Federated split deep neural network computation

Abstract

Federated machine learning (FML) is combined with user equipment (UE) split deep neural network (DNN) calculations. The plurality of UEs are assigned to a federated learning group for performing split execution of the neural network (106) such that the computing system performs a respective first portion (106A) of the neural network and each UE in the federated learning group performs a respective second portion (106B) of the neural network. An initial indication of a neural network configuration (114A-M) of a second portion of the neural network is sent to UEs in the federated learning group. A parameter of a second portion of the neural network is updated based on update information (126-M) for the respective second portion of the neural network received from the at least two UEs, and an update indication (130) of the updated parameter is sent to UEs in the federated learning group.

Description

Background Technology

[0001] The use of deep neural networks (DNNs) is expected to play a significant role in fifth-generation (5G) advanced and sixth-generation (6G) cellular standards. For example, DNNs can be used for image rendering to support extended reality (XR) use cases such as virtual reality (VR) and augmented reality (AR).

[0002] However, user equipment (UE) executing such applications may have constraints affecting the UE's ability to execute DNNs, such as constraints on processing power, available memory, and available battery power. These constraints can be temporary, such as available battery power, current memory or processor usage, or permanent, such as the UE's hardware specifications. Additionally, the UE allows users to control access to data. Summary of the Invention

[0003] In a first aspect, this disclosure provides a computer-implemented method executed by a computing system, the method comprising: receiving a request for neural network configuration from a user equipment (UE); assigning the UE to a federated learning group comprising a plurality of UEs executing a split execution of the neural network based on the request, such that the computing system executes a corresponding first portion of the neural network for each UE in the federated learning group, and each UE in the federated learning group executes a corresponding second portion of the neural network; sending an initial instruction for neural network configuration of the second portion of the neural network to the plurality of UEs in the federated learning group; receiving corresponding update information for the corresponding second portion of the neural network executed at the plurality of UEs from at least two of the plurality of UEs in the federated learning group; updating parameters of the second portion of the neural network based on the update information; and issuing an update instruction for the updated parameters of the second portion of the neural network to the plurality of UEs in the federated learning group.

[0004] In a further aspect, this specification provides a computer-implemented method executed by a user equipment (UE), the method comprising: sending a request for neural network configuration to a network node; in response to the request, receiving an initial indication of neural network configuration for a second part of a neural network to be executed locally at the UE; receiving from the network node intermediate output of a neural network model generated by a first part of the neural network; processing the intermediate output by the second part of the neural network model to generate a model output of the neural network; generating an update to the second part of the neural network based on the model output and using a local update process; and sending the update to the network node.

[0005] Further aspects of this disclosure provide systems, apparatus, and computer-readable media for implementing any of the methods described herein.

[0006] The systems, methods, and apparatuses described herein enable one or more of the following advantages. Combining federated learning with split deep neural networks (DNNs) ensures that sensitive data portions (such as private user data) can be retained locally at the UE during both DNN inference and DNN training, while allowing the UE to utilize network resources when executing the DNN, for example, to execute potentially larger and / or more complex DNNs than could be executed using only the local UE. Attached Figure Description

[0007] Figure 1A An overview of example systems and methods for assigning UEs to federated learning groups is shown; Figure 1B An overview of example systems and methods for performing split DNNs is shown; Figure 1C An overview of example systems and methods for updating split DNNs using federated machine learning is shown; Figure 2 An overview of a further example system for performing split DNN is shown, where the UE-side split of the DNN is not the same for all UEs in the federated learning group; Figure 3 A signaling diagram for performing federated learning using a split DNN architecture is shown; Figure 4A An example of a signaling graph for performing user plane data splitting at a two-part DNN is shown; Figure 4B A further example of a signaling graph for performing user plane data splitting at a three-part DNN is shown; Figure 4C A further example of a signaling graph for performing user plane data splitting at a two-part DNN is shown; Figure 4D A further example of a signaling graph for performing user plane data splitting at a three-part DNN is shown; Figure 5A An example of a federated machine learning process for splitting a two-part DNN is shown; Figure 5B An example of a federated machine learning process for a three-part split DNN is shown; Figure 5C Further examples of federated machine learning processes for splitting DNNs into two and / or three parts are shown; Figure 6 A flowchart is shown for an example network-side method for performing federated learning using a split DNN architecture; Figure 7A flowchart is shown for an example UE-side method for performing federated learning using a split DNN architecture; Figure 8 An example system for executing any of the methods described is shown; and Figure 9 An example computer-readable medium is shown. Detailed Implementation

[0008] This specification describes systems, methods, and apparatuses that combine federated machine learning (FML) with user equipment (UE) split deep neural network (DNN) computation, where FML addresses UE privacy constraints, and split computation addresses both UE local computational capabilities and UE privacy constraints. In addition to addressing privacy and power constraints, combining FML with split DNN also helps improve network efficiency when updating DNN models.

[0009] Figures 1A to 1C An overview of an example system and method for building and executing a split DNN-based FML is shown. The system includes network entities 102 (also referred to herein as “network nodes” or “network elements”), such as base stations, edge devices, or core network components, and multiple user equipment (UEs) 104A-N. The system is configured to use a DNN 106 to build and execute an FML, the execution of which is split between network entity 102 and each of the multiple UEs 104A-M in a federated machine learning group 110 (“FML group”).

[0010] Figure 1A An overview of a sample system and method 100A for assigning UE 104A-M to Federated Machine Learning (FML) group 110 is shown. Each of the multiple UEs 104A-N sends a request 112A-N for configuring DNN 106 to network entity 102. (About...) Figure 3 The details of request 112A-N and the conditions that can trigger a request from UE 104A-N are described in further detail.

[0011] Based on requests 112A-N, network entity 102 determines whether to configure UE 104A-N for split execution of DNN 106. The network entity groups multiple UE 104A-Ms into group 110 for FML based on, for example, the configuration of UE 104A-M in FML group 110 for split execution of DNN 106 and / or the similarity of capabilities of UE 104A-Ms in that group.

[0012] In the example shown, a UE 104N is not assigned to FML group 110, for example, because the unassigned UE 104N has different capabilities relative to UE 104A-M in FML group 110 and / or has been assigned a different configuration for split execution of DNN 106 by network entity 102. If a further FML group (not shown) containing UEs with similar capabilities to the unassigned UE 104N already exists, the unassigned UE 104N is assigned to that further FML group. If network entity 102 has not yet established such a group, UE 104N remains unassigned to an FML group until network entity 102 determines that one or more further UEs (not shown) requesting DNN 106 architecture configuration are suitable for grouping with 104N to form a further FML group. Furthermore, if a UE in FML group 110 undergoes a capability change, network entity 102 may regroup that UE into another FML group.

[0013] For UE 104A-M, which has been assigned the split DNN 106 configuration by the network entity, network entity 102 splits the DNN 106 into a first part 106A and a second part 106B based on request 112A-M, and sends configuration data 114A-M to UE 104A-M in FML group 110. At inference time, in the example shown, network entity 102 executes a local copy 108A-M (also called a "replica") of the first part 106A (i.e., the input part of the DNN) and each UE in UE 104A-N executes a local copy 108A-M (also called a "replica") of the second part 106B (i.e., the output part of the DNN), as per [reference to...]. Figure 1B and Figure 4A As described. In some alternative examples, each UE in UE 104A-M executes a local copy 108A-M of the first part 106B of the DNN (i.e., the input part of the DNN) and network entity 102 executes the second part of the DNN (i.e., the output part of the DNN), as described in... Figure 4C As described.

[0014] In the example shown, DNN 106 is split into two parts, 106A and 106B, for UE 104A-M in FML group 110. However, in some alternative examples, DNN 106 can be split into an initial / input part, an intermediate / middle part, and an output / final part. In some implementations, the initial and final parts are performed by UE 104A-M, while the intermediate part is performed by network entity 102. Regarding... Figure 4BAn example of this configuration is described. In an alternative implementation, network entity 102 executes the initial and final parts of the DNN, with the intermediate parts executed by UE 104A-N. (About...) Figure 4D An example of this type of configuration is described.

[0015] Network entity 102 determines whether to group multiple UEs 104A-M into FML group 110 based on request 112A and the structure and split of DNN 106. In some examples, UEs with the same split position (i.e., UEs 104A-M with the same architecture in their UE-side DNN 108) are assigned to the same FML group 110. In some examples, UEs whose UE-side DNN splits share at least a threshold number of layers with each other are assigned to the same FML group 110, as shown in the example regarding... Figure 2 As described. Alternatively or additionally, UEs with similar capabilities and / or local conditions are assigned to the same FML group 110.

[0016] In some implementations, if an FML group 110 has already been established for UEs 104A-M with similar attributes (e.g., the same split location, similar capabilities, etc.), network entity 102 will assign the newly requesting UE (not shown) to that existing group. If no suitable FML group is available, network entity 102 will establish another FML group and assign the newly requesting UE (as well as other requesting UEs with similar attributes) to that other FML group.

[0017] After requesting that UE 104A-M has been assigned to FML group 110, network entity 102 sends configuration message 114A-M to UE 104A-M in FML group 110. For example, to save network overhead, in some implementations, network entity 102 multicasts configuration message 114A-M to UE 104A-M in FML group 110. In some examples, network entity 102 uses unicast messaging, such as when a UE joins an established FML group 110 or hands over between FML groups 110, or when a UE performs a cell handover or other mobility procedure. In some examples, configuration message 114A-M is sent as a Non-Access Stratum (NAS) layer message. A separate configuration message 114N is also sent to UE 104N that is not assigned to FML group 110, for example, indicating that UE 104N is assigned to a further FML group or is not assigned to any FML group.

[0018] In some implementations, configuration messages 114A-N include an indication of the architecture of the second part 106B of the DNN. For example, configuration message 114A-M indicates the network structure of the replica 108A-M to be formed at each UE 104A-M.

[0019] In some implementations, configuration message 114A-M includes FML configuration parameters indicating the attributes of the FML procedures that the FML group 110 identity and / or federal UE 104A-M should use. (Regarding...) Figure 3 Examples of FML configuration parameters are described in further detail.

[0020] In some implementations, UE 104A-M can execute multiple split DNN parts. For example, each split DNN part performs a different function for the UE and is therefore assigned to more than one FML group. For example, for each DNN that the UE is executing, the UE can be assigned to a different FML group. For example, the FML group of the first DNN contains UE 104A-M, while different subsets of UE 104A-B also implement a second DNN, which can be a split or complete (non-split) DNN separate from the network entity.

[0021] Figure 1B An overview of an example system and method 100B for performing split DNN 106 is shown. In the examples shown, the UE-side portion of neural networks 108A-M has the same architecture for each UE 104A-M in FML group 110. However, in some examples, a proper subset of UE 104A-M in FML group 110 performs one or more additional layers, as per [reference to...]. Figure 2 Further detailed description. For simplicity, a single FML group 110 is shown.

[0022] When performing split execution on neural network 106 using UE 104A-M, network entity 102 uses the first part 106A of the neural network to process the input data 116A-M associated with UE 104A-M to generate the corresponding intermediate output 120A-M, i.e., the output of the intermediate layer 118 of neural network 106. In some implementations, the input data 116A-M is generated / captured at UE 104A-M and sent from UE 104A-M to network entity 102, for example, as per [reference to...]. Figure 4A As described. In some implementations, the input data originates from network entity 102, or from the core network, for example, as per [the description of...]. Figure 4D As described.

[0023] Given an input, intermediate output 120A-M is sent from network entity 102 to the corresponding UE 104A-M. The corresponding UE 104A-M receives intermediate output 120A-M and processes it using a local version of the second part of DNN 108A-M (also referred to as a "copy" of the second part 106B of DNN 106) to generate the corresponding network output 122A-M.

[0024] exist Figure 1B In the example shown, DNN 106 is split into two parts. However, in some implementations, the network entity splits the DNN into three or more parts. Regarding... Figure 4B and Figure 4D An example of such a three-part DNN configuration is described.

[0025] Figure 1C An overview of an example system and method 100C for using federated machine learning to update split DNN 106 is shown.

[0026] Each UE 104A-M in FML group 110 is based on the corresponding output data 122A-M generated at the UE (see [link]). Figure 1B The local update process 124A-M is executed to generate update information 126A-M. In some implementations, the update process 124A-M at each UE 124A-M includes determining the gradient of the parameters (i.e., weights and / or biases) of the objective function relative to the UE-side split of the DNN 108A-M, for example, using backpropagation of gradients. In some implementations, the local update process 124A-M includes applying optimization routines (such as stochastic gradient descent) to the objective function to determine updates to the values ​​of the parameters of the UE-side split of the DNN 108A-M. In some examples, the local update process 124A-M is a supervised learning process (e.g., based on an objective function that compares the network output 122A-M with corresponding ground truth data, i.e., a labeled training dataset), a semi-supervised learning process (e.g., using a mixture of labeled and unlabeled training data), or an unsupervised learning process (e.g., using unlabeled training data).

[0027] Update information 126A-M from each UE 104A-M in FML group 110 is sent to network entity 102. In some examples, each UE 104A-M sends a NAS layer message containing its corresponding update information 126A-N. Network entity 102 uses the update information 126A-M to perform update procedure 128 (also referred to as the "global update procedure") on the second part 106B of the DNN (i.e., the network-side version / copy of the DNN split on the UE side) using FML technology.

[0028] In some implementations, update information 126A-M includes updated parameter values ​​for the UE-side split of DNN 108A-N. For example, in some implementations, update information 126A-M includes updated parameter values ​​for the UE-side split of DNN 108A-M. In other alternative implementations, update information 126A-M includes the difference between the current parameter values ​​and the updated values ​​for the UE-side split of DNN 108A-M, i.e., the "delta" determined by the local update process 124A-M. The network entity combines update information 126A-M from UEs in FML group 110, for example, by averaging or weighted averaging, to determine the update for the second part 106B of the DNN.

[0029] In some implementations, update information 126A-N includes local gradients of the parameters of the objective function relative to replicas 108A-M. Network entity 102 combines the local gradients from UEs 104A-M in the FML group and uses the combined gradients to determine updates to the parameters of the second part 106B of the DNN, for example, using an optimization routine such as stochastic gradient descent. In some implementations, network entity 102 uses the local gradients to update the first part 106A of the DNN, for example, by backpropagating the local gradients through the first part 106A of the DNN and applying the optimization routine.

[0030] In one example where DNN 106 has three split parts (see...) Figure 4B and 4D In some implementations, the network-side portion of the DNN sends gradient information to the UE 104A-M, which uses this gradient data to perform a local update process on the preceding part of the DNN. The UE 104A-M then sends the local update information of the preceding part of the DNN back to the network entity 102. The network entity 102 uses this local update information to update the network-side version of the preceding part of the DNN using the FML procedure.

[0031] Figure 2 An overview of a further example system / method 200 for performing DNN splitting is shown, where the UE-side splitting of DNN 208A-M is not identical for all UEs 204A-M in FML group 210. For ease of presentation, only two UEs, 202A and 204M, are shown. Each UE 208A-M in FML group 210 performs a first set (also referred to as the "common layer set") of one or more DNN layers common across FML group 210. Figure 1B In the example, all UE 104A-Ms in FML group 110 only execute these common layers. However, in Figure 2In the example, each UE in one or more UEs 204A in FML group 210 executes one or more additional layers of the DNN. In other words, all UEs 204A-M in FML group 210 execute a common set of layers that reflect a copy of the second part 206B of split DNN 206, and a proper subset of UEs 204A in FML group 210 executes one or more additional layers of split DNN 206.

[0032] The method is as described above. Figure 1B As described, except that the intermediate output 220A of UE 204A, which performs the additional layer, is obtained from an earlier intermediate layer 218A than the intermediate output 220M of UE 204M, which only performs the common layer, from the first part 206A of the DNN. In the FML update process (corresponding to...) Figure 1C During this period, the network entity issues updates related to the common layer 206B received from all UEs 204A-M in FML group 110 to update the common layer 206B of the split DNN 206. In some examples, additional update information is received from UE 204A that performs the additional layer. Network entity 202 uses this additional update information to update the version of the additional layer 218M maintained by the network entity.

[0033] Figure 3 A high-level signaling diagram for performing federated learning using a split DNN architecture is shown. The example shown illustrates two UEs, 304A and 304B, but typically multiple UEs can be utilized. Operations 312, 340, 314, and 342 are related to... Figure 1A The methods described correspond to those described above. Operations 344A and 344B are related to... Figure 1B The methods described correspond to those described. Operations 324-332 and about Figure 1C The methods described correspond to this.

[0034] Each UE 304A, 304B sends a corresponding request for DNN configuration 312A, 312B to network node / entity 302. For example, requests 312A, 312B from UE 304A, 304B are triggered when one or more conditions are met. For example, requests 312A, 312B may be triggered when UE 304A, 304B determines that it cannot / is no longer able to perform DNN locally, for example, due to a change in local conditions at the UE. Alternatively or additionally, requests 312A, 312B may be triggered when changes to the DNN constraints of UE 304A, 304B cause UE 304A, 304B to no longer support the current DNN configuration (e.g., the DNN configuration previously provided by network entity 302). Alternatively or additionally, when a UE that has already performed a split execution of DNN 306 with a different network entity (not shown) joins the cell of network entity 302, requests 312A and 312B from UEs 304A and 304B are triggered.

[0035] Requests 312A and 312B are sent, for example, as messages in the Non-Access Stratum (NAS) layer of the protocol stack (e.g., as Information Elements (IEs)). One or more messages may form the basis of a request, and a request may include any of several parameters, such as explicit requests, the current DNN-specific capabilities of UE 104, or DNN-specific parameters, and may provide implicit requests.

[0036] In some implementations, requests 312A and 312B for DNN architecture configuration identify a specific DNN, for example, the request may contain an identifier of the DNN being executed for the requested split. Alternatively, requests 312A and 312B for DNN architecture configuration identify the functions requested by UEs 304A and 304B, such as image / audio classification, image / audio generation, network resource prediction, augmented reality, image / video / audio enhancement (such as super-resolution and / or denoising), video structure reconstruction, or similar functions. Based on the identified functions, network node 302 selects the DNN to use from a plurality of available DNNs.

[0037] In some implementations, the requests 312A and 312B for configuring the DNN architecture include an explicit indication of the location of the DNN split, i.e., the UE 304A and 304B that issued the request determines the split location and transmits it to network entity 302 as part of the requests 312A and 312B.

[0038] In some alternative implementations, requests 312A and 312B for DNN architecture configuration include UE capability information or UE auxiliary information, as defined by 3GPP, indicating one or more capabilities of the requested UE 304A and 304B. In some implementations, the UE capability information or UE auxiliary information is provided separately to requests 312A and 312B as part of another message (i.e., a non-DNN-specific message). Examples of UE capabilities include UE hardware specifications, such as processor speed and / or memory size.

[0039] Alternatively or additionally, requests 312A and 312B may include indications of DNN-related constraints corresponding to UEs 304A and 304B. DNN-related constraints may include one or more of the following: the number / type of DNN nodes supported by the UE; the type of DNN layers supported by the UE; the type of convolutional layers supported by the UE (e.g., the size of the supported convolutional filters); and / or fully connected layer capabilities (e.g., the maximum number of nodes supported in a fully connected layer).

[0040] Alternatively or additionally, requests 312A and 312B indicate one or more local UE conditions. Examples of such conditions include one or more of the following: current processor and / or memory availability / usage; current power level / availability / usage; signal strength at the UE location; and / or UE temperature. In such an implementation, network entity 102 determines the location of each UE's split based on capability data and / or local conditions received from the respective UEs 304A and 304B.

[0041] UE capability information, UE assistance information, local UE conditions, and / or DNN-related constraints are collectively referred to as "UE constraints" or "UE local constraints." In some examples, UE constraints include privacy constraints.

[0042] Network node 302 determines the split architecture of 340 DNN for each requesting UE 304A, 304B, and determines whether FML can be supported for a set of requesting UEs 304A, 304B.

[0043] In the implementation of requests 312A and 312B including the splitting position of the DNN, network node 302 uses the splitting position to determine the relationship with... Figure 1B or Figure 2 A consistent DNN splitting architecture for each request UE 304A, 304B.

[0044] In an implementation where network node 302 is provided with UE constraints (as part of requests 312A, 312B, or otherwise), network node 302 determines the splitting location of the DNN. For example, network node 302 uses a lookup table accessible to network node 302 to determine the DNN splitting location based on the UE constraints. In an alternative implementation, network node 302 determines the resource requirements for each of a plurality of candidate splitting locations. The network node then selects a splitting location from the plurality of candidate splitting locations based on the UE constraints compared to the resource requirements. For example, network node 302 selects the splitting location that yields the minimum network-side splitting that satisfies the UE constraints.

[0045] In some implementations, network node 302 determines whether FML can be supported for a group of requesting UEs 304A and 304B based on the determined split positions of UEs 304A and 304B. For example, an FML group is created from multiple UEs 304A and 304B that have the same split position of their DNNs (i.e., their network-side portions of the DNNs are the same and their UE-side portions are the same), as per [reference to...]. Figures 1A to 1C As described. Alternatively, FML groups are created from multiple UEs (304A, 304B) that meet one or more similarity criteria (e.g., having at least a threshold number of common layers and / or fewer than a threshold number of distinct layers) split from the UE-side DNN, for example, as per [reference to...]. Figure 2 As described.

[0046] Alternatively, network node 302 determines whether FML can be supported for a group of requesting UEs 304A and 304B based on the UE constraints of UEs 304A and 304B. For example, network node 302 groups multiple UEs 304A and 304B with similar (or identical) UE constraints into FML groups.

[0047] After determining the network architecture and FML grouping at 340, network node 302 causes the initial indication of the neural network configuration 314 of the second part of the neural network to be transmitted to requesting UEs 304A and 304B. In some implementations, the initial indication of the neural network configuration 314 is multicast to UEs 304A and 304B that have already been grouped into the same FML group. In examples where UEs join an existing FML group, the initial indication of the neural network configuration 314 is, for example, unicast to the UEs joining the FML group.

[0048] In the implementation of requests 312A and 312B that include the split location, the initial indication of the neural network configuration 314 may include a confirmation message that the network node 302 has accepted the split requested by the UE.

[0049] In the implementation of network node 302 determining the DNN splitting location, the initial indication of neural network configuration 314 includes an indication of the architecture of the UE-side replica of the split DNN part.

[0050] In some examples, the initial indication of neural network configuration 314 includes the splitting location of the DNN, such as an indication of the layer where the DNN splitting occurs, like the layer number. Alternatively, the initial indication of neural network configuration 314 includes an indication of the layers that UE 304A, 304B will execute, such as an identifier of one or more layers of the DNN that UE 304A, 304B will execute.

[0051] Alternatively, in some implementations, the initial indication of the neural network configuration 314 includes explicit indications of the network structure of each part of the DNN to be executed by the UEs 304A and 304B. For example, the initial indication of the neural network configuration 314 indicates the type and / or structure of each layer of the DNN split on the UE side, the parameter values ​​(i.e., weights and / or biases) of the nodes in the second part of the DNN, and / or the activation function used in the DNN. In some implementations, the initial indication of the neural network configuration 314 includes references to predefined DNN layers stored in memory on the UEs 304A and 304B. In some such implementations, the initial indication of the neural network configuration 314 includes instructions for modifying predefined DNN layers, such as updating the weights and / or biases of predefined layers stored at the UEs 304A and 304B.

[0052] Network node 302 causes the transmission of FML configuration 342 to request UEs 304A and 304B. FML configuration 342 includes parameters related to the FML process. For example, the FML configuration includes an identifier of the FML group to which the UE has been assigned. FML configuration 342 alternatively or additionally includes, for example, an update type indicating what type of update UE 104A-M should issue to network entity 102, such as updated parameter values, changes to parameter values, loss function values, and / or the gradient of the objective function relative to network parameters. Alternatively or additionally, FML configuration 342 instructs UE 104A-N to issue one or more triggers for local updates to network entity 102, such as a time period or the number of local update iterations.

[0053] In some implementations, network node 302 multicasts FML configuration 342 to UEs 304A and 304B. For example, when an FML group is initialized, network node 302 uses multicast to all UEs 304A and 304B assigned to the FML group. The network node also uses multicast to send updates to FML configuration 342 to UEs 304A and 304B in the FML group. In some examples, for subsequent UEs added to the FML group, network node 302 unicasts FML configuration 342 to the newly added UEs.

[0054] Although the initial instructions for neural network configuration 314 and FML configuration 342 are in Figure 3 While shown as separate messages, it should be understood that in some examples, they are combined into a single message that indicates both the architecture of the DNN split on the UE side and the FML configuration.

[0055] In some implementations, each UE 304A, 304B sends an acknowledgment message (not shown) to network entity 302 to indicate that the UE 304A, 304B is ready for split execution of DNN.

[0056] After configuration, network node 302 and UEs 304A and 304B perform DNN splitting and execution of 344A and 344B. (About...) Figures 4A to 4C The DNN splitting process is described in further detail. Generally, network entity 302 performs one or more splits of the DNN, and UEs 304A and 304B perform one or more further splits of the DNN.

[0057] Network node 302 and UEs 304A and 304B in the federated learning group execute federated machine learning procedure 346 to update the UE-side DNN split at network node 302. Regarding Figure 5A Figure 5D provides a further detailed example of the signaling for the federated machine learning process 346. Generally, UEs 304A and 304B perform corresponding local updates on their respective UE-side DNN replicas to generate corresponding local update data. UEs 304A and 304B transmit their corresponding local update data to network node 302, which uses the local update data to update the network entity versions split by the UE-side DNN.

[0058] Following FML procedure 346, network node 302 sends update instruction 332 to UEs 304A and 304B in the federated learning group, indicating an update to the UE-side DNN split. Update instruction 332 can be multicast or unicast to the UEs in the FML group. UEs 304A and 304B use update instruction 332 to update their local copies of the UE-side DNN split.

[0059] Figure 4A An example of high-level signaling diagram 444A for performing user plane data splitting at a two-part DNN is shown. This diagram is related to... Figure 1B The method described in [the document] corresponds to and is detailed in [the document]. Figure 3 The steps introduced as element 344 in the middle.

[0060] Network node 402 performs a 446-fold splitting of the DNN on the input data set 416 to generate an intermediate DNN output 420, such as the values ​​of node activations at intermediate layers of the DNN. The network node sends the intermediate output 420 to UE 404, which performs a 448-fold splitting of the DNN on the intermediate output data to generate the final output of the DNN. In some implementations, UE 404 may perform one or more actions based on the final output. For example, UE 404 causes the final output to be rendered on a display associated with the UE.

[0061] In some implementations, the input data set 416 is sent from UE 404 to network node 402, for example, as data plane communication. For instance, UE 404 collects the input data set using one or more sensors of UE 404 and / or one or more user inputs to UE 404. As an example, the input data may reflect the UE's positioning or heading information, and the final output data may be a video image of the UE oriented in a context reflecting the positioning or heading of the input data.

[0062] Alternatively or additionally, at least a portion of the input data 416 originates from the network side, for example, collected by network node 402 or a further network entity. Continuing with the previous example, the final output data video image may be part of a 360-degree background image received as input data from the application server.

[0063] Figure 4B An example of a high-level signaling graph for performing user plane data splitting at a three-part DNN is shown. This graph corresponds to an example where the DNN is split into three parts. This graph is... Figure 1B The method described in [the document] corresponds to and is detailed in [the document]. Figure 3 The steps introduced as element 344 in the middle. Figure 4B and Figure 4C Such examples can provide users with additional privacy because the input data does not leave UE 404.

[0064] Figure 4B Examples are similar to Figure 4AIn this example, the DNN is further split on the UE side, meaning the split DNN has an initial / input portion on the UE side, an intermediate portion on the network side, and a final / output portion on the network side. Input data at UE 404 is fed into a local copy of the initial portion of the DNN, which processes 450 input data 416 to generate an initial (first) intermediate output. The UE sends the initial intermediate output 452 to network entity 402. Network entity 402 feeds the initial intermediate output into the intermediate portion of the DNN, which processes 452 the initial intermediate output to generate a further (second) intermediate output. Network entity 402 sends this further intermediate output 454 to UE 404. UE 404 feeds the further intermediate output into a local copy of the final portion of the DNN, which processes 456 further intermediate output data to generate the final output of the DNN.

[0065] Figure 4C A further example of a high-level signaling graph for performing user plane data splitting at a two-part DNN is shown. In this example, when with Figure 4A In contrast, the DNN split is reversed, that is, the first (input) part of the DNN is executed by UE 404, and the second (output) part of the DNN is executed by network entity 402.

[0066] The first part of the DNN at UE 404 performs step 446 on the input data generated / captured by UE 404 to generate intermediate output data. UE 404 sends the intermediate output data to network entity 402 at step 420. Network entity 402 performs step 448 on the intermediate output data to generate the second part of the DNN to generate the DNN's output data. In some implementations, network entity 402 sends output data 466 to UE 404. In some implementations, network entity 402 retains the output data 466.

[0067] Figure 4D A further example of a high-level signaling graph for performing user plane data splitting at a three-part DNN is shown. In this example, when with Figure 4B In contrast, the DNN split is reversed, that is, the first (input) part of the DNN is executed by network entity 402, the second (intermediate) part of the DNN is executed by UE 404, and the third (output) part of the DNN is executed by network entity 402.

[0068] This method is as follows: Figure 4AThe process proceeds as described, up to and including the second part of the DNN performed by UE 404 (448). The output of the second part of the DNN is considered further intermediate output data and is sent (458) to network entity 402. Network entity 402 performs the third part of the DNN (460) on the further intermediate output data to generate the DNN output data (466). In some implementations, network entity 402 sends the output data (466) to UE 404. In some implementations, network entity 402 retains the output data (466).

[0069] Figure 5A An example of a high-level signaling graph for performing FML on a two-part split DNN is shown. Element 546 with Figure 3 The corresponding element is 346. For convenience, a single UE 504 is shown, but typically multiple UEs participate in the FML process and operate in parallel using their own local input data.

[0070] UE 504 performs a local update process 524 on the UE-side split of the DNN to generate local update information 526. For example, UE 504 determines the value of the loss / objective function based on the data output by the DNN during the split execution of the DNN, and applies at least a portion of an optimization routine (such as stochastic gradient descent) to the loss / objective function to determine the update information of the UE-side split of the DNN. In the example using supervised learning, the loss / objective function compares the output of the DNN with the true value output.

[0071] In some examples, update information 526 includes an indication of the updated values ​​of the parameters of the DNN's UE-side split, such as the updated value of the parameter itself or the difference between the current value and the updated value of the parameter. Alternatively or additionally, update information 526 includes the gradient of the loss / objective function relative to the parameters of the DNN's UE-side split, for example, as determined by backpropagation using the gradients of the DNN's UE-side split. In some implementations, update information 526 includes one or more loss function values, i.e., the values ​​of the loss function are computed using the output data output at UE 504.

[0072] UE 504 sends local update information 526 to network node 502. Each UE in the FML group (not shown) that has already performed local update procedure 524 may perform a similar operation at different times. Network node 502 uses the local update information 526 received from multiple UEs to perform aggregated updates 528 of the UE-side split stored at network node 502. For example, if the local update information 526 includes indications of updated values ​​of parameters of the DNN for the UE-side split, network node 502 averages the indications of updated values ​​of the parameters over the set of received update information 526 and applies these updates to the UE-side split DNN stored at network node 502.

[0073] In the implementation where update information 526 includes the gradient of the loss / objective function, network node 502 can average the gradient values ​​over the received update information set 526 and use the average gradient values ​​to determine the parameter updates of the UE-side split DNN, for example, using a gradient ascent or descent process. In some examples, network node 502 backpropagates the average gradient values ​​through the network-side DNN split to determine the gradient of the loss function relative to the parameters of the network-side DNN split, and uses these gradients to perform updates 530 to the network-side DNN split parameters, for example, using a gradient ascent or descent process.

[0074] Figure 5B An example of a high-level signaling graph for performing FML on a tripartite splitting DNN is shown. Tripartite splitting DNN and related... Figure 4B The described DNN splitting corresponds to this.

[0075] UE 504 performs a local update process 524 on the UE-side split of the DNN to generate local update information 526 for the final part of the DNN split, and sends the local update information 526 to network node 502, such as regarding Figure 5A As described.

[0076] In an implementation where the local update information 526 includes an indication of the updated values ​​of the parameters of the final UE-side split of the DNN, rather than gradient information, network node 502 performs actions as described above. Figure 5A The described update aggregation. In this implementation, the network-side split and the initial UE-side split of the DNN are not updated; only the final UE-side split of the DNN is updated.

[0077] However, in the implementation shown, the local update information 526 includes gradient information. Network node 502 uses the gradient information from the final part of the UE-side split to update the network-side DNN splitting parameters 530, for example, as... Figure 5AAs described in [the document]. Additionally, network node 502 backpropagates the gradient of the loss function through the initial portion of the UE-side split of the DNN to determine the gradient of the parameters of the loss function relative to the initial portion of the UE-side split of the DNN. Network node 502 uses these gradients to determine an update 562 for the initial portion of the UE-side split. Network node 502 sends the determined update 562 for the initial portion of the UE-side split to UE 504 in the federated learning group, for example, as [example missing]. Figure 3 As described in operation 332. UE 502 uses update 562 to update the UE-side split of the copy of their corresponding DNN.

[0078] Figure 5C Further examples of high-level signaling graphs for performing FML on two-part split DNNs or three-part split DNNs are shown. Two-part split DNNs and... Figure 4C The DNN split shown corresponds to this: the input part of the DNN is executed by UE 504, and the output part of the DNN is executed by network node 502. The three-part split DNN corresponds to... Figure 4D The DNN split shown corresponds to the following: the input part of the DNN is executed by network node 502, the middle part of the DNN is executed by UE 504, and the output side of the DNN is executed by network node 502.

[0079] In this example, the DNN output occurs at network node 502. Therefore, in some implementations, network node 502 determines the parameter updates for the neural network without performing any local update process at UE 504 in the FML group. The entire DNN update process 568 is performed at network node 502. For example, network node 502 applies an optimization routine such as stochastic gradient descent to the loss function determined based on the DNN output. Network node 502 sends the determined UE-side split updates to UE 504 in the federated learning group, for example, as... Figure 3 As described in Operation 332. UE 502 uses these updates to update the UE-side split replicas of their corresponding DNNs.

[0080] However, in some implementations, the ground truth data corresponding to the DNN output is stored locally at UE 504. To avoid privacy issues arising from sharing such ground truth data, network node 502 sends the DNN output 566 to UE 504. UE 504 determines one or more loss function values ​​based on the ground truth data and the network output, and sends these loss function values ​​back to network node 502. Network node 502 then uses the loss function values ​​to perform the DNN update process 568. Network node 502 sends the determined UE-side split updates to UE 504 in the federated learning group, for example, as... Figure 3 As described in operation 332. UE 504 uses update 562 to update the UE-side split replicas of their corresponding DNNs.

[0081] Figure 6 A flowchart of an example network-side method for performing federated learning using a split DNN is shown. This method is performed by network nodes / entities, as shown in Figure 1 and... Figure 8 The network node / entity described. For convenience, the method is described as being executed by the system. In some implementations, the method corresponds at least in part to any one or more of the methods described with respect to Figures 1 through 5.

[0082] At operation 612, the system receives a request for neural network configuration from the UE. In some examples, operation 612 is related to... Figure 1A element 112 and / or Figure 3 The element 312 corresponds to this.

[0083] At operation 640, the system assigns the UE to a federated learning group based on the request. In some examples, operation 612 is related to... Figure 3 The element 340 corresponds to this.

[0084] At operation 614, the system sends initial instructions for the neural network configuration of the second part of the neural network to multiple UEs in the federated learning group. In some examples, operation 614 is related to... Figure 1A element 114 and / or Figure 3 The elements 314 and / or 342 correspond to each other.

[0085] At operation 626, the system receives update information from at least two of the multiple UEs in the federated learning group regarding the corresponding second part of the neural network executed at those multiple UEs. In some examples, operation 626 is related to... Figure 1C Elements 126A-M and / or Figures 5A to 5B The element 526 corresponds to this.

[0086] At operation 628, the system updates the parameters of the second part of the neural network based on the updated information. In some examples, operation 626 is... Figure 1C The elements 128 and / or Figures 5A to 5B The element 528 corresponds to this.

[0087] At operation 632, the system issues update instructions for the updated parameters of the second part of the neural network to multiple UEs in the federated learning group. In some examples, operation 632 is related to... Figure 1C element 132 and / or Figure 3 The element 332 corresponds to this.

[0088] Figure 7 A flowchart of an example UE-side method for performing federated learning using a split DNN is shown. This method is performed by the UE, as shown in Figure 1 and... Figure 8 The UE is described. In some implementations, this method corresponds at least in part to any one or more of the methods described with respect to Figures 1 through 5.

[0089] At operation 712, the UE sends a request for neural network configuration to itself. In some examples, operation 712 is related to... Figure 1A element 112 and / or Figure 3 The element 312 corresponds to this.

[0090] At operation 714, the UE, in response to the request, receives an initial instruction regarding the neural network configuration of the second part of the neural network to be executed locally at that UE. In some examples, operation 714 is related to... Figure 1A element 114 and / or Figure 3 The elements 314 and / or 342 correspond to each other.

[0091] At operation 720, the UE receives the intermediate output of the neural network model generated by the first part of the neural network from the network node. In some examples, operation 720 is related to... Figure 1B Element 120A-M, Figure 4A Element 420 Figure 4B element 454 and / or Figure 4D The element 420 corresponds to this.

[0092] At operation 722, the UE processes the intermediate output through the second part of the neural network model to generate the model output of the neural network. In some examples, operation 722 is related to... Figure 1B Element 122A-M, Figure 4A Element 448 Figure 4B Elements 456 and / or Figure 4D The element 448 corresponds to this.

[0093] At operation 724, the UE generates an update for the second part of the neural network based on the model output and using a local update process. In some examples, operation 724 is... Figure 1C Elements 124A-M and / or Figures 5A to 5B The element 524 corresponds to this.

[0094] At operation 726, the UE sends an update to the network node. In some examples, operation 726 is related to... Figure 1C Elements 126A-M and / or Figures 5A to 5B The element 526 corresponds to this.

[0095] Figure 8A schematic example of a computing system / equipment 800 for performing any of the described methods, operations, or processes and / or for implementing any of the described systems, units, and / or apparatuses is shown. The computing system / equipment 800 shown is an example of a computing device or platform. Those skilled in the art will understand that other types of computing devices / systems / platforms can be used alternatively to implement the described methods, such as distributed computing systems. In some examples, the computing system / equipment 800 is a UE, a subsystem of the UE, a network node / entity, and / or a subsystem of a network node / entity.

[0096] Equipment (or system) 800 includes one or more processors 802 (e.g., CPUs). The one or more processors 802 control the operation of other components of the system / equipment 800. The system / equipment 800 may be part of a computing device, computing system, distributed computing system, cloud computing platform, etc., for implementing the functionality of the system / equipment and / or one or more methods / operations / processes as described. For example, the one or more processors 802 include general-purpose processors. The one or more processors 802 may be single-core or multi-core devices. The one or more processors 802 may include a central processing unit (CPU) or a graphics processing unit (GPU). Alternatively, the one or more processors 802 include dedicated processing hardware, such as a RISC processor or programmable hardware with embedded firmware. In some examples, multiple processors are included. In some embodiments, the one or more processors 802 are part of a distributed computing system such as a cloud computing system and / or cloud computing platform.

[0097] The system / equipment includes a memory system or memory 804, which includes working memory or volatile memory 806. One or more processors access the volatile memory 806 to process data and control the storage of data in the memory. The volatile memory 806 may include any type of RAM, such as static RAM (SRAM), dynamic RAM (DRAM), or may include flash memory, such as an SD card. In some embodiments, memory 804 and / or one or more volatile memories 806 comprise a plurality of memories 804 forming part of a distributed computing system such as a cloud computing system and / or a cloud computing platform.

[0098] The system / equipment includes non-volatile memory 808. Non-volatile memory 808 stores computer-readable instructions in the form of an operating or operating system instruction set 809a and / or computer-readable instructions 809b for controlling the operation of processor 802. When executed on one or more processors 802, these instructions cause the processor to implement the described methods, processes, operations, and / or functionalities. Depending on the application requirements, non-volatile memory 808 can be any type of memory, such as read-only memory (ROM), flash memory, SD drive, magnetic drive memory, or disk drive memory. In some embodiments, non-volatile memory 808 includes multiple non-volatile memories 808 forming part of a distributed computing system such as a cloud computing system and / or cloud computing platform.

[0099] One or more processors 802 are configured to execute operation instructions 809a and / or software instructions 809b, causing the system / equipment to perform any of the described methods or processes. Operation instructions 809a include, for example, code related to hardware components of the system / equipment 800 (i.e., drivers), and code related to the basic operation of the system / equipment 800. Generally, one or more processors 802 execute one or more of the operation instructions 809a and / or software instructions 809b, which are persistently or semi-persistently stored in non-volatile memory 808, using volatile memory 806 to temporarily store data generated during the execution of the operation instructions 809a and / or software instructions 809b.

[0100] In some implementations, one or more processors 802 are connected to a network interface 810, which includes a transmitter (TX) and a receiver (RX) for communicating with other equipment and systems over a network. In some examples, one or more processors 802 are connected to a user interface (UI) 812 for user or operator input to instruct or use the computing system and / or output data therefrom. In some examples, one or more processors 802 are connected to a display 814 for displaying output to a user or operator. At least one processor 802, along with at least one memory 804 and computer program code 809a, 809b, are arranged to cause the computing system 800 to perform at least at least one operation, method, and / or process, for example, as shown with respect to Figures 1 to 10. Figure 7 The schematic diagram, flowchart, or operation described in any of the above and its related features is disclosed.

[0101] Figure 9An example non-transitory computer-readable medium 900 according to some implementations is shown. The non-transitory medium 900 includes a computer-readable storage medium 902 and / or an input / output mechanism 904 to enable a computing system 800 to access the computer-readable medium 902. Although in this example the non-transitory medium is a USB stick, this is merely an example, and the invention is not limited thereto. Those skilled in the art will understand that the non-transitory medium 900 can be any other type of computer-readable medium or medium, such as a CD, DVD, USB stick, Blu-ray disc, flash drive, etc., and / or any other computer-readable medium as required by the application. The non-transitory medium 900 stores computer program code, causing the device to perform, for example, actions related to Figures 1 to 12. Figure 7 The flowcharts and schematic diagrams and their associated features disclose one or more of the methods, operations, or processors of any prior process.

[0102] The described methods or processes can be implemented as digital electronic circuit systems, integrated circuit systems, specially designed ASICs (Application-Specific Integrated Circuits), computer hardware, firmware, software, and / or combinations thereof. These may include computer program products (such as software stored on, for example, a disk, optical disk, memory, or programmable logic device) containing computer-readable instructions that, when executed by a computer, such as regarding… Figure 8 The described method causes the computer to perform one or more of the described methods.

[0103] Any system feature described can also be provided as a method or process feature, and vice versa. As used herein, apparatus plus functional features can alternatively be expressed according to their corresponding structures. Specifically, method aspects can be applied to system aspects, and vice versa.

[0104] Furthermore, any, some, and / or all features of one aspect may be applied in any suitable combination to any, some, and / or all features of any other aspect. It should also be understood that specific combinations of the various features described and defined in any aspect of this disclosure may be independently implemented and / or supplied and / or used.

[0105] Although several embodiments have been shown and described, those skilled in the art will understand that changes may be made to these embodiments without departing from the principles of this disclosure, the scope of which is defined in the claims and their equivalents.

Claims

1. A computer-implemented method executed by a computing system, the method comprising: Receive (312) a request for neural network configuration from the user equipment (UE); Based on the request, the UE is assigned (340) to a federated learning group comprising multiple UEs performing split execution of a neural network, such that the computing system executes a corresponding first part of the neural network for each UE in the federated learning group, and each UE in the federated learning group executes a corresponding second part of the neural network; Send (314) an initial instruction for the neural network configuration of the second part of the neural network to the plurality of UEs in the federated learning group; Receive corresponding update information for the corresponding second part of the neural network executed at the plurality of UEs from at least two of the plurality of UEs in the federated learning group; The parameters of the second part of the neural network are updated based on the updated information; as well as An update instruction for the updated parameters of the second part of the neural network is issued to the plurality of UEs in the federated learning group.

2. The method of claim 1, further comprising: Receive further requests for neural network configuration from the UE; as well as Based on the further request for neural network configuration, the further UEs are assigned to the federated learning group.

3. The method of any one of claims 1 or 2, wherein the method further comprises: Based on the input data from the UE, the first part of the neural network is executed to generate an intermediate neural network output; as well as This causes the intermediate neural network output to be transmitted to the UE.

4. The method of any of the preceding claims, wherein the update information includes one or more updated parameter values ​​of the second part of the neural network and / or one or more changes to the parameter values ​​of the second part of the neural network.

5. The method of any of the preceding claims, wherein the update information includes one or more gradients of the objective function relative to the parameters of the second part of the neural network.

6. The method of claim 5, further comprising updating one or more parameters of the first part of the neural network based on the one or more gradients of the objective function relative to the parameters of the second part of the neural network.

7. The method as described in any of the preceding claims, wherein: The request for neural network configuration includes one or more local UE condition capability information of the UE, auxiliary information of the UE, and / or indications of the local conditions of the UE; and Assigning the UE to a federated learning group includes determining that the local UE conditions fall within the range of capabilities and / or local conditions associated with the federated learning group.

8. The method of claim 7, wherein the one or more local UE conditions include one or more of the following: UE capability information, UE assistance information, UE processing availability and / or capability information; DNN-specific processing availability and / or capability information; UE power information; and / or UE thermal information.

9. The method of any of the preceding claims, wherein the initial indication of the neural network configuration is the same for each UE in the federated learning group.

10. The method according to any one of claims 1 to 7, wherein: The corresponding second part of the neural network executed by the plurality of UEs in the federated learning group has a set of one or more common layers across the federated learning group; and A proper subset of the UEs in the federated learning group has a corresponding second part of the neural network, the corresponding second part of the neural network including one or more additional neural network layers that are not present in the set of one or more common layers.

11. The method of any preceding claim, wherein the initial indication of the neural network configuration includes an indication of one or more of the following: the identity and / or type of one or more layers of the second portion of the neural network; multiple weights and / or bias values ​​of nodes of the second portion of the neural network; the identity of the federated learning group to which the UE is assigned; the split position of the neural network; one or more triggers for sending the update information to the computing system; the type and / or format of the update information; and / or the type of the local update process for generating the update information.

12. The method of any of the preceding claims, wherein the computing system is a cellular network entity.

13. The method of any of the preceding claims, wherein the corresponding update information is obtained from a local update procedure at each of the plurality of UEs in the federated learning group.

14. A computer-implemented method executed by a user equipment (UE), the method comprising: Send a request for neural network configuration to the network node; In response to the request, an initial instruction is received regarding the neural network configuration of the second part of the neural network to be executed locally at the UE; Receive intermediate outputs of the neural network model generated by the first part of the neural network from the network nodes; The intermediate output is processed by the second part of the neural network model to generate the model output of the neural network; Based on the model output, an update to the second part of the neural network is generated using a local update process; as well as The update is sent to the network node.

15. The method of claim 14, further comprising: Receive one or more federated updates for the second part of the neural network from the network node; as well as The received one or more federated updates are applied to the second part of the neural network model.

16. The method of any one of claims 14 or 15, wherein the method further comprises sending the input data of the first portion of the neural network to the network node before receiving the intermediate output.

17. The method of any one of claims 14 or 15, wherein the method further comprises, before receiving the intermediate output: The input data is processed by the third part of the neural network to generate an initial intermediate output; and The initial intermediate output is sent to the network node for input into the first part of the neural network.

18. The method of any one of claims 14 to 17, wherein the request for neural network configuration includes one or more local conditions of the UE.

19. The method of claim 18, wherein the one or more local conditions of the UE include one or more of the following: UE capability information, UE assistance information, UE processing availability and / or capability information; DNN-specific processing availability and / or capability information; UE power information; and / or UE thermal information.

20. The method of any one of claims 14 to 19, wherein the request for configuring the neural network specifies the splitting location of the neural network.

21. The method as described in any one of claims 14 to 20, The initial indication of the neural network configuration of the second part of the neural network includes one or more of the following: data indicating the architecture of the second part of the neural network; data indicating multiple weights of the second part of the neural network; and / or data indicating multiple biases of the second part of the neural network.

22. The method of any one of claims 14 to 21, wherein the initial indication of the neural network configuration of the second portion of the neural network includes one or more further neural network layers for execution by the UE, in addition to a common second portion of the neural network shared among multiple UEs in a federated learning group.

23. The method of any one of claims 14 to 22, wherein the initial indication of the neural network configuration of the second portion of the neural network includes federated learning configuration data.

24. The method of claim 23, wherein the federated learning configuration data includes one or more of the following: the identity of the federated learning group to which the UE is assigned; one or more triggers for sending the update information to the computing system; the type and / or format of the update information; and / or the type of the local update process used to generate the update information.

25. The method of any one of claims 14 to 24, wherein the update comprises one or more of the following: a set of updated network weights and / or biases of the second portion of the neural network; and / or a set of gradients of the loss function relative to the weights and / or biases of the second portion of the neural network.

26. An apparatus comprising: Network interface; One or more processors, said one or more processors being coupled to the network interface; as well as A memory that stores computer-readable instructions that, when executed by the one or more processors, cause the apparatus to perform the method as described in any one of claims 1 to 13.

27. A user equipment, comprising: One or more antennas; One or more processors; as well as A memory that stores computer-readable instructions that, when executed by the one or more processors, cause the user equipment to perform the method as described in any one of claims 14 to 25.

28. A computer program product comprising computer-readable instructions that, when executed by a means including a network interface and one or more processors coupled to the network interface, cause the means to perform the method as claimed in any one of claims 1 to 25.

Images