Method for selecting adaptive model parts for collaborative continual learning using UE groups
The method for selecting adaptive model parts in collaborative continual learning addresses the challenges of catastrophic forgetting and resource constraints by using UE groups to manage adaptive parts and latent replays, enabling efficient adaptation and participation in continual learning for edge devices.
Patent Information
- Application Number
- PCT/IB2024/054770
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-06-01
- Filing Date
- 2024-05-16
- Publication Date
- 2024-12-05
AI Technical Summary
Current machine learning models face challenges in continual learning at the edge, particularly due to catastrophic forgetting and inefficient retraining, which is computationally expensive and not suitable for edge devices with limited storage and processing capabilities.
The method involves selecting adaptive model parts for collaborative continual learning using UE groups, where a frozen part is shared among UEs, and adaptive parts are trained locally, with a network device managing the selection based on performance indicators and processing capabilities to minimize the number of models and latent replays required.
This approach allows for efficient adaptation to changing conditions while reducing the computational and storage burden on edge devices, preventing catastrophic forgetting and enabling UEs with limited resources to participate in continual learning by leveraging more capable peers.
Smart Images

Figure IB2024054770_05122024_PF_FP_ABST
Abstract
Description
Method For Selecting Adaptive Model Parts For Collaborative Continual Learning Using UE GroupsRELATED APPLICATION
[0001] This application claims priority to US provisional Application No. 63 / 470249 filed June 1, 2023, which is incorporated herein by reference in its entirety.TECHNICAL FIELD
[0002] The examples and non-limiting example embodiments relate generally to communications and, more particularly, to a method for selecting adaptive model parts for collaborative continual learning using UE groups.BACKGROUND
[0003] It is known to implement edge computing devices, such as terminal devices, in a communication network.SUMMARY
[0004] Various aspects of examples of the invention are set out in the claims.
[0005] In accordance with one aspect, a terminal device comprising: at least one processor; and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the terminal device at least to: transmit, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receive at least one adaptive part model of at least one terminal device; determine at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmit, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receive, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determine to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0006] In accordance with one aspect, a method comprising: transmitting, from a terminal device to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0007] In accordance with one aspect, a non-transitory computer-readable medium comprising program instructions stored thereon for performing at least the following: transmitting, from a terminal device to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0008] In accordance with one aspect, a terminal device comprising means for performing: transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintainat least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0009] In accordance with one aspect, a network device comprising: at least one processor; and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the network device at least to: receive, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determine at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receive, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determine a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmit, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0010] In accordance with one aspect, a method comprising: receiving, at a network device and from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at leastone adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0011] In accordance with one aspect, a non-transitory computer-readable medium comprising program instructions stored thereon for performing at least the following: receiving, at a network device and from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0012] In accordance with one aspect, a network device comprising means for performing: receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The foregoing aspects and other features are explained in the following description, taken in connection with the accompanying drawings.
[0014] FIG. 1 is a block diagram of one possible and non-limiting system in which the example embodiments may be practiced.
[0015] FIG. 2 shows a representation of a continuous learning model where some layers are frozen while others are trained on new data.
[0016] FIG. 3 is an illustration of a scenario where inference is performed at the UE side with UEs having ML models with frozen and adaptive parts.
[0017] FIG. 4 is a high level description of an AP selection method.
[0018] FIG. 5 shows signaling for the herein described ML model AP selection method in a scenario where there are two UEs.
[0019] FIG. 6 shows a performance threshold based criterion applied at the MLO / NW for dropping APs valid only for local UE data and not sufficiently generalized.
[0020] FIG. 7 shows criterion for cross-validation of APs.
[0021] FIG. 8 is an example apparatus configured to implement the examples described herein.
[0022] FIG. 9 shows a representation of an example of non-volatile memory media used to store instructions that implement the examples described herein.
[0023] FIG. 10 is an example method, based on the examples described herein.
[0024] FIG. 11 is an example method, based on the examples described herein.DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS
[0025] Turning to FIG. 1, this figure shows a block diagram of one possible and nonlimiting example in which the examples may be practiced. A user equipment (UE) 110, radio access network (RAN) node 170, and network element(s) 190 are illustrated. In the example of FIG. 1, the user equipment (UE) 110 is in wireless communication with a wireless network 100. A UE is a wireless device that can access the wireless network 100. The UE 110 includes one or more processors 120, one or more memories 125, and one or more transceivers 130 interconnected through one or more buses 127. Each of the one or more transceivers 130 includes a receiver, Rx, 132 and a transmitter, Tx, 133. The one or more buses 127 may be address, data, or control buses, and may include any interconnection mechanism, such as a series of lines on a motherboard or integrated circuit, fiber optics or other optical communication equipment, and the like. The one or more transceivers 130 are connected to one or more antennas 128. The one or more memories 125 include computer program code 123. The UE 110 includes a module 140, comprising one of or both parts 140-1 and / or 140- 2, which may be implemented in a number of ways. The module 140 may be implemented in hardware as module 140-1, such as being implemented as part of the one or more processors120. The module 140-1 may be implemented also as an integrated circuit or through other hardware such as a programmable gate array. In another example, the module 140 may be implemented as module 140-2, which is implemented as computer program code 123 and is executed by the one or more processors 120. For instance, the one or more memories 125 and the computer program code 123 may be configured to, with the one or more processors 120, cause the user equipment 110 to perform one or more of the operations as described herein. The UE 110 communicates with RAN node 170 via a wireless link 111.
[0026] The RAN node 170 in this example is a base station that provides access for wireless devices such as the UE 110 to the wireless network 100. The RAN node 170 may be, for example, a base station for 5G, also called New Radio (NR). In 5G, the RAN node 170 may be a NG-RAN node, which is defined as either a gNB or an ng-eNB. A gNB is a node providing NR user plane and control plane protocol terminations towards the UE, and connected via the NG interface (such as connection 131) to a 5GC (such as, for example, the network element(s) 190). The ng-eNB is a node providing E-UTRA user plane and control plane protocol terminations towards the UE, and connected via the NG interface (such as connection 131) to the 5GC. The NG-RAN node may include multiple gNBs, which may also include a central unit (CU) (gNB-CU) 196 and distributed unit(s) (DUs) (gNB-DUs), of which DU 195 is shown. Note that the DU 195 may include or be coupled to and control a radio unit (RU). The gNB-CU 196 is a logical node hosting radio resource control (RRC), SDAP and PDCP protocols of the gNB or RRC and PDCP protocols of the en-gNB that control the operation of one or more gNB-DUs. The gNB-CU 196 terminates the Fl interface connected with the gNB-DU 195. The Fl interface is illustrated as reference 198, although reference 198 also illustrates a link between remote elements of the RAN node 170 and centralized elements of the RAN node 170, such as between the gNB-CU 196 and the gNB- DU 195. The gNB-DU 195 is a logical node hosting RLC, MAC and PHY layers of the gNB or en-gNB, and its operation is partly controlled by gNB-CU 196. One gNB-CU 196 supports one or multiple cells. One cell may be supported with one gNB-DU 195, or one cell may be supported / shared with multiple DUs under RAN sharing. The gNB-DU 195 terminates the Fl interface 198 connected with the gNB-CU 196. Note that the DU 195 is considered to include the transceiver 160, e.g., as part of a RU, but some examples of this may have the transceiver 160 as part of a separate RU, e.g., under control of and connected to the DU 195. The RAN node 170 may also be an eNB (evolved NodeB) base station, for LTE (long term evolution), or any other suitable base station or node.
[0027] The RAN node 170 includes one or more processors 152, one or more memories 155, one or more network interfaces (N / W I / F(s)) 161, and one or more transceivers 160 interconnected through one or more buses 157. Each of the one or more transceivers 160 includes a receiver, Rx, 162 and a transmitter, Tx, 163. The one or more transceivers 160 are connected to one or more antennas 158. The one or more memories 155 include computer program code 153. The CU 196 may include the processor(s) 152, one or more memories 155, and network interfaces 161. Note that the DU 195 may also contain its own memory / memories and processor(s), and / or other hardware, but these are not shown.
[0028] The RAN node 170 includes a module 150, comprising one of or both parts 150-1 and / or 150-2, which may be implemented in a number of ways. The module 150 may be implemented in hardware as module 150-1, such as being implemented as part of the one or more processors 152. The module 150-1 may be implemented also as an integrated circuit or through other hardware such as a programmable gate array. In another example, the module 150 may be implemented as module 150-2, which is implemented as computer program code 153 and is executed by the one or more processors 152. For instance, the one or more memories 155 and the computer program code 153 are configured to, with the one or more processors 152, cause the RAN node 170 to perform one or more of the operations as described herein. Note that the functionality of the module 150 may be distributed, such as being distributed between the DU 195 and the CU 196, or be implemented solely in the DU 195.
[0029] The one or more network interfaces 161 communicate over a network such as via the links 176 and 131. Two or more gNBs 170 may communicate using, e.g., link 176. The link 176 may be wired or wireless or both and may implement, for example, an Xn interface for 5G, an X2 interface for LTE, or other suitable interface for other standards.
[0030] The one or more buses 157 may be address, data, or control buses, and may include any interconnection mechanism, such as a series of lines on a motherboard or integrated circuit, fiber optics or other optical communication equipment, wireless channels, and the like. For example, the one or more transceivers 160 may be implemented as a remote radio head (RRH) 195 for LTE or a distributed unit (DU) 195 for gNB implementation for 5G, with the other elements of the RAN node 170 possibly being physically in a different location from the RRH / DU 195, and the one or more buses 157 could be implemented in part as, for example, fiber optic cable or other suitable network connection to connect the other elements(e.g., a central unit (CU), gNB-CU 196) of the RAN node 170 to the RRH / DU 195. Reference 198 also indicates those suitable network link(s).
[0031] A RAN node / gNB can comprise one or more TRPs to which the methods described herein may be applied. FIG. 1 shows that the RAN node 170 comprises two TRPs, TRP 51 and TRP 52. The RAN node 170 may host or comprise other TRPs not shown in FIG. 1.
[0032] A relay node in NR is called an integrated access and backhaul node. A mobile termination part of the IAB node facilitates the backhaul (parent link) connection. In other words, the mobile termination part comprises the functionality which carries UE functionalities. The distributed unit part of the IAB node facilitates the so called access link (child link) connections (i.e. for access link UEs, and backhaul for other IAB nodes, in the case of multi-hop IAB). In other words, the distributed unit part is responsible for certain base station functionalities. The IAB scenario may follow the so called split architecture, where the central unit hosts the higher layer protocols to the UE and terminates the control plane and user plane interfaces to the 5G core network.
[0033] It is noted that the description herein indicates that “cells” perform functions, but it should be clear that equipment which forms the cell may perform the functions. The cell makes up part of a base station. That is, there can be multiple cells per base station. For example, there could be three cells for a single carrier frequency and associated bandwidth, each cell covering one-third of a 360 degree area so that the single base station’s coverage area covers an approximate oval or circle. Furthermore, each cell can correspond to a single carrier and a base station may use multiple carriers. So if there are three 120 degree cells per carrier and two carriers, then the base station has a total of 6 cells.
[0034] The wireless network 100 may include a network element or elements 190 that may include core network functionality, and which provides connectivity via a link or links 181 with a further network, such as a telephone network and / or a data communications network (e.g., the Internet). Such core network functionality for 5G may include location management functions (LMF(s)) and / or access and mobility management function(s) (AMF(S)) and / or user plane functions (UPF(s)) and / or session management function(s) (SMF(s)). Such core network functionality for LTE may include MME (mobility management entity) / SGW (serving gateway) functionality. Such core network functionality may include SON (self- organizing / optimizing network) functionality. These are merely example functions that maybe supported by the network element(s) 190, and note that both 5G and LTE functions might be supported. The RAN node 170 is coupled via a link 131 to the network element 190. The link 131 may be implemented as, e.g., an NG interface for 5G, or an SI interface for LTE, or other suitable interface for other standards. The network element 190 includes one or more processors 175, one or more memories 171, and one or more network interfaces (N / W I / F(s)) 180, interconnected through one or more buses 185. The one or more memories 171 include computer program code 173. Computer program code 173 may include SON and / or MRO functionality 172.
[0035] The wireless network 100 may implement network virtualization, which is the process of combining hardware and software network resources and network functionality into a single, software-based administrative entity, or a virtual network. Network virtualization involves platform virtualization, often combined with resource virtualization. Network virtualization is categorized as either external, combining many networks, or parts of networks, into a virtual unit, or internal, providing network-like functionality to software containers on a single system. Note that the virtualized entities that result from the network virtualization are still implemented, at some level, using hardware such as processors 152 or 175 and memories 155 and 171, and also such virtualized entities create technical effects.
[0036] The computer readable memories 125, 155, and 171 may be of any type suitable to the local technical environment and may be implemented using any suitable data storage technology, such as semiconductor based memory devices, flash memory, magnetic memory devices and systems, optical memory devices and systems, non-transitory memory, transitory memory, fixed memory and removable memory. The computer readable memories 125, 155, and 171 may be means for performing storage functions. The processors 120, 152, and 175 may be of any type suitable to the local technical environment, and may include one or more of general purpose computers, special purpose computers, microprocessors, digital signal processors (DSPs) and processors based on a multi-core processor architecture, as nonlimiting examples. The processors 120, 152, and 175 may be means for performing functions, such as controlling the UE 110, RAN node 170, network element(s) 190, and other functions as described herein.
[0037] In general, the various example embodiments of the user equipment 110 can include, but are not limited to, cellular telephones such as smart phones, tablets, personal digital assistants (PDAs) having wireless communication capabilities, portable computers havingwireless communication capabilities, image capture devices such as digital cameras having wireless communication capabilities, gaming devices having wireless communication capabilities, music storage and playback devices having wireless communication capabilities, internet appliances including those permitting wireless internet access and browsing, tablets with wireless communication capabilities, head mounted displays such as those that implement virtual / augmented / mixed reality, as well as portable units or terminals that incorporate combinations of such functions. The UE 110 can also be a vehicle such as a car, or a UE mounted in a vehicle, a UAV such as e.g. a drone, or a UE mounted in a UAV. The user equipment 110 may be terminal device, such as mobile phone, mobile device, sensor device etc., the terminal device being a device used by the user or not used by the user.
[0038] UE 110, RAN node 170, and / or network element(s) 190, (and associated memories, computer program code and modules) may be configured to implement (e.g. in part) the methods described herein, including a method for selecting adaptive model parts for collaborative continual learning using UE groups. Thus, computer program code 123, module 140-1, module 140-2, and other elements / features shown in FIG. 1 of UE 110 may implement user equipment related aspects of the examples described herein. Similarly, computer program code 153, module 150-1, module 150-2, and other elements / features shown in FIG. 1 of RAN node 170 may implement gNB / TRP related aspects of the examples described herein. Computer program code 173 and other elements / features shown in FIG. 1 of network element(s) 190 may be configured to implement network element related aspects of the examples described herein.
[0039] Having thus introduced a suitable but non-limiting technical context for the practice of the example embodiments, the example embodiments are now described with greater specificity.
[0040] Continual learning (CL) is to freeze part of the ML model and the other part is adapted for retraining. This approach is good for hardware limited devices such as mobile user terminals. The new knowledge is applied only to the adapted layers while frozen layers continue holding the old knowledge. Continual learning may be applicable to 3GPP UE with appropriate modifications specific to the UE chipset platforms.
[0041] Training a machine learning (ML) model on the edge (e.g., on light computing devices such as smartphones) is highly desirable in several applications where privacy, lackof network connection and fast adaptation are real constraints. With improving UE capabilities, it is becoming increasingly possible to train ML models at the edge.
[0042] When an ML model is trained and deployed at the network edge, it still needs to improve its learning over a period of time due to changing learning conditions. Retraining the ML model from scratch is one option, however, such a learning scheme assumes that all samples are available during the training phase and, therefore, requires the retraining of the network parameters on the entire dataset in order to adapt to changes in the data distribution. When trained on sequential tasks with samples becoming progressively available over time, the performance of conventional neural network models significantly decreases on previously learned tasks as new tasks are learned. Retraining from scratch pragmatically addresses catastrophic forgetting, however this methodology is very inefficient and hinders the learning of novel data in real time as retraining a full ML model is computationally quite expensive and may not be a suitable option for edge learning.
[0043] It is therefore important to study novel techniques to manage storage and computational requirements carefully as edge devices cannot have the same storage and computation capabilities as network servers can have with training data stored in the cloud.
[0044] Continual learning is defined as a fundamental idea in machine learning in which models continuously learn and evolve based on the input of increasing amounts of data while retaining previously learned knowledge. The continual learning (CL) concept falls under the categories stateful and online training.
[0045] Stateful (re)training (training or retraining) is when the model continues training on new data (fine-tuning). Stateful training and retraining mostly refers to data iteration i.e., using the same model architecture and features but training from new data and allowing updating (fine-tuning) the model with less data. Stateful (re)training uses automated training procedures.
[0046] Stateless (re)training is when the model is trained from scratch each time. Stateless training or retraining involves changing the model architecture or adding a new feature, requiring training the new model from scratch i.e., a model iteration. Stateless (re)training can use manual or automated training procedures.
[0047] The ability to continually learn over time by accommodating new knowledge whileretaining previously learned models is referred to as continual or lifelong learning. Such a continuous learning task has represented a long-standing challenge for machine learning and neural networks.
[0048] The main issue of computational models regarding lifelong learning is that they are prone to catastrophic forgetting or catastrophic interference, i.e., training a model with new information interferes with previously learned knowledge. This phenomenon typically leads to an abrupt performance decrease or, in the worst case, to the old knowledge being completely overwritten by the new knowledge. This forgetting occurs because parametric models, including neural networks, assume that data is independent and identically distributed (iid). In offline settings, models can simulate the notion of iid experiences by shuffling data. However, in continual learning settings, the data stream is evolving in a non-iid manner over time, which causes catastrophic forgetting of previous knowledge.
[0049] Latent Replays
[0050] Referring to FIG. 2, in deep neural networks, the layers close to the input 204, such as input layer 206, Hl 208, H2 210, and H3 212, usually perform low-level feature extraction and, after a proper pre-training on a large dataset, their weights are quite stable and reusable across applications. On the other hand, higher layers, such as HN 214, HN+1 216, and output layer 218, tend to extract more discriminant features and their tuning is often important to maximize accuracy.
[0051] To minimize (or remove) catastrophic forgetting, a latent replay technique is used where, replay mechanisms are used (such as latent replay 202) for storing a representation 203 of previous data to mix with new data when updating the network. This representation 203 can be in form of past raw (input) data, or some weights or activation functions representing features. To keep the representation 203 stable and the stored replays valid, a part (layers below the latent replay such as layers 206, 208, 210, 212) of the neural network is frozen 220 for retraining while leaving the layers above (214, 216, 218) free to learn (retrain during adaptive stage 222). This helps avoid catastrophic forgetting at low memory cost. Thus, FIG. 2 illustrates an example of CL where some layers are frozen for retraining while other layers perform full retraining.
[0052] There are two extreme strategies for ML model re-training when inference is performed at the edge or UE side, namely individual models and federated learning.
[0053] 1. Individual models have a different set of conditions for the same task, (e.g., CSI compression, CSI prediction, beam management, handover management in different cells). Disadvantages of individual models include there being too many model downloads at the UE if UE input conditions change fast and the model is trained at the NW.
[0054] 2. In federated learning, one model is valid for a larger set of input conditions. Disadvantages of federated learning include that one model may not fit all conditions if conditions change substantially in only one subgroup while conditions are relatively stable in other subgroups. A very large retraining effort is required unnecessary to retrain an FL model.
[0055] There is a trade-off in which continual learning at UE side is implemented with one (potentially larger) part fixed and one (potentially smaller) part retrainable.
[0056] However, CL helps if new conditions are not very different from the ones the model was originally trained on. In a typical scenario, input conditions may change slightly as compared to original conditions, but this change is not homogeneous across all UE groups. In this scenario, instead of retraining a full model, CL can be applied by allowing to have different adaptive parts (AP) for different UEs (or group of UEs) based on their input conditions to adapt to different conditions, resulting in lower retraining effort as compared to full retraining. However, this creates a problem of having too many ML models in the system with same fixed parts and different adaptive parts, an undesirable feature for model management, as all or too many UEs need to process latent replays with high processing power.
[0057] The examples described herein provide a solution to the problem of how to select a small number of adaptive parts in CL for different UEs with the same frozen part in order to keep complexity of the solution (in terms of number of different adaptive parts and different replay memories) reduced while still ensuring an accurate ML model with regard to changing conditions, such as radio and traffic conditions.
[0058] On one hand, this allows generalization of the ML model to a set of local UE conditions while still keeping intact main characteristics of the ML model through the frozen part; on the other hand, it allows the UE with temporally inadequate CL support (processing, memory) for usage of latent replays to participate in the CL procedure, with indirect help from other more capable UEs, and avoids the ‘catastrophic forgetting’ problem.
[0059] There are fundamental techniques to perform continual or lifelong learning. However, there is no mechanism known on how to coordinate CL procedures for distributed edge nodes in a wireless network.
[0060] As illustrated in FIG. 3, the scenario discussed is based on the following assumptions. The adaptive stage architecture is the same for all UEs (110, 110-2, 110-3), gNB 170 and / or the network (170, 190) has a fixed / frozen part 310 (which is downloaded by each UE (110, 110-2, 110-3)) and inference is performed at the UE side, and the UE (110, 110-2, 110-3) can store adaptive parts (320, 330) and latent replays 302 locally for CL. Thus, FIG. 3 is an illustration of a scenario where inference is performed at the UE side with UEs having ML models with frozen and adaptive parts. The frozen part 310 is the same for all UEs (110, 110-2, 110-3) but the APs (320, 330) can be different.
[0061] Referring to FIG. 3 and to FIG. 4, the solution described herein comprises items 1- 8:
[0062] 1 (410). If the ML model is not performing well, an ML model managing entity, called the ML model Orchestrator (MLO) 340 at the gNB / NW (170, 190) decides to trigger CL at the UE (110, 110-2, 110-3).
[0063] 2 (410). Based on UE input conditions (radio and non-radio), different UEs (110, 110-2, 110-3) may start training their different adaptive parts (320, 330) on the local data available to them.
[0064] 3 (410). UEs feedback their AP parameters to the NW for validation. They can optionally send information regarding their ability to process latent replays.
[0065] 4 (420). The MLO or serving gNB validates performance of all (or most) models on a validation data set and selects a subset of models. This step has multiple parts: monitoring, and validation, and decision or action to mitigate performance degradation.
[0066] 5 (430). Cross UE Validation: The UE’s APs which are determined as meeting the selected performance criterion in step 3 (use case specific, scenario and configuration specific) are further evaluated by other UEs with their local input data sets and performance results are fed back to the NW.
[0067] 6 (430). The corresponding UEs are selected to participate in the collaborative CLprocedure.
[0068] 7 (440). Model Similarity Check: The MLO further tries to select the APs usable by the largest number of participating UEs based on a ‘performance similarity criterion’ described herein. To reduce the number of models in operation, models very similar in performance are merged such that only one of them is retained in the system (and discarding one of them). In another embodiment, instead of functional performance, model parameter comparison can be performed to evaluate similarity between two models.
[0069] 8. A limited number of UE AP models are selected in step 6 and all participating UEs are informed of the AP models to be used. All the UEs using the same APs do not need to store latent replays locally, and only a subset of UEs can keep latent replays with them. This helps UEs to perform CL even if they are temporarily low on compute power (processing and / or memory) to process latent replays.
[0070] Thus, FIG. 4 depicts a high-level flow chart for the herein described AP selection method. Initially M UEs’ APs are trained locally at participating UEs in the collaborative CL procedure and inference results and / or performance KPIs are sent to the gNB where the MLO further processes these results to decide how many distinct ML models are sufficient to be used in the system without compromising much on performance.
[0071] In particular, at 410 M adaptive parts are received at the MLO from different UEs. At 420, models with poor performance on validation data are eliminated. At 430, models suited to a very small number of UEs (cross validation) are eliminated. At 440, models based on similarity in performance can be eliminated (with only distinct K<M models retained).
[0072] FIG. 5 shows core signaling steps for implementing the herein described method for performing CL with one frozen part and a distinct set of adaptive parts where selection of adaptive parts depends on similarity of functional performance of the ML model. Examples of ML functionality could be ML assisted handover, ML assisted uplink power control, ML assisted beam management, etc. with key performance indicator (KPIs) including indicators such as successful handover rate for ML assisted handover, and average throughput measure for ML assisted uplink power control, packet delays, service QoS, or beam selection accuracy.
[0073] Scenario
[0074] It is assumed that UEs are running ML models at the edge and performing local inference to assist a communication functionality as described above. Typically, ML model inference results deteriorate over time due to possible change in inference input conditions and the model needs retraining. However, performing model training from a scratch is not ideal and continuous learning is employed as explained herein. CL can be performed with one frozen part and one AP but it is possible to use distinct APs, each trained on local UE data. On one side, this will improve UE performance (more suitable to local conditions) but on the other side, it will make the ML model management difficult as not all UEs may have enough memory for storing latent replays and UEs will need to download different APs when conditions change (e.g., due to mobility).
[0075] Described herein is a method to select a set of distinct APs for CL, which allows a UE to account for their local input conditions. UEs with insufficient memory to support a local AP can still benefit from the herein described method.
[0076] Method
[0077] The herein described method comprises the following main steps as shown in FIG. 5, where UE 1 (110) and UE 2 (110-2) are two UEs performing local inference on ML models. These ML models can be assumed to be fully trained at the NW side or locally at the UE side but it is assumed that all UEs perform inference using the same ML models at the beginning.
[0078] 0. UEs send feedback on performance results (periodically or triggered based) by sending performance KPIs to the network.
[0079] 1. On the network side (e.g. at gNB 170), a model management orchestrator (MLO) evaluates functional KPIs and decides if ML model retraining is required or not.
[0080] 2. If model CL training is required, the NW triggers CL training by sending a configuration message. This message may include information such as the type of training is CL, and which adaptive layers are retrained on new data. It is assumed that UEs already have information on latent replays before starting CL. Alternatively, when the NW does not have full information about the AP in the UE, the NW can still trigger the CL in the UEs without any additional configuration information.
[0081] 3. UEs re-train their APs based on new data and implementation specific conditions.After the re-training procedure ends: 3a. UEs send AP model parameters to the NW, 3b. Assuming that some of UEs may not have temporary ability to process latent replays for CL, those UEs optionally can inform the NW. This information is used later by the MLO at the NW side.
[0082] 4. MLO uses validation data collected over a larger distribution to validate the UE AP part. As shown in FIG. 6, if the UE AP part (combined with frozen layers which have been assumed to be known at the NW side) performance is below a functional KPI threshold (determined at 630), at 650 the model is dropped and not validated for further use in the system. All the models that are validated are processed further in the next step.
[0083] 5 and 5.5. Cross UE Validation. Though validation on a validation data set is already performed in step 4 above, data distribution over training data space might have been different from the distribution represented by the NW validation data set. Therefore, introduced herein is a further validation on a ‘cross UE’ data set, which requires UEs to validate all (or most) other APs and send at least one functional performance KPI back to the NW. An MLO policy determines how many UEs perform this cross validation for each AP. One embodiment of this procedure is described with reference to FIG. 7. At step 5, the UE2 110-2 receives the model to perform inference, and at step 5.5, the UE2 110-2 performs the inference.
[0084] 6. The UE (e.g. UE2 110-2) reports performance results and / or indicators based on cross validation for each of the APs it receives after performing local inference on its local training data set.
[0085] 7. Based on cross validation in steps 5, 5.5, and 6, the MLO discards some of the models failing cross validation as they are difficult to generalize across many UEs.
[0086] 8. Model Similarity Check. Based on cross validation, the UE AP models which can be generalized to a larger group of UEs are accepted.
[0087] 8a. UEs showing similar performance on particular APs are grouped together. This is further explained below. In each group, not all UEs need to process the latent replay used to avoid catastrophic forgetting in CL, where if some of the UEs do not possess compute power (already communicated to MLO in step 3(b)), only one (or smaller group of UEs) can maintain latent replays locally and help other UEs if needed to avoid catastrophic forgetting when CL is performed next time (triggered by the NW).
[0088] 8b. In an additional embodiment, if all of the UEs grouped together in a group do not have (e.g. temporarily) sufficient compute capability, their group can be merged with some other group as CL cannot be performed without processing latent replays.
[0089] 9. UEs selected to process latent replay data are informed and group AP model parameters are transferred to all UEs.
[0090] 10. After receiving AP parameters in Step 9, 10a. UEs with sufficient processing power and memory maintain latent replays for further CL, and 10b. UEs with insufficient processing power or memory are not selected to process latent replays and just receive AP parameters.
[0091] Two UEs are shown in FIG. 5 for illustration. The signaling can be extended to scenarios with more than two UEs.
[0092] FIG. 6 shows a performance threshold based criterion applied at the MLO or NW for dropping APs valid only for local UE data and that are not sufficiently generalized.
[0093] At 610, the UE sends feedback of one or more performance KPIs and adaptive model parts based on local data. At 620, inference is started based on a validation data set. At 630, it is determined whether the functional performance of the one or more key performance indicators is above a threshold. If at 630 it is determined that the functional performance indicated by the one or more key performance indicators is above the threshold (e.g. “Yes”), at 640 the models are saved for further processing. If at 630 it is determined that the functional performance indicated by the one or more key performance indicators is below the threshold (e.g. “No”), at 650 the model is determined to be not reliable for further processing.
[0094] Example Embodiment: Criterion for Cross UE Validation and Model Similarity
[0095] FIG. 7 shows in detail how the model similarity check and selection is made using cross UE validation as shown in FIG. 5 (steps 6, 7, 8(a)). At 701, UEs send feedback of performance of one or more KPIs based on selected AP models. At 710, the model similarly check and selection using cross UE validation starts.
[0096] For a particular model, say Ml, a cross validation is performed on local data by X UEs. There is a minimum performance threshold, e.g., % of successful handover decisions,set at the MLO. At 720, a model will be accepted to be used in the system (as similar models) if it has better performance than a threshold, called Th2, for at least N% of UEs who cross validated this model on local data sets. This condition signifies that an AP model can certainly (or most likely) be used by many UEs in the system and not all of them need to train their own AP. Here, N is a configurable parameter that the MLO selects based on UEs using an ML model for a particular functionality. If a large number of UEs are present in the system, a reasonably low value should be chosen as a large value would request a large number of UEs to perform cross validation on other UE adaptive parts and it is a considerable overhead. On the other hand, for a smaller number of UEs, a reasonably larger value of N can be selected to ensure that cross validated AP parts can be generalized to most local conditions (if they pass the cross validation test). As indicated at 730, a model is determined to be not reliable if it does not have better performance than the threshold Th2 for at least N% of UEs who cross validated this model on local data sets.
[0097] There could be many APs which pass the cross-validation condition but do not show similar performance. Then, in the next step (740), mean performance of all such models is evaluated and (1-3):
[0098] 1) If Diff(Mean_model_performance between two APs)>delta (determined at 740), then the models are sufficiently different, and both should be retained (760). Diff (.) is a function used to determine a difference between at least two items, such as two values, parameters, or models. For example, Diff (3, 1) may be result in 3-1 = 2.
[0099] 2) If the above condition is false, then any of the APs can be discarded (750). After performing this step, only distinct APs are left.
[0100] 3) The next step (770) is to group UEs which can use one particular AP part. One example embodiment is to group UEs such that their performance variance is below delta2 for a particular model, i.e., Var(model_performance for a UE for a particular model)<delta2, where Var(.) is a function is to determine variance of a model’s performance based on a performance key performance indicator (KPI).
[0101] Thus, FIG. 7 shows criterion for cross-validation of APs. The method indirectly hints that UEs experience similar local conditions. At 780, the method stops when all the UEs are grouped.
[0102] Alternate Embodiment:
[0103] Instead of evaluating similarity of two APs based on their functionality performance, an alternative is to directly evaluate similarity between parameters of two APs and if a similarity metric (e.g., Euclidean distance) is below a threshold, the models are considered to be similar. In another embodiment, a combination of both a direct parameter similarity check and a performance similarity check can be used.
[0104] Additional Embodiment:
[0105] If a large number of UEs are using one particular AP (and ML model), and parameters delta, delta2 are carefully selected, then it is expected that each group of UEs using the same AP will have enough number of UEs with capability to process latent replays. However, if as an extreme case, there is a group with UEs with no UE able to process latent replays, two solutions can be provided (1-2):
[0106] 1. Latent replays for such group are stored at the MLO side and provided to the UEs when needed for the CL procedure.
[0107] 2. The group of UEs is merged with another UE group using a different AP. Performance may suffer, but the MLO can assign UEs to the group which is closest in performance to this particular group to minimize performance loss.
[0108] Note that latent replays can be provided to the UEs using existing side link transmission or unicast downlink transmission.
[0109] Advantages and technical effects of the herein described methods include that distributed CL with AP selection allows generalization of the model to a set of local conditions while still keeping intact main characteristics of the ML model through the frozen part. The herein described method allows UEs with limited memory to store latent replays to participate in CL as other UEs can help with storage to avoid catastrophic forgetting.
[0110] The examples described herein are related to AI / ML solutions for the air-interface which utilize online training and adaptations. As such these solutions are likely to be in the scope of NR Release 19+ and 6G. 3GPP is to promote the introduction of RL / CL concepts for Release 19. 6G ML solutions are more likely to adopt RL / CL / adaptation principles.
[0111] FIG. 8 is an example apparatus 800, which may be implemented in hardware, configured to implement the examples described herein. The apparatus 800 comprises at least one processor 802 (e.g. an FPGA and / or CPU), one or more memories 804 including computer program code 805, the computer program code 805 having instructions to carry out the methods described herein, wherein the at least one memory 804 and the computer program code 805 are configured to, with the at least one processor 802, cause the apparatus 800 to implement circuitry, a process, component, module, or function (implemented with control module 806) to implement the examples described herein, including a method for selecting adaptive model parts for collaborative continual learning using UE groups. The memory 804 may be a non-transitory memory, a transitory memory, a volatile memory (e.g. RAM), or a non-volatile memory (e.g. ROM). CL 830 of the control module implements the herein described aspects related to selection of adaptive model parts for collaborative continual learning using UE groups.
[0112] The apparatus 800 includes a display and / or RO interface 808, which includes user interface (UI) circuitry and elements, that may be used to display aspects or a status of the methods described herein (e.g., as one of the methods is being performed or at a subsequent time), or to receive input from a user such as with using a keypad, camera, touchscreen, touch area, microphone, biometric recognition, one or more sensors, etc. The apparatus 800 includes one or more communication e.g. network (N / W) interfaces (I / F(s)) 810. The communication I / F(s) 810 may be wired and / or wireless and communicate over the Internet / other network(s) via any communication technique including via one or more links 824. The link(s) 824 may be the link(s) 131 and / or 176 from FIG. 1. The link(s) 131 and / or 176 from FIG. 1 may also be implemented using transceiver(s) 816 and corresponding wireless link(s) 826. The communication I / F(s) 810 may comprise one or more transmitters or one or more receivers.
[0113] The transceiver 816 comprises one or more transmitters 818 and one or more receivers 820. The transceiver 816 and / or communication I / F(s) 810 may comprise standard well-known components such as an amplifier, filter, frequency-converter, (de)modulator, and encoder / decoder circuitries and one or more antennas, such as antennas 814 used for communication over wireless link 826.
[0114] The control module 806 of the apparatus 800 comprises one of or both parts 806-1 and / or 806-2, which may be implemented in a number of ways. The control module 806 maybe implemented in hardware as control module 806-1, such as being implemented as part of the one or more processors 802. The control module 806-1 may be implemented also as an integrated circuit or through other hardware such as a programmable gate array. In another example, the control module 806 may be implemented as control module 806-2, which is implemented as computer program code (having corresponding instructions) 805 and is executed by the one or more processors 802. For instance, the one or more memories 804 store instructions that, when executed by the one or more processors 802, cause the apparatus 800 to perform one or more of the operations as described herein. Furthermore, the one or more processors 802, the one or more memories 804, and example algorithms (e.g., as flowcharts and / or signaling diagrams), encoded as instructions, programs, or code, are means for causing performance of the operations described herein.
[0115] The apparatus 800 to implement the functionality of control 806 may be UE 110, RAN node 170 (e.g. gNB), or network element(s) 190. Thus, processor 802 may correspond to processor(s) 120, processor(s) 152 and / or processor(s) 175, memory 804 may correspond to one or more memories 125, one or more memories 155 and / or one or more memories 171, computer program code 805 may correspond to computer program code 123, computer program code 153, and / or computer program code 173, control module 806 may correspond to module 140-1, module 140-2, module 150-1, and / or module 150-2, and communication I / F(s) 810 and / or transceiver 816 may correspond to transceiver 130, antenna(s) 128, transceiver 160, antenna(s) 158, N / W I / F(s) 161, and / or N / W I / F(s) 180. Alternatively, apparatus 800 and its elements may not correspond to either of UE 110, RAN node 170, or network element(s) 190 and their respective elements, as apparatus 800 may be part of a self- organizing / optimizing network (SON) node or other node, such as a node in a cloud.
[0116] The apparatus 800 may also be distributed throughout the network (e.g. 100) including within and between apparatus 800 and any network element (such as a network control element (NCE) 190 and / or the RAN node 170 and / or the UEs, such as UE 110, UE 110-2, and UE 110-3).
[0117] Interface 812 enables data communication and signaling between the various items of apparatus 800, as shown in FIG. 8. For example, the interface 812 may be one or more buses such as address, data, or control buses, and may include any interconnection mechanism, such as a series of lines on a motherboard or integrated circuit, fiber optics or other optical communication equipment, and the like. Computer program code (e.g.instructions) 805, including control 806 may comprise object-oriented software configured to pass data or messages between objects within computer program code 805. The apparatus 800 need not comprise each of the features mentioned, or may comprise other features as well. The various components of apparatus 800 may at least partially reside in a common housing 828, or a subset of the various components of apparatus 800 may at least partially be located in different housings, which different housings may include housing 828.
[0118] FIG. 9 shows a schematic representation of non-volatile memory media 900a (e.g. computer / compact disc (CD) or digital versatile disc (DVD)) and 900b (e.g. universal serial bus (USB) memory stick) and 900c (e.g. cloud storage for downloading instructions and / or parameters 902 or receiving emailed instructions and / or parameters 902) storing instructions and / or parameters 902 which when executed by a processor allows the processor to perform one or more of the steps of the methods described herein.
[0119] FIG. 10 is an example method 1000, based on the example embodiments described herein. At 1010, the method includes transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay. At 1020, the method includes receiving, from the network device, at least one adaptive part model of at least one terminal device. At 1030, the method includes determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device. At 1040, the method includes transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device. At 1050, the method includes receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator. At 1060, the method includes determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter. Method 1000 may be performed with UE 110, UE 110-2, UE 110-3, or apparatus 800.
[0120] FIG. 11 is an example method 1100, based on the example embodiments described herein. At 1110, the method includes receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay. At 1120, the method includes determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device. At 1130, the method includes transmitting, to the at least one terminal device, the at least one adaptive part model of the atleast one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold. At 1140, the method includes receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device. At 1150, the method includes determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold. At 1160, the method includes determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device. At 1170, the method includes transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device. At 1180, the method includes transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold. Method 1100 may be performed with RAN node 170 (which RAN node 170 may be a gNB or comprise an MLO) or apparatus 800.
[0121] The following examples are provided and described herein.
[0122] Example 1. A terminal device including: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the terminal device at least to: transmit, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receive, from the network device, at least one adaptive part model of at least one terminal device; determine at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmit, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receive, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determine to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0123] Example 2. The terminal device of example 1, wherein the at least one latent replay of the at least one adaptive part model of the at least one terminal device comprises a stored representation of the at least one adaptive part model of the at least one terminal device.
[0124] Example 3. The terminal device of any of examples 1 to 2, wherein the determining of the at least one performance indicator of the at least one adaptive part model of the at least one terminal device comprises performing inference using the at least one adaptive part model of the at least one terminal device, wherein the inference is performed with a training set local to the terminal device.
[0125] Example 4. The terminal device of any of examples 1 to 3, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from the network device, an indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device; wherein the determining to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group is based on the indication received from the network device to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device.
[0126] Example 5. The terminal device of example 4, wherein the indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device is received from the network device at least partially based on the transmitted indication that the terminal device is capable of maintaining the latent replay.
[0127] Example 6. The terminal device of any of examples 1 to 5, wherein the determining to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group is based at least partially on the transmitted indication that the terminal device is capable of maintaining the latent replay.
[0128] Example 7. The terminal device of any of examples 1 to 6, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from the network device, the at least one latent replay of the at least one adaptive part model of the at least one terminal device, when a processing capability of the terminal device is less than a processing capability threshold.
[0129] Example 8. The terminal device of any of examples 1 to 7, wherein the instructions,when executed by the at least one processor, cause the terminal device at least to: transmit, to a terminal device within the group of the at least one terminal device, the at least one latent replay of the at least one adaptive part model of the at least one terminal device, when a processing capability of the terminal device is less than a processing capability threshold, and when a processing capability of the terminal device within the group is greater than the processing capability threshold.
[0130] Example 9. The terminal device of any of examples 1 to 8, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from a terminal device within the group of the at least one terminal device, the at least one latent replay of the at least one adaptive part model of the at least one terminal device, when a processing capability of the terminal device is greater than a processing capability threshold, and when a processing capability of the terminal device within the group is less than the processing capability threshold.
[0131] Example 10. The terminal device of any of examples 7, 8, or 9, wherein the processing capability of the terminal device comprises a memory capability of the terminal device, the processing capability of the terminal device comprises a memory capability of the terminal device, and the processing capability threshold comprises a memory capability threshold.
[0132] Example 11. The terminal device of any of examples 1 to 10, wherein the at least one adaptive part model comprises a machine learning model.
[0133] Example 12. The terminal device of any of examples 1 to 11, wherein the terminal device comprises the at least one terminal device.
[0134] Example 13. The terminal device of any of examples 1 to 12, wherein the at least one performance indicator is based on at least one of: a handover success rate, a handover failure rate, at least one throughput measure, at least one packet delay, a quality of service indicator, or beam selection accuracy.
[0135] Example 14. A network device including: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the network device at least to: receive, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determine at least one firstperformance indicator of at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receive, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determine a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmit, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0136] Example 15. The network device of example 14, wherein the at least one latent replay of the at least one adaptive part model of the at least one terminal device comprises a stored representation of the at least one adaptive part model of the at least one terminal device.
[0137] Example 16. The network device of any of examples 14 to 15, wherein the instructions, when executed by the at least one processor, cause the network device at least to: transmit, to the at least one terminal device, the indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device based on the receiving, from the at least one terminal device, the indication that the at least one terminal device is capable of maintaining the latent replay.
[0138] Example 17. The network device of any of examples 14 to 16, wherein the instructions, when executed by the at least one processor, cause the network device at least to: transmit, to the at least one terminal device, the at least one latent replay of the at least one adaptive part model parameter of the at least one terminal device, when the processing capability of the at least one terminal device is less than the processing capability threshold.
[0139] Example 18. The network device of any of examples 14 to 17, wherein the instructions, when executed by the at least one processor, cause the network device at least to: merge the group of the at least one terminal device with another group of the at least one terminal device, when the at least one processing capability of the at least one terminal device of the group is less than the processing capability threshold, and when the at least one processing capability of the at least one terminal device of the another group is greater than the processing capability threshold.
[0140] Example 19. The network device of any of examples 14 to 18, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to accept the at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one adaptive part model of the at least one terminal device is greater than the second performance indicator threshold for a threshold percentage of a number of the at least one terminal device.
[0141] Example 20. The network device of example 19, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine the threshold percentage based on a number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received; determine the threshold percentage to be a first percentage that is lower than a second percentage, when the number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received is greater than a terminal device number threshold; and determine the threshold percentage to be the second percentage that is higher than the first percentage, when the number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received is less than the terminal device number threshold.
[0142] Example 21. The network device of any of examples 14 to 20, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to accept a first one of the at least one adaptive part model of the at least one terminal device having a first performance indicator of the at least one second performance indicator, and a second one of the at least one adaptive part model of the at least one terminal device having a second performance indicator of the at least one second performance indicator, when a difference between the first performance indicator and the secondperformance indicator is greater than a difference threshold; and determine to not accept the first one of the at least one adaptive part model or the second one of the at least one adaptive part model, when the difference between the first performance indicator and the second performance indicator is less than the difference threshold.
[0143] Example 22. The network device of any of examples 14 to 21, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to accept a first one of the at least one adaptive part model of the at least one terminal device having a first parameter, and a second one of the at least one adaptive part model of the at least one terminal device having a second parameter, when a difference between the first parameter and the second parameter is greater than a similarity threshold; and determine to not accept the first one of the at least one adaptive part model or the second one of the at least one adaptive part model, when the difference between the first parameter and the second parameter is less than the similarity threshold.
[0144] Example 23. The network device of any of examples 14 to 22, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine a difference between the at least one second performance indicator of the at least one adaptive part model of a first terminal device and the at least one second performance indicator of the at least one adaptive part model of a second terminal device; and determine the first terminal device and the second terminal device to be in the group, when the difference between the at least one second performance indicator of the at least one adaptive part model of the first terminal device and the at least one second performance indicator of the at least one adaptive part model of the second terminal device is less than a group performance threshold.
[0145] Example 24. The network device of any of examples 14 to 23, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine a variance of the at least one second performance indicator of an adaptive part model of a terminal device of the at least one terminal device; determine the terminal device to be in the group, when the variance of the at least one second performance indicator of the adaptive part model of the terminal device is less than a variance threshold.
[0146] Example 25. The network device of any of examples 14 to 24, wherein the at least one processing capability of the at least one of the at least one terminal device comprises atleast one memory capability of the at least one terminal device, and the processing capability threshold comprises a memory capability threshold.
[0147] Example 26. The network device of any of examples 14 to 25, wherein the at least one first performance indicator and the at least one second performance indicator are based on at least one of: a handover success rate, a handover failure rate, at least one throughput measure, at least one packet delay, a quality of service indicator, or beam selection accuracy.
[0148] Example 27. The network device of any of examples 14 to 26, wherein the at least one adaptive part model comprises a machine learning model.
[0149] Example 28. The network device of any of examples 14 to 27, wherein the at least one terminal device comprises a plurality of terminal devices.
[0150] Example 29. The network device of any of examples 14 to 28, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device, using data different from data used to train the at least one adaptive part model of the at least one terminal device.
[0151] Example 30. The network device of example 29, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device using at least one sample of the data used to determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device; wherein the at least one sample is from a first state space larger than a second state space comprising at least one sample of the data used to train the at least one adaptive part model of the at least one terminal device.
[0152] Example 31. The network device of any of examples 14 to 30, wherein the instructions, when executed by the at least one processor, cause the network device at least to: train the at least one adaptive part model of the at least one terminal device with data used train the at least one adaptive part model of the at least one terminal device.
[0153] Example 32. The network device of any of examples 14 to 31, wherein the network device comprises a gNB or a machine learning orchestrator.
[0154] Example 33. The network device of any of examples 14 to 32, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to not accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is less than the second performance indicator threshold.
[0155] Example 34. A method including: transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving, from the network device, at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0156] Example 35. A method including: receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of thegroup of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0157] Example 36. An apparatus including: means for transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay; means for receiving, from the network device, at least one adaptive part model of at least one terminal device; means for determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; means for transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; means for receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and means for determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0158] Example 37. An apparatus including: means for receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; means for determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; means for transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; means for receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; means for determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; means for determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; means for transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the atleast one terminal device; and means for transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0159] Example 38. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations including: transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving, from the network device, at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
[0160] Example 39. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations including: receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model ofthe at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
[0161] References to a ‘computer’, ‘processor’, etc. should be understood to encompass not only computers having different architectures such as single / multi-processor architectures and sequential or parallel architectures but also specialized circuits such as field- programmable gate arrays (FPGAs), application specific circuits (ASICs), signal processing devices and other processing circuitry. References to computer program, instructions, code etc. should be understood to encompass software for a programmable processor or firmware such as, for example, the programmable content of a hardware device whether instructions for a processor, or configuration settings for a fixed-function device, gate array or programmable logic device etc.
[0162] The memories as described herein may be implemented using any suitable data storage technology, such as semiconductor based memory devices, flash memory, magnetic memory devices and systems, optical memory devices and systems, non-transitory memory, transitory memory, fixed memory and removable memory. The memories may comprise a database for storing data.
[0163] As used herein, the term ‘circuitry’ may refer to the following: (a) hardware circuit implementations, such as implementations in analog and / or digital circuitry, and (b) combinations of circuits and software (and / or firmware), such as (as applicable): (i) a combination of processor(s) or (ii) portions of processor(s) / software including digital signal processor(s), software, and memories that work together to cause an apparatus to perform various functions, and (c) circuits, such as a microprocessor(s) or a portion of a microprocessor(s), that require software or firmware for operation, even if the software or firmware is not physically present. As a further example, as used herein, the term ‘circuitry’ would also cover an implementation of merely a processor (or multiple processors) or a portion of a processor and its (or their) accompanying software and / or firmware. The term ‘circuitry’ would also cover, for example and if applicable to the particular element, a baseband integrated circuit or applications processor integrated circuit for a mobile phone ora similar integrated circuit in a server, a cellular network device, or another network device.
[0164] It should be understood that the foregoing description is only illustrative. Various alternatives and modifications may be devised by those skilled in the art. For example, features recited in the various dependent claims could be combined with each other in any suitable combination(s). In addition, features from different example embodiments described above could be selectively combined into a new example embodiment. Accordingly, this description is intended to embrace all such alternatives, modifications and variances which fall within the scope of the appended claims.
[0165] The following acronyms and abbreviations that may be found in the specification and / or the drawing figures are given as follows (the abbreviations and acronyms may be appended with each other or with other characters using e.g. a dash, hyphen, slash, or number, and may be case insensitive):3GPP third generation partnership project4G fourth generation5G fifth generation5GC 5G core network6G sixth generationAl artificial intelligenceAMF access and mobility management functionAP adaptive partASIC application- specific integrated circuitCD compact / computer discCL continual learningCPU central processing unitCSI channel state informationCU central unit or centralized unitDSP digital signal processorDVD digital versatile disc eNB evolved Node B (e.g., an LTE base station)EN-DC E-UTRAN new radio - dual connectivity en-gNB node providing NR user plane and control plane protocol terminations towards the UE, and acting as a secondary node in EN-DCE-UTRA evolved universal terrestrial radio access, i.e., the LTE radio access technologyE-UTRAN E-UTRA networkFl interface between the CU and the DUFL federated learningFPGA field-programmable gate array gNB base station for 5G / NR, i.e., a node providing NR user plane and control plane protocol terminations towards the UE, and connected via the NG interface to the 5GCHl hidden layer 1H2 hidden layer 2H3 hidden layer 3HN hidden layer NIAB integrated access and backhaulI / F interface iid independent and identically distributedI / O input / outputKP key performanceKPI key performance indicatorLMF location management functionLTE long term evolution (4G)MAC medium access controlML machine learningMLO ML model orchestratorMME mobility management entityMRO mobility robustness optimizationNCE network control element ng or NG new generation ng-eNB new generation eNBNG-RAN new generation radio access networkNR new radioNW networkN / W networkPDA personal digital assistantPDCP packet data convergence protocolPHY physical layerQoS quality of serviceRAM random access memoryRAN radio access networkRLC radio link controlRL reinforcement learningROM read-only memoryRRC radio resource controlRU radio unitRx receiver or receptionSDAP service data adaptation protocolSGW serving gatewaySMF session management functionSON self-organizing / optimizing networkTh thresholdTRP transmission reception pointTx transmitter or transmissionUAV unmanned aerial vehicleUE user equipment (e.g., a wireless, typically mobile device)UI user interfaceUPF user plane functionUSB universal serial busX2 network interface between RAN nodes and between RAN and the core networkXn network interface between NG-RAN nodes
Claims
CLAIMSWhat is claimed is:
1. A terminal device comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the terminal device at least to: transmit, to a network device, an indication that the terminal device is capable of maintaining a latent replay; receive, from the network device, at least one adaptive part model of at least one terminal device; determine at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmit, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receive, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determine to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
2. The terminal device of claim 1, wherein the at least one latent replay of the at least one adaptive part model of the at least one terminal device comprises a stored representation of the at least one adaptive part model of the at least one terminal device.
3. The terminal device of any of claims 1 to 2, wherein the determining of the at least one performance indicator of the at least one adaptive part model of the at least one terminal device comprises performing inference using the at least one adaptive partmodel of the at least one terminal device, wherein the inference is performed with a training set local to the terminal device.
4. The terminal device of any of claims 1 to 3, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from the network device, an indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device; wherein the determining to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group is based on the indication received from the network device to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device.
5. The terminal device of claim 4, wherein the indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device is received from the network device at least partially based on the transmitted indication that the terminal device is capable of maintaining the latent replay.
6. The terminal device of any of claims 1 to 5, wherein the determining to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group is based at least partially on the transmitted indication that the terminal device is capable of maintaining the latent replay.
7. The terminal device of any of claims 1 to 6, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from the network device, the at least one latent replay of the at least one adaptive part model of the at least one terminal device, when a processing capability of the terminal device is less than a processing capability threshold.
8. The terminal device of any of claims 1 to 7, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: transmit, to a terminal device within the group of the at least one terminal device, the at least one latent replay of the at least one adaptive part model of the atleast one terminal device, when a processing capability of the terminal device is less than a processing capability threshold, and when a processing capability of the terminal device within the group is greater than the processing capability threshold.
9. The terminal device of any of claims 1 to 8, wherein the instructions, when executed by the at least one processor, cause the terminal device at least to: receive, from a terminal device within the group of the at least one terminal device, the at least one latent replay of the at least one adaptive part model of the at least one terminal device, when a processing capability of the terminal device is greater than a processing capability threshold, and when a processing capability of the terminal device within the group is less than the processing capability threshold.
10. The terminal device of any of claims 7 to 9, wherein the processing capability of the terminal device comprises a memory capability of the terminal device, the processing capability of the terminal device comprises a memory capability of the terminal device, and the processing capability threshold comprises a memory capability threshold.
11. The terminal device of any of claims 1 to 10, wherein the at least one adaptive part model comprises a machine learning model.
12. The terminal device of any of claims 1 to 11, wherein the terminal device comprises the at least one terminal device.
13. The terminal device of any of claims 1 to 12, wherein the at least one performance indicator is based on at least one of: a handover success rate, a handover failure rate, at least one throughput measure, at least one packet delay, a quality of service indicator, or beam selection accuracy.
14. A network device comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the network device at least to:receive, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determine at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receive, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determine to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determine a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmit, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmit, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
15. The network device of claim 14, wherein the at least one latent replay of the at least one adaptive part model of the at least one terminal device comprises a stored representation of the at least one adaptive part model of the at least one terminal device.
16. The network device of any of claims 14 to 15, wherein the instructions, whenexecuted by the at least one processor, cause the network device at least to: transmit, to the at least one terminal device, the indication to maintain the at least one latent replay of the at least one adaptive part model of the at least one terminal device based on the receiving, from the at least one terminal device, the indication that the at least one terminal device is capable of maintaining the latent replay.
17. The network device of any of claims 14 to 16, wherein the instructions, when executed by the at least one processor, cause the network device at least to: transmit, to the at least one terminal device, the at least one latent replay of the at least one adaptive part model parameter of the at least one terminal device, when the processing capability of the at least one terminal device is less than the processing capability threshold.
18. The network device of any of claims 14 to 17, wherein the instructions, when executed by the at least one processor, cause the network device at least to: merge the group of the at least one terminal device with another group of the at least one terminal device, when the at least one processing capability of the at least one terminal device of the group is less than the processing capability threshold, and when the at least one processing capability of the at least one terminal device of the another group is greater than the processing capability threshold.
19. The network device of any of claims 14 to 18, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to accept the at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one adaptive part model of the at least one terminal device is greater than the second performance indicator threshold for a threshold percentage of a number of the at least one terminal device.
20. The network device of claim 19, wherein the instructions, when executed by the at least one processor, cause the network device at least to:determine the threshold percentage based on a number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received; determine the threshold percentage to be a first percentage that is lower than a second percentage, when the number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received is greater than a terminal device number threshold; and determine the threshold percentage to be the second percentage that is higher than the first percentage, when the number of the at least one terminal device from which the at least one second performance indicator of the at least one adaptive part model of at least one terminal device was received is less than the terminal device number threshold.
21. The network device of any of claims 14 to 20, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to accept a first one of the at least one adaptive part model of the at least one terminal device having a first performance indicator of the at least one second performance indicator, and a second one of the at least one adaptive part model of the at least one terminal device having a second performance indicator of the at least one second performance indicator, when a difference between the first performance indicator and the second performance indicator is greater than a difference threshold; and determine to not accept the first one of the at least one adaptive part model or the second one of the at least one adaptive part model, when the difference between the first performance indicator and the second performance indicator is less than the difference threshold.
22. The network device of any of claims 14 to 21, wherein the instructions, when executed by the at least one processor, cause the network device at least to:determine to accept a first one of the at least one adaptive part model of the at least one terminal device having a first parameter, and a second one of the at least one adaptive part model of the at least one terminal device having a second parameter, when a difference between the first parameter and the second parameter is greater than a similarity threshold; and determine to not accept the first one of the at least one adaptive part model or the second one of the at least one adaptive part model, when the difference between the first parameter and the second parameter is less than the similarity threshold.
23. The network device of any of claims 14 to 22, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine a difference between the at least one second performance indicator of the at least one adaptive part model of a first terminal device and the at least one second performance indicator of the at least one adaptive part model of a second terminal device; and determine the first terminal device and the second terminal device to be in the group, when the difference between the at least one second performance indicator of the at least one adaptive part model of the first terminal device and the at least one second performance indicator of the at least one adaptive part model of the second terminal device is less than a group performance threshold.
24. The network device of any of claims 14 to 23, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine a variance of the at least one second performance indicator of an adaptive part model of a terminal device of the at least one terminal device; determine the terminal device to be in the group, when the variance of the at least one second performance indicator of the adaptive part model of the terminal device is less than a variance threshold.
25. The network device of any of claims 14 to 24, wherein the at least one processing capability of the at least one of the at least one terminal device comprises at least onememory capability of the at least one terminal device, and the processing capability threshold comprises a memory capability threshold.
26. The network device of any of claims 14 to 25, wherein the at least one first performance indicator and the at least one second performance indicator are based on at least one of: a handover success rate, a handover failure rate, at least one throughput measure, at least one packet delay, a quality of service indicator, or beam selection accuracy.
27. The network device of any of claims 14 to 26, wherein the at least one adaptive part model comprises a machine learning model.
28. The network device of any of claims 14 to 27, wherein the at least one terminal device comprises a plurality of terminal devices.
29. The network device of any of claims 14 to 28, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device, using data different from data used to train the at least one adaptive part model of the at least one terminal device.
30. The network device of claim 29, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device using at least one sample of the data used to determine the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device; wherein the at least one sample is from a first state space larger than a second state space comprising at least one sample of the data used to train the at least one adaptive part model of the at least one terminal device.
31. The network device of any of claims 14 to 30, wherein the instructions, when executed by the at least one processor, cause the network device at least to:train the at least one adaptive part model of the at least one terminal device with data used train the at least one adaptive part model of the at least one terminal device.
32. The network device of any of claims 14 to 31, wherein the network device comprises a gNB or a machine learning orchestrator.
33. The network device of any of claims 14 to 32, wherein the instructions, when executed by the at least one processor, cause the network device at least to: determine to not accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is less than the second performance indicator threshold.
34. A method comprising: transmitting, from a terminal device to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving, from the network device, at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
35. A method comprising:receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
36. A terminal device comprising: means for transmitting, to a network device, an indication that the terminal device is capable of maintaining a latent replay; means for receiving, from the network device, at least one adaptive part modelof at least one terminal device; means for determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; means for transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; means for receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and means for determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at least one parameter.
37. A network device comprising: means for receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; means for determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; means for transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; means for receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; means for determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least oneterminal device is greater than a second performance indicator threshold; means for determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; means for transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and means for transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
38. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations comprising: transmitting, from a terminal device to a network device, an indication that the terminal device is capable of maintaining a latent replay; receiving, from the network device, at least one adaptive part model of at least one terminal device; determining at least one performance indicator of the received at least one adaptive part model of the at least one terminal device; transmitting, to the network device, the at least one performance indicator of the at least one adaptive part model of the at least one terminal device; receiving, from the network device, at least one parameter of the at least one adaptive part model of the at least one terminal device of a group of the at least one terminal device, based at least partially on the at least one performance indicator; and determining to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group based on the received at leastone parameter.
39. A non-transitory program storage device readable by a machine, tangibly embodying a program of instructions executable by the machine for performing operations, the operations comprising: receiving, from at least one terminal device, an indication that the at least one terminal device is capable of maintaining a latent replay; determining at least one first performance indicator of at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, the at least one adaptive part model of the at least one terminal device, when the at least one first performance indicator of the at least one adaptive part model of the at least one terminal device is greater than a first performance indicator threshold; receiving, from the at least one terminal device, at least one second performance indicator of the at least one adaptive part model of the at least one terminal device; determining to accept at least one of the at least one adaptive part model of the at least one terminal device, when the at least one second performance indicator of the at least one of the at least one adaptive part model of the at least one terminal device is greater than a second performance indicator threshold; determining a group of the at least one terminal device, based on the accepted at least one of the at least one adaptive part model of the at least one terminal device; transmitting, to the at least one terminal device, at least one parameter of the at least one adaptive part model of the at least one terminal device of the group of the at least one terminal device; and transmitting, to the at least one terminal device, an indication to maintain at least one latent replay of the at least one adaptive part model of the at least one terminal device of the group, when at least one processing capability of the at least one terminal device is greater than a processing capability threshold.
