to the same group.

[0202] 5. The electronic device as described in Example 4, wherein the grouping is determined by the federated learning server based on model distance.

[0203] 6. The electronic device as described in Example 4, wherein the grouping is determined by the first UE or the second UE based on a model distance.

[0204] 7. The electronic device as described in Example 4, wherein the processing circuit is further configured to: communicate with the second UE using resources specific to the group.

[0205] 8. An electronic device as described in Example 7, wherein the processing circuit is further configured to: receive resource configuration information from a network side device, and the resource configuration information indicates side link resources specific to the group for communication between the first UE and the second UE.

[0206] 9. An electronic device as described in Example 8, wherein the side link resources include a resource pool specific to the group, and the resource pool is allocated based on a training task of federated learning.

[0207] 10. An electronic device as described in Example 7, wherein the processing circuit is further configured to: in response to determining that the first UE and the second UE belong to the same group, maintain an initial communication channel between the first UE and the second UE; and in response to determining that the first UE and the third UE do not belong to the same group, deactivate the initial communication channel between the first UE and the third UE.

[0208] 11. An electronic device as described in Example 6, wherein the second UE is the UE in the group that completes the current training round of federated learning the fastest.

[0209] 12. An electronic device as described in Example 1, wherein the processing circuit is further configured to: in response to a specific value among the one or more values ​​representing the model difference being greater than a predetermined transmission threshold, include the specific value in the model difference sent; and in response to the specific value being not greater than the predetermined transmission threshold, not include the specific value in the model difference sent.

[0210] 13. An electronic device as described in Example 1, wherein the processing circuit is further configured to: receive a global artificial intelligence model from the federated learning server, wherein the global artificial intelligence model is generated by at least aggregating the model difference received from the first UE and the second artificial intelligence model received from the second UE.

[0211] 14. An electronic device for a network side, wherein the electronic device comprises: a processing circuit, wherein the processing circuit is configured to: receive a first artificial intelligence model from a first UE, wherein the first artificial intelligence model is trained by the first UE in a federated learning process; receive a model difference from a second UE, wherein the model difference represents a difference between the first artificial intelligence model and a second artificial intelligence model trained by the second UE in the federated learning process; and generate a global artificial intelligence model based on aggregating at least the first artificial intelligence model and the model difference.

[0212] 15. The electronic device of embodiment 14, wherein the model difference is received when the uplink quality of the second UE is less than a predetermined link quality condition.

[0213] 16. The electronic device as described in Example 14, wherein the model difference is received when a model distance corresponding to the model difference is less than a predetermined model distance threshold.

[0214] 17. The electronic device as described in Example 16, wherein the first UE and the second UE belong to the same group.

[0215] 18. The electronic device of embodiment 17, wherein the grouping is determined by the processing circuit based on a model distance.

[0216] 19. The electronic device as described in Example 17, wherein the grouping is determined by the first UE or the second UE based on a model distance.

[0217] 20. The electronic device of embodiment 17, wherein the processing circuit is further configured to: allocate resources specific to the grouping to the grouping.

[0218] 21. An electronic device as described in Example 20, wherein the resources include side link resources specific to the group for communication between different UEs in the group.

[0219] 22. An electronic device as described in Example 21, wherein the side link resources include a resource pool specific to the group, and the resource pool is allocated based on the training task of the federated learning process.

[0220] 23. An electronic device as described in Example 18, wherein the second UE is the UE in the group that completes the current training round of federated learning the fastest.

[0221] 24. The electronic device of embodiment 14, wherein a specific value among the one or more values ​​representing the model difference is received when the specific value is greater than a predetermined transmission threshold.

[0222] 25. An electronic device as described in Example 14, wherein the processing circuit is further configured to: send the global artificial intelligence model to the first UE and the second UE.

[0223] 26. A method performed by a first UE, comprising: obtaining a first artificial intelligence model trained by the first UE; receiving a second artificial intelligence model trained by a second UE; determining a model difference between the second artificial intelligence model and the first artificial intelligence model; and sending the model difference to a federated learning server.

[0224] 27. A method performed by an electronic device on the network side, comprising: receiving a first artificial intelligence model from a first UE, the first artificial intelligence model being trained by the first UE in a federated learning process; receiving a model difference from a second UE, the model difference representing a difference between the first artificial intelligence model and a second artificial intelligence model trained by the second UE in the federated learning process; and generating a global artificial intelligence model based on aggregating at least the first artificial intelligence model and the model difference.

[0225] 28. A computer-readable storage medium storing one or more instructions, wherein when the one or more instructions are executed by one or more processing circuits of an electronic device, the electronic device executes the method as described in any one of Embodiments 26-27.

[0226] 29. A computer program product, comprising a computer program, which, when executed by a processor, performs the method as described in any one of embodiments 26-27.

[0227] 30. An apparatus comprising means for performing the method of any one of embodiments 26-27.

Claims

1. An electronic device for a first UE, wherein: The electronic device comprises: A processing circuit, the processing circuit being configured to: Obtaining a first artificial intelligence model trained by the first UE; receiving a second artificial intelligence model trained by a second UE; determining a model difference between the second artificial intelligence model and the first artificial intelligence model; and The model differences are sent to a federated learning server.

2. The electronic device according to claim 1, wherein: The model difference is sent to the federated learning server when the uplink quality of the first UE is lower than a predetermined link quality condition.

3. The electronic device according to claim 1, wherein: The model difference is sent to the federated learning server when a model distance corresponding to the model difference is less than a predetermined model distance threshold.

4. The electronic device according to claim 3, wherein: The first UE and the second UE belong to the same group.

5. The electronic device according to claim 4, wherein: The grouping is determined by the federated learning server based on model distance.

6. The electronic device according to claim 4, wherein: The grouping is determined by the first UE or the second UE based on a model distance.

7. The electronic device according to claim 4, wherein: The processing circuit is further configured to: Communicate with the second UE using resources specific to the packet.

8. The electronic device according to claim 7, wherein: The processing circuit is further configured to: Resource configuration information is received from a network-side device, the resource configuration information indicating a side link resource specific to the packet for communication between the first UE and the second UE.

9. The electronic device according to claim 8, wherein: The side link resources include a resource pool specific to the group, and the resource pool is allocated based on a training task of federated learning.

10. The electronic device according to claim 7, wherein: The processing circuit is further configured to: In response to determining that the first UE and the second UE belong to the same group, maintaining an initial communication channel between the first UE and the second UE; as well as In response to determining that the first UE and the third UE do not belong to the same group, an initial communication channel between the first UE and the third UE is deactivated.

Images