Apparatuses, methods and computer programs for updating two-sided model
The method of updating the first side model at the UE and controlling its use based on compatibility with the network-side model addresses incompatibility issues in two-sided models, ensuring optimal performance by ensuring compatibility and meeting performance criteria.
Patent Information
- Application Number
- PCT/EP2024/057699
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-10
- Filing Date
- 2024-03-22
- Publication Date
- 2024-11-14
AI Technical Summary
Conventional procedures for updating two-sided models, comprising a first side model at User Equipment (UE) and a second side model at the network, often result in incompatibility issues, leading to suboptimal joint performance and failure to meet desired performance criteria.
The solution involves updating the first side model at the UE and sending update information to the network node, receiving compatibility information on the updated model's compatibility with the second side model, and controlling the use of the updated model based on this information to ensure compatibility and optimal performance.
This approach ensures that the updated first side model remains compatible with the second side model, maintaining or improving the joint performance of the two-sided model, thereby meeting performance criteria and avoiding incompatibility issues.
Smart Images

Figure EP2024057699_14112024_PF_FP_ABST
Abstract
Description
[0001] APPARATUSES, METHODS AND COMPUTER PROGRAMS FOR UPDATING TWO-SIDED MODEL
[0002] TECHNOLOGICAL FIELD
[0003] Examples of the present disclosure relate to updating a two-sided model. Various examples provide apparatuses, methods and computer programs for use in updating a two- sided model. Some examples, though without prejudice to the foregoing, relate to updating a two-sided model, a first side model of which (located at User Equipment, UE) is configured to encode information (such as channel measurements or Channel State Information, CSI), and a second side model of which (located at the network) is configured to decode the encoded information.
[0004] BACKGROUND
[0005] Two-sided models typically comprise a first side model, located at a first device, and a second side model, located at a second separate / remote device. The first and second side models are paired and configured so as to jointly perform a task, such as performing joint inference. For example, first and second side models of a two-sided model may be Artificial Intelligence, Al, / Machine Learning, ML, models provided at a UE and at the network, NW (e.g. at a Radio Access Network, RAN, node or core node of the network). The paired AI / ML models may be configured to perform inference (e.g. encoding and decoding) jointly across the UE and the network. In this regard a first part of inference is firstly performed by the first side model at the UE, and then the remaining part of the inference is performed by the second side model at the network (or vice versa).
[0006] Conventional procedures for updating a two-sided model are not always optimal. Updating one side of a two-sided model may have adverse effects on the overall joint performance of the two-sided model, e.g. the two-sided model comprising an updated first side model in combination with a (non-updated) second side model. An updated side of the two-sided model may no longer be compatible with the other side the two-sided model. Performance of a two-sided model comprising an updated side and a (non-updated) other side may not meet required / desired performance criteria.
[0007] In some circumstances, it can be desirable to provide an improved UE, network node, method and computer program for use in updating a two-sided model. In some circumstances, it can be desirable to avoid incompatibility of an update side of a two-sided model with the other side of two-sided model. In some circumstances, it can be desirable to update a side of two-sided model such that it is compatible, in use, with the other side of two-sided model.
[0008] The listing or discussion of any prior-published document or any background in this specification should not necessarily be taken as an acknowledgement that the document or background is part of the state of the art or is common general knowledge. One or more aspects / examples of the present disclosure may or may not address one or more of the background issues.
[0009] BRIEF SUMMARY
[0010] The scope of protection sought for various embodiments of the invention is set out by the claims.
[0011] According to various, but not necessarily all, examples of the disclosure there are provided examples as claimed in the appended claims. Any examples and features described in this specification that do not fall underthe scope of the independent claims are to be interpreted as examples useful for understanding various embodiments of the invention.
[0012] According to at least some examples of the disclosure there is provided a User Equipment, UE, comprising: means for updating a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at the UE and a second side model of the two-sided model is located at a node of a network; means for sending, to the network node, first side model update information for enabling the network node to determine the updated first side model; means for receiving, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and means for controlling use of the updated first side model based at least in part on the compatibility information.
[0013] According to various, but not necessarily all, examples of the disclosure there is provided a method comprising: updating a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; sending, to the network node, first side model update information for enabling the network node to determine the updated first side model; receiving, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and controlling use of the updated first side model based at least in part on the compatibility information.
[0014] According to various, but not necessarily all, examples of the disclosure there is provided a computer program comprising instructions, which when executed by an apparatus, cause the apparatus to perform: updating a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; sending, to the network node, first side model update information for enabling the network node to determine the updated first side model; receiving, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and controlling use of the updated first side model based at least in part on the compatibility information.
[0015] According to various, but not necessarily all, examples of the disclosure there is provided an apparatus comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: update a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; send, to the network node, first side model update information for enabling the network node to determine the updated first side model; receive, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and control use of the updated first side model based at least in part on the compatibility information.
[0016] According to various, but not necessarily all, examples of the disclosure there is provided a non-transitory computer readable medium encoded with instructions that, when executed by at least one processor, causes at least the following to be performed: update a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; send, to the network node, first side model update information for enabling the network node to determine the updated first side model; receive, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and control use of the updated first side model based at least in part on the compatibility information.
[0017] The following portion of this 'Brief Summary’ section describes various features that can be features of any of the examples described in the foregoing portion of the 'Brief Summary’ section mutatis mutandis. The description of a function should additionally be considered to also disclose any means suitable for performing that function, or any instructions stored in at least one memory that, when executed by at least one processor, cause an apparatus to perform that function.
[0018] In some but not necessarily all examples the two-sided model comprises at least one selected from the group of: a machine learning, ML, model, a pair of side models comprising a first side model and a second side model jointly configured to perform joint inference, wherein an output from the first model is input into the second model, an autoencoder, and the first and second side models being configured to perform encoding and decoding of channel measurements.
[0019] In some but not necessarily all examples, the first side model update information comprises at least one selected from the group of: the updated first side model; an identifier of the updated first side model, a description of the changes made to the first side model to form the updated first side model, information indicative of the configuration, structure or parameters of the updated first side model, an indication that an update of the first side model has occurred, and information for enabling the network node to update a local copy of the first side model to form a local copy of the updated first side model.
[0020] In some but not necessarily all examples, the compatibility information comprises: an indication of whether a type and / or format of an output from the updated first side model is compatible with a type and / or format of an input of the second side model; and / or an indication of a configuration of the second side model.
[0021] In some but not necessarily all examples, the UE further comprises means for sending, to the network node: information input to the updated first side model, and information output from the updated first side model; and wherein the compatibility information is based at least in part on: the input information and the output information.
[0022] In some but not necessarily all examples, the first side model is an encoder configured to encode information and the second side model is a decoder configured to decode the encoded information, and wherein the compatibility information is indicative of whether and to what extent the second side model decoder is able to recover the information encoded by the updated first side model encoder. In some but not necessarily all examples, the compatibility information is indicative of whether and to what extent the joint performance of the updated first side model and the second side model meets one or more performance criteria.
[0023] In some but not necessarily all examples, the one or more performance criteria comprises at least one selected from the group of: an accuracy of the joint performance of the updated first side model and the second side model; a reliability of the joint performance of the updated first side model and the second side model; an amount of processing resources required in the joint performance of the updated first side model and the second side model; and an amount of network resources required in the joint performance of the updated first side model and the second side model.
[0024] In some but not necessarily all examples, the control of the use of the updated first side model comprises at least one of: discarding, retaining, applying, or further updating the updated first side model.
[0025] In some but not necessarily all examples, the update of the first side model comprises performing knowledge distillation from a teacher first side model to a student first side model to form the updated first side model, and the UE further comprises means for sending knowledge distillation information from the knowledge distillation process to the network node.
[0026] In some but not necessarily all examples, the knowledge distillation information comprises at least one selected from the group of: knowledge distillation specific performance key performance indicators; distillation loss; latent space metrics; correlation metrics; configuration of the updated first side model; information indicative of a model structure of the updated first side model; and information indicative of model parameters of the updated first side model.
[0027] In some but not necessarily all examples, the knowledge distillation information is sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the updated first side model and the second side model.
[0028] According to at least some examples of the disclosure there is provided a network node comprising: means for updating a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at the network node and a first side model of the two-sided model is located at a User Equipment, UE; means for receiving first side model information for enabling the network node to determine the first side model of the two-sided model; means for determining compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; means for controlling use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0029] According to various, but not necessarily all, examples of the disclosure there is provided a method comprising: updating a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receiving first side model information for enabling the network node to determine the first side model of the two-sided model; determining compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and controlling use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0030] According to various, but not necessarily all, examples of the disclosure there is provided a computer program comprising instructions, which when executed by an apparatus, cause the apparatus to perform: updating a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receiving first side model information for enabling the network node to determine the first side model of the two-sided model; determining compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and controlling use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0031] According to various, but not necessarily all, examples of the disclosure there is provided an apparatus comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to: update a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receive first side model information for enabling the network node to determine the first side model of the two-sided model; determine compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and control use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0032] According to various, but not necessarily all, examples of the disclosure there is provided a non-transitory computer readable medium encoded with instructions that, when executed by at least one processor, causes at least the following to be performed: update a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receive first side model information for enabling the network node to determine the first side model of the two-sided model; determine compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and control use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0033] The following portion of this 'Brief Summary’ section describes various features that can be features of any of the examples described in the foregoing portion of the 'Brief Summary’ section mutatis mutandis. The description of a function should additionally be considered to also disclose any means suitable for performing that function, or any instructions stored in at least one memory that, when executed by at least one processor, cause an apparatus to perform that function.
[0034] In some but not necessarily all examples, the two-sided model comprises at least one selected from the group of: a machine learning, ML, model, a pair of side models comprising a first side model and a second side model jointly configured to perform joint inference, wherein an output from the first model is input into the second model, an autoencoder, and the first and second side models being configured to perform encoding and decoding of channel measurements.
[0035] In some but not necessarily all examples, the first side model information comprises at least one selected from the group of: the first side model; an identifier of the first side model, and information indicative of the configuration, structure or parameters of the first side model.
[0036] In some but not necessarily all examples, the compatibility information comprises: an indication of whether a type and / or format of an output from the first side model is compatible with a type and / or format of an input of the updated second side model; and / or an indication of a configuration of the updated second side model.
[0037] In some but not necessarily all examples, the network node further comprises means for receiving, from the network node: information input to the first side model, and information output from the first side model; and wherein the compatibility information is determined based at least in part on: the input information and the output information.
[0038] In some but not necessarily all examples, the first side model is an encoder configured to encode information and the second side model is a decoder configured to decode the encoded information, and wherein the compatibility information is indicative of whether and to what extent the updated second side model decoder is able to recover the information encoded by the side model encoder.
[0039] In some but not necessarily all examples, the network node further comprising means for determining one or more performance criteria of the joint performance of the first side model and the updated second side model; and wherein the compatibility information is indicative of whether and to what extent the joint performance of the first side model and the updated second side model meets the one or more performance criteria.
[0040] In some but not necessarily all examples, the one or more performance criteria comprises at least one selected from the group of: an accuracy of the joint performance of the first side model and the updated second side model; a reliability of the joint performance of the first side model and the updated second side model; an amount of processing resources required in the joint performance of the first side model and the updated second side model; and an amount of network resources required in the joint performance of the first side model and the updated second side model. In some but not necessarily all examples, the control of the use of the updated second side model comprises at least one of: discarding, retaining, applying, or further updating the updated second side model.
[0041] In some but not necessarily all examples, the update of the second side model comprises performing knowledge distillation from a teacher second side model to a student second side model to form the updated second side model, and the network node comprises means for sending knowledge distillation information from the knowledge distillation process to the UE.
[0042] In some but not necessarily all examples, the knowledge distillation information comprises at least one selected from the group of: knowledge distillation specific performance key performance indicators; distillation loss; latent space metrics; correlation metrics; configuration of the updated first side model; information indicative of a model structure of the updated second side model; and information indicative of model parameters of the updated second side model.
[0043] In some but not necessarily all examples, the knowledge distillation information is sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the first side model and the updated second side model.
[0044] While the above examples of the disclosure and optional features are described separately, it is to be understood that their provision in all possible combinations and permutations is contained within the disclosure. It is to be understood that various examples of the disclosure can comprise any or all of the features described in respect of other examples of the disclosure, and vice versa. Also, it is to be appreciated that any one or more or all of the features, in any combination, may be implemented by / comprised in / performable by an apparatus, a method, and / or computer program instructions as desired, and as appropriate.
[0045] BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Some examples will now be described with reference to the accompanying drawings in which:
[0047] Fig. 1 schematically illustrates an example of a radio telecommunications network suitable for use with examples of the present disclosure;
[0048] Fig. 2 schematically illustrates an example of a two-sided model for use with examples of the present disclosure;
[0049] Fig. 3 schematically illustrates another example of a two-sided model for use with examples of the present disclosure; Fig. 4 schematically illustrates an example of a method in accordance with the subject matter described herein;
[0050] Fig. 5 schematically illustrates another example of a method in accordance with the subject matter described herein;
[0051] Fig. 6 schematically illustrates a further another example of a method in accordance with the subject matter described herein;
[0052] Fig. 7 schematically illustrates an example of a signaling diagram in accordance with the subject matter described herein;
[0053] Fig. 8 schematically illustrates another example of a signaling diagram in accordance with the subject matter described herein;
[0054] Fig. 9 schematically illustrates a further example of a signaling diagram in accordance with the subject matter described herein;
[0055] Fig. 10 schematically illustrates a yet further example of a signaling diagram in accordance with the subject matter described herein;
[0056] Fig. 11 schematically illustrates an example of knowledge distillation;
[0057] Fig. 12 schematically illustrates an example of an apparatus in accordance with the subject matter described herein; and
[0058] Fig. 13 schematically illustrates an example of a computer program in accordance with the subject matter described herein.
[0059] The figures are not necessarily to scale. Certain features and views of the figures can be shown schematically or exaggerated in scale in the interest of clarity and conciseness. For example, the dimensions of some elements in the figures can be exaggerated relative to other elements to aid explication. Similar reference numerals are used in the figures to designate similar features. For clarity, all reference numerals are not necessarily displayed in all figures.
[0060] In the drawings (and description) a similar feature may be referenced by the same three- digit number. In the drawings (and description), an optional subscript to the three-digit number can be used to differentiate different instances of similar features. Therefore, a three-digit number without a subscript can be used as a generic reference and the three- digit number with a subscript can be used as a specific reference.
[0061] ABREVIATIONS / DEFINITIONS
[0062] 3GPP 3rd Generation Partnership Project
[0063] 5G 5th Generation
[0064] ACK Acknowledgement
[0065] Al Artificial Intelligence
[0066] CSI Channel State Information
[0067] CSI-RS Channel State Information Reference Signals gNB 5G / NR base station
[0068] KD Knowledge Distillation KPI Key Performance Indicator
[0069] MAC-CE Media Access Control - Control Element
[0070] ML Machine Learning
[0071] NR New Radio
[0072] NACK Negative Acknowledgement
[0073] NW Network
[0074] RAN Radio Access Network
[0075] RRC Radio Resource Control
[0076] TRP Transmission Reception Point
[0077] UE User Equipment
[0078] DETAILED DESCRIPTION
[0079] FIG. 1 schematically illustrates an example of a network 100 suitable for use with examples of the present disclosure. The network comprises a plurality of network nodes including: terminal nodes 110 (also referred to as User Equipment, UE), access nodes 120 (also referred to as Transmission Reception Points, TRPs; Radio Access Network, RAN, nodes, or base stations), and one or more core network nodes 130. The terminal nodes 110 and access nodes 120 communicate with each other. The one or more core network nodes 130 may, in some but not necessarily all examples, communicate with each other. The one or more access nodes 120 may, in some but not necessarily all examples, communicate with each other.
[0080] The network 100 is in this example a radio telecommunications network, i.e. a RAN, in which at least some of the terminal nodes 110 and access nodes 120 communicate with each other using transmission / reception of radio waves.
[0081] The RAN 100 may be a cellular network comprising a plurality of cells 122 each served by an access node 120. The access nodes 120 comprise cellular radio transceivers. The terminal nodes 110 comprise cellular radio transceivers.
[0082] In the particular example illustrated, the network 100 may be a New Radio, NR, network of the Third Generation Partnership Project, 3GPP, and its fifth generation, 5G, technology. In other examples, the network 100 may be a Next Generation, NG, i.e. sixth generation, 6G, Radio Network that is currently under development (i.e. an evolution of the NR network and its 5G technology).
[0083] The interfaces between the terminal nodes 110 and the access nodes 120 are radio interfaces 124 (e.g., Uu interfaces). The interfaces between the access nodes 120 and one or more core nodes 130 are backhaul interfaces 128 (e.g., S1 and / or NG interfaces).
[0084] Depending on the exact deployment scenario, the access nodes 120 may be RAN nodes such as NG-RAN nodes. NG-RAN nodes may be gNodeBs, gNBs, that provide NG user plane and control plane protocol terminations towards the UE. The gNBs connected by means of NG interfaces to a 5G Core (5GC), more specifically to an Access and Mobility Management Function, AMF, by means of an NG Control Plane, NG-C, interface and to a User Plane Function, UPF, by means of an NG User Plane, NG-U, interface. The access nodes 120 may be interconnected with each other by means of Xn interfaces 126.
[0085] The cellular network 100 could be configured to operate in licensed frequency bands, or unlicensed frequency bands (not least such as: unlicensed bands that rely upon a transmitting device to sense the radio resources / medium before commencing transmission, such as via a Listen Before Talk, LBT, procedure; and a 60GHz unlicensed band where beamforming may be required in order to achieve required coverage).
[0086] The access nodes 120 may be deployed in an NG standalone operation / scenario. The access nodes 120 may be deployed in a NG non-standalone operation / scenario. The access nodes may be deployed in a Carrier Aggregation operation / scenario. The access nodes 120 may be deployed in a dual connectivity operation / scenario, i.e., Multi Radio Access Technology - Dual Connection, MR-DC.
[0087] In such non-standalone / dual connectivity deployments, the access nodes 120 may be interconnected to each other by means of X2 or Xn interfaces, and connected to an Evolved Packet Core, EPC, by means of an S1 interface or to the 5GC by means of a NG interface.
[0088] The terminal nodes 110 are network elements in the network that terminate the user side of the radio link. They are devices allowing access to network services. The terminal nodes 110 may be referred to as User Equipment, UE, mobile terminals or mobile stations. The term 'User Equipment’ may be used to designate mobile equipment comprising a smart card for authentication / encryption etc such as a Subscriber Identity Module, SIM. In other examples, the term 'User Equipment’ is used to designate a location / position tag, a hyper / smart tag or a mobile equipment comprising circuitry embedded as part of the user equipment for authentication / encryption such as software SIM.
[0089] The access nodes 120 are network elements in the network responsible for radio transmission and reception in one or more cells 122 to or from the terminal nodes 110. The access nodes 120 are the network termination of a radio link. An access node 120 may be implemented as a single network equipment, or have a split architecture that is disaggregated / distributed over two or more RAN nodes, such as a Central Unit, CU, a Distributed Unit, DU, a Remote Radio Head-end, RRH, using different functional-split architectures and different interfaces.
[0090] Where the access node 120 has a disaggregated (split) architecture, the access node 120 may comprise one or more distributed units, gNB-DU, and a centralized unit, gNB-CU (not shown in FIG. 1 ). The gNB-CU is a logical node configured to host a Radio Resource Connection, RRC, layer and other layers of the access node 120. The gNB-CU controls the operation of one or more gNB-DUs. The gNB-DU is a logical node configured to host Radio Link Control, RLC, protocol layer, Medium Access Control, MAC, layer and Physical, PHY, layer of the access node 120. The gNB-DU may communicate via a dedicated interface (e.g. an F1 interface) to a RRC layer hosted by the gNB-CU. One gNB-DU may support one or multiple cells 122.
[0091] In the following description, a terminal node 110 will be referred to simply as UE, and an access node / RAN node 120, core node 130 or other node / server of the network will be referred to simply as the network, NW, or a network node.
[0092] Fig. 2 schematically illustrates an example of a two-sided model 201 for use with examples of the present disclosure. The two-sided model comprises: a first side model 201 located at / running at a UE 110, and a second side model 2012located at / running at NW 120, 130.
[0093] Outputs from the UE’s first side model are input into the NW’s second side model.
[0094] The side models at UE and NW may be Artificial Intelligence, Al, / Machine Learning, ML, models, that are prepared and trained such that it is necessary to get the inference from both of them consequently in order to obtain a desired final output.
[0095] In some examples, the two-sided model comprises paired AI / ML Models over which joint inference is performed, where joint inference comprises AI / ML inference whose inference is performed jointly across the UE and the network, i.e. the first part of inference is firstly performed by UE and then the remaining part is performed by gNB, or vice versa.
[0096] The two side models of the two-sided model could be either high dependency models (i.e. side models whose architecture has high dependency) or low dependency models (i.e. side models whose architecture has low dependency).
[0097] 1 . High dependency models.
[0098] An example of high dependency models / high dependency architecture is an autoencoder, which has an encoder part (e.g. at a UE) as a first side model and a decoder part (e.g. at a NW) as a second side model.
[0099] In an auto encoder, it is not generally possible to train / update / retrain the decoder side model without input from the encoder side model (forward pass). Likewise, it is not generally possible to train / update / retrain the encoder side model without backpropagation from the decoder side model (backward pass) which holds the final loss function and the ground truth in case of supervised learning. At a time of training, the UE and NW may have the full autoencoder (i.e. the UE may have both the encoder side model as well as the decoder side model, and likewise the NW may have both the decoder side model as well as the encoder side model). Whereas, at inference, just the encoder part is used at the UE, and just the decoder part is used at the NW.
[0100] Since the UE may initially have the full autoencoder (i.e. an encoder with version N, and also a compatible decoder that matches the network side decoder with version M), it may be possible to retrain the encoder part as it has the decoder part (same version M as at the network). Similarly, since the NW may initially have the full autoencoder (i.e. a decoder with version M, and also a compatible encoder with version N that matches the UE side encoder), it may be possible to retrain the decoder part as it has the encoder as well (same version N of encoder at the UE side).
[0101] The UE can retrain its encoder to a new version N’, but it has to share such information about the new encoder version N’ with the NW. Similarly, the network can retrain its decoder to new version M’ but it has to share such information about the new decoder version M’ with the UE. If the UE updates its model to version N’ from version N, this new version N’ must remain compatible with the existing model at the NW side (i.e. M) and vice versa. Accordingly, in this architecture, there is a strong dependency between the model parts.
[0102] 2 - Low dependency models
[0103] An example of low dependency models / low dependency architecture is two separate models that have with input dependencies. For instance, two models may be provided at different parts of the network (UE and NW), and an output of one side model is used as an input to the second side model. Each side model has its own output layer with a loss function (which differs from an auto encoder that has a loss function at the decoder only).
[0104] Even though the two models are separate models, since the output from one is used as input to the other, then version control and compatibility needs to be checked if one of the models is updated to a new / different version. In this architecture, there is no strong dependency between the models, only dependency on the output of the first model part. Due to this, retraining could be done independently, e.g. making use of performance Key Performance Indicator, KPI, metrics.
[0105] In the low dependency case, the performance KPIs may comprise one or more or: accuracy, F1 -score, mean absolute error, or mean squared error. The ground truth or label information may not be shared in the low dependency case. In one example of a low dependency case, the first part of the two-sided model performs inference while the other / second part of the two-sided model monitors and / or influencing a decision making process. In high dependency case, the performance KPIs may comprise one or more of: intermediate normalized mean squared error, squared generalized cosine similarity (SGCS), a correlation between input and output of the first node, etc., and also the above-mentioned low dependency KPIs.
[0106] Fig. 3 schematically illustrates another example of a two-sided model 201 for use with examples of the present disclosure. The two-sided model comprises first UE side model M1 2011 and a second NW side model M2 2012.
[0107] The UE part M1 has an encoder and a quantizer Q. The NW part M2 has an inverse quantizer and a decoder. In some setups, the quantizer can be part of the encoder and decoder (not shown in the Figure), in other cases, those operations can be separated.
[0108] The first side model M1 performs an inference. It receives information / data X as an input and outputs Y, where Y = M1 (X). In effect, Y is an intermediate result of the two-sided model.
[0109] Then the second side model M2 performs an inference. It receives the output Y from M1 and outputs x, where x=M2(Y).
[0110] In some examples, where the two-sided model is applied to an autoencoder for encoding / compressing channel measurements, the input X could be channel measurements (e.g. CSI reports) measured by a UE, and the encoded CSI reports are transmitted to the NW to be decoded to extract x (i.e. seek to recover X).
[0111] An accuracy of the joint inference can be estimated as the error between x and X.
[0112] Incompatibility issues:
[0113] If M1 were changed (e.g. updated to a new / different version), then M2 may not be able to produce x to within a predetermined / acceptable accuracy threshold, e.g. such diff(X, x) < 1 e-12. Likewise, if M2 were updated to a new / different version, then when updated M2 is used in conjunction with M1 (i.e. using Mi ’s output as updated M2’s input), updated M2 may not be able to produce x to within a predetermined / acceptable accuracy threshold.
[0114] Training:
[0115] In a two-sided model, for example, an autoencoder which comprises of an encoder and a decoder, it is highly desirable to enable model adaptability (e.g. retraining / updating the two- sided model) with new environments over the course of time, such as to accommodate / be optimised for new contexts. Differing training collaboration types for two-sided models require differing retraining / updating procedures, as set out below. Type 1 Joint training with model transfer: An update of one part / side model (e.g. either encoder or decoder) in one entity / device (e.g. UE or NW) would need to be transferred to the other entity / device.
[0116] • Type 2 Joint training without model transfer: An update of one part / side model in one entity / device would need to be informed to the other entity / device simultaneously with the update taking effect (so that the other entity may resume its operation effectively without any performance degradation).
[0117] • Type 3 Separate training: An update of one part / side model in one entity / device would need to be indicated to the other entity. The other entity / device would need to operate effectively without any performance degradation.
[0118] Moreover, for any model update, particularly in Type 2 or Type 3, each part / side model would need to be compatible with each other, such as with regards to output, input, configuration, structure, KPIs and so on.
[0119] As will be discussed below, examples of the present disclosure seeks to address issues relating to an update of a two-sided model occurring on either of the two side models (e.g. either at the UE side model or NW side model). By way of an example, before an update, an existing side model at the UE (version N) may be compatible with an existing side model at the NW (having version M, which is compatible with the UE’s side model version N). However, under an "extendability” condition of 3GPP ["Extendability: to train new UE-side model compatible with NW-side model in use; Or to train new NW-side model compatible with UE-side model in use”], after an update of one side model of a two-sided model, the updated side model (located at one entity / device) must remain compatible with the preexisting other side model (located at the other entity / device). For example, if a UE updates its side model to version N+1 from version N, this new version (N+1 ) of the UE’s side model must remain compatible with the existing version M of the NW’s side model (and vice versa).
[0120] Examples of the disclosure seek to provide a side model training procedure wherein a new / updated / re-trained UE side model is able to co-operate with NW’s side model, e.g. to co-operate transparently or require no / almost no change required in the NW’s side model. Likewise, the other way is also possible. That is a new / updated / re-trained NW side model is able to co-operate with all UE-side models transparently or require no / almost no change in the UE-side models. It being appreciated that the NW’s side model may be configured to co-operate with / pair up with plural UE side models.
[0121] As an example use case, a two-sided model could be an autoencoder for a CSI compression, where an encoder-decoder pair of side models (at a UE and NW respective) is used to compress CSI measurements of CSI reports: the CSI is first compressed with an encoder at the UE side; the compressed data is then transmitted to the NW (e.g. gNB) and thereafter used as input for the decoder to retrieve the initial CSI measurements. Thus, a model update in either encoder (UE part model) or in decoder (gNB part model) may impact the performance of the other side (and / or the joint performance of the (partially updated) two-sided model) if a similar / corresponding / complementary update does not take place on the other side model.
[0122] Various examples of the present disclosure seek to provide a framework which satisfies the extendability criteria as defined in 3GPP: to train new UE-side model compatible with NW-side model in use; or to train a new NW-side model compatible with UE-side model in use.
[0123] Fig. 4 schematically illustrates an example of a method in accordance with the subject matter described herein.
[0124] A two-sided model is provided via a first side model 201 located at a UE 110, and a second side model 2012located at NW 120, 130.
[0125] The two-sided model may comprise at least one selected from the group of: a machine learning, ML, model, a pair of side models comprising a first side model and a second side model jointly configured to perform joint inference, wherein an output from the first model is input into the second model, an autoencoder, and the first and second side models being configured to perform encoding and decoding of channel measurements.
[0126] The first and second side models are initially matched / paired with one another. In this regard they are configured so as to be compatible with each other and to jointly perform an operation / inference (not least for example encoding and decoding). The initial / original versions of the first and second side models are initially configured / trained so as to meet one or more performance metrics in their joint performance of the operation / inference.
[0127] In block 401 , the UE updates its first side model 2011 to form an updated first side model 201r.
[0128] The UE sends, to the NW, a signal / message 402 that indicates first side model update information 403 for enabling the NW to determine the updated first side model (e.g. enable the NW to receive / obtain / establish its own copy of the updated first side model).
[0129] In this regard, the first side model update information may comprise at least one selected from the group of: the updated first side model (such that NW would receive, from the UE, its own copy of the updated first side model); an identifier of the updated first side model (such that the NW could download / obtain for itself a copy of the updated first side model, e.g. from a data storage storing a pre-existing matching version of the updated first side model), a description of the changes made to the first side model to form the updated first side model (such that the NW could establish, for itself, the updated first side model, e.g. by the NW duly updating a copy of the first side model so as to form its own copy of the updated first side model), information indicative of the configuration, structure or parameters of the updated first side model (such that the NW could create / reconstruct its own copy of the updated first side model), an indication that an update of the first side model has occurred (so that the NW can be aware that the UE’s first side model has been updated), and information for enabling the NW to update a local copy of the first side model (i.e. local to the NW) to form a local copy of the updated first side model (for instance an indication of the training data used to update / train the first side model).
[0130] In block 404, the NW determines whether and to what extent the updated first side model M1 ’ is compatible with the second side model M2. For example, where the updated first side model is an updated encoder and the second side model is a decoder, an assessment of the compatibility of the two side models may comprise a determination as to whether the decoder is capable of doing the decoding process, i.e. recovering the information encoded via the updated encoder, as well as a determination of the accuracy of the recovered information.
[0131] In this regard, following receipt of the model update information, the NW can determine the updated first side model M1 ’, e.g. via one of the above-mentioned processes, such that the NW has its own local copy of the updated first side model M1 ’ which it can use to perform an evaluation and / or monitoring of the compatibility of the updated first side model M1 ’ with the NW’s (non-updated) second side model M2.
[0132] The evaluation of the compatibility of the updated first side model M1 ’ with the NW’s (nonupdated) second side model M2 may comprise determining whether a type and / or format of an output from the updated first side model is compatible with a type and / or format of an input of the second side model. This may be done since, when one side model is updated / modified, its output format may change (e.g. higher dimension / different type). Since the output of the one side model is input to the other side model, a check should be made that the first side model’s output format is compatible with the second side model, such that that second model inference can be run on the second side model. Once the output format is confirmed as compatible, and the second inference is run with the second side model, then an evaluation / estimate of the accuracy of the overall two-sided model can be made. The evaluation and / or monitoring of the compatibility of the updated first side model M1 ’ with the NW’s (non-updated) second side model M2 may comprise performance validation, namely the NW determining whether and to what extent the joint performance of the updated first side model and the second side model meets one or more performance criteria (e.g. if it meets pre-determined threshold values of performance criteria metrics).
[0133] Such performance criteria may comprise one or more of: an accuracy of the joint performance of the updated first side model and the second side model; a reliability of the joint performance of the updated first side model and the second side model; an amount of processing resources required in the joint performance of the updated first side model and the second side model; and an amount of network resources required in the joint performance of the updated first side model and the second side model.
[0134] With regards to the determination of such performance criteria, they may be determined during use / operation of the first updated side model, i.e. during use / operation of a first updated side model with a second side model (wherein the second side model is configured so as to be compatible with the first side model, i.e. an earlier version of the first updated side model prior to it being updated). In other words, the compatibility / performance criteria may be determined. The inventors have appreciated that the performance depends also on channel conditions. Advantageously, by determining performance criteria during use of the two side models, the compatibility of the updated first side model and the second side model can be assessed under real world conditions, and in particular real world channel conditions. Likewise, an evaluation and monitoring of the amount of network resources (not least network overheads), and computer processing resources, involved in the real world use of the model can be determined. For example, this may enable a determination as to whether, in spite of the updated side model meeting accuracy-based criteria, nevertheless employing the updated side model necessitates too much processing / network resources, and hence a resources-based criteria is not met.
[0135] By way of an example, consider a "part-updated” two-side model somewhat similar to that of Fig. 3 albeit it comprising: an updated first side model M1 ’ configured to encode input information X to output encoded information Y’, i.e. Y’= M1 ’(X); and a (non-updated) second side model M2 configured to decode information encoded by the (non-updated) first side model M1 , i.e. an earlier version of the first side model M1 prior to M1 ’s updating to becoming the updated first side model M1 ’. In this regard, the second side model M2 is configured to decode Y to output x, i.e. x=M2(Y) = M2(M1 X). During use / operation of the first updated side model M1 the UE would send, to the NW, an output Y’ from M1 ’. Y’ can be considered to be an intermediate result of a "part-updated” two-sided model that comprises: M1 ’ and M2. The sending of Y’ from the UE to NW via a wireless channel means that the overall performance of the two-sided model with regard to the encoding, transmission / reception, and decoding is subject to (and may be affected by) channel conditions during the transmission / reception. Evaluating and monitoring compatibility during use, i.e. real world use under real world channel conditions, may provide an improved indication of compatibility of the updated side model with the other side model such as with regards to its accuracy and reliability under differing / real world channel conditions.
[0136] The NW, could then use its second side model M2 to decode Y’ to seek to recover X, namely decode Y’ to produce x’. In order to evaluate the accuracy / reliability of the joint performance of the encoding / decoding of the two-sided model, the NW would need a copy of the initial X (the ground truth) that it could then compare to its output x’ (its inferred result). Accordingly, in addition to sending Y’ (an encoded version of the initial input information / data X, encoded by the first updated side model), the UE also sends a copy of X (e.g. a non-encoded copy of X). In other words, the UE would separately send, to the NW, information X (i.e. a copy of the information that is input to the updated first side model), and information Y’ (i.e. the output from the updated first side model), so that the NW can use the same to determine the compatibility information. The sending of both X and Y, by the UE to the NW for the NW to use in determining the compatibility information, is not shown in the Figure.
[0137] Following on from the NW’s determination of the compatibility of the first updated side model with the second side model, the NW determines compatibility information 406 based on the same (e.g. such compatibility information being based, at least in part, on the results of the above discussed compatibility evaluation(s) / monitoring).
[0138] The NW sends a signal / message 405 indicative of the compatibility information 406, wherein the compatibility information is indicative of whether and to what extent the updated first side model is compatible with the second side model.
[0139] In block 407, the UE controls use of the updated first side model based at least in part on the compatibility information.
[0140] Such control of the use of the updated first side model may comprise at least one of: discarding the updated first side model (e.g. based on the compatibility information indicating a compatibility metric below a first threshold value - in which case the UE may revert to using the former / pre-updated first side model) retaining / applying the updated first side model (e.g. using / continuing to use the new update first side model based on the compatibility information indicating a compatibility metric above a second threshold value), and further updating the updated first side model (e.g. based on the compatibility information indicating a compatibility metric crossing a third threshold value, or having a value between the first and second threshold values).
[0141] Optionally, in block 408, the NW may determine, based at least in part on the compatibility information, to update its own second side model 2012to form an updated second side model 2012' based at least in part on the compatibility information, for example based on the compatibility information indicating a compatibility metric crossing a threshold value.
[0142] In some examples, the updating of the first side model in block 401 comprises performing a knowledge distillation process (which is discussed further with respect to FIG. 11). Knowledge distillation is a process in which knowledge is transferred from a large model to a smaller one. While large models (such as very deep neural network or ensembles of many models) have higher knowledge capacity than small models, this capacity might not be fully utilized. It can be just as computationally expensive to evaluate a model even if it utilizes little of its knowledge capacity. Knowledge distillation transfers knowledge from a large model to a smaller model without loss of validity.
[0143] In some examples, knowledge from a (large) teacher first side model can be distilled and conveyed to a (smaller) student first side model, whereupon the student first side model thereby forms the updated first side model.
[0144] The knowledge distillation information may comprise information relating to one or more of: knowledge distillation specific performance key performance indicators; distillation loss; latent space metrics; correlation metrics; configuration of the updated first side model; information indicative of a model structure of the updated first side model; and information indicative of model parameters of the updated first side model.
[0145] The UE may send knowledge distillation information to the NW. The knowledge distillation information may be sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the updated first side model and the second side model. For instance, the sending of the knowledge distillation information may be based on the compatibility information indicating a compatibility metric crossing a threshold value. Such a threshold value may be set so as to be indicative of a compatibility level that is not so low that the updated first side model ought to be discarded (and that too much updating / training of the second model would be required in order to reach an acceptable level of compatibility with the updated first side model), but that an acceptable amount of updating / training of the second model would be required in order to improve / optimise the level of compatibility and seek to improve KPI metrics of the joint model. Fig. 5 schematically illustrates another example of a method in accordance with the subject matter described herein. This method is somewhat similar to that of the method of FIG. 4. However, rather than the UE updating its first side model M1 in block 401 (as in Fig. 4); instead in FIG. 5 the NW updates its second side model M2 in block 501. In view of the similarity, various details and aspects of the description provided with reference to Fig. 4 are relevant and applicable to FIG. 5, mutatis mutandis, and shall not be repeated for the sake of brevity.
[0146] In Fig. 5, a two-sided model is provided via a first side model 201 located at a UE 110, and a second side model 2012located at NW 120, 130.
[0147] In block 501 , the NW updates its second side model 2012to form an updated second side model 2012'.
[0148] The UE sends, to the NW, a signal / message 502 that indicates first side model information 403 for enabling the NW to determine the UE’s first side model (e.g. enable the NW to receive / obtain / establish its own copy of the UE’s first side model).
[0149] In this regard, the first side model information may comprise at least one selected from the group of: the first side model (such that NW would receive, from the UE, its own copy of the first side model); an identifier of the first side model (such that the NW could obtain for itself a copy of the first side model, e.g. from a data storage storing a pre-existing matching version of the first side model), information indicative of the configuration, structure or parameters of the first side model (such that the NW could create its own copy of the first side model).
[0150] In block 504, the NW determines whether and to what extent the updated second side model M2’ is compatible with the first side model M1 .
[0151] In this regard, the NW may evaluate and monitor a compatibility of the updated first side model M1 ’ with the NW’s (non-updated) second side model M2 may comprise determining whether a type and / or format of an output from the updated first side model is compatible with a type and / or format of an input of the second side model.
[0152] The evaluation and / or monitoring of the compatibility of the NW’s updated second side model M2’ with the UE’s (non-updated) first side model M1 may comprise the NW carrying out performance validation, namely the NW determining whether and to what extent the joint performance of the updated second side model and the first side model meets one or more performance criteria (e.g. if it meets pre-determined threshold values of performance criteria metrics). Similar to that discussed with respect to Fig. 4, the determination of the compatibility may involve the UE sending X itself (i.e. the initial input information / data) along with Y (the output of the first model Y=M1 (X)) for the NW to use in determining the compatibility information. Such sending of X and Y is not shown in FIG. 5.
[0153] Following on from the NW’s determination of the compatibility of the second updated side model with the first side model, the NW determines compatibility information 506 based on the same (e.g. such compatibility information being based, at least in part, on the results of the above discussed compatibility evaluation(s) / monitoring).
[0154] The NW sends a signal / message 505 indicative of the compatibility information 506, wherein the compatibility information is indicative of whether and to what extent the updated second side model is compatible with the first side model.
[0155] In block 407, the NW controls use of the updated second side model based at least in part on the compatibility information.
[0156] Optionally, in block 508, the UE may determine, based at least in part on the received compatibility information, to update its own first side model 2011 to form an updated second side model 201 based at least in part on the compatibility information, for example based on the compatibility information indicating a compatibility metric crossing a threshold value.
[0157] Similar to that discussed with respect to Fig. 4, the updating of the second side model in block 401 may comprise performing a knowledge distillation process. The NW may send knowledge distillation information to the UE. The knowledge distillation information may be sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the updated first side model and the second side model.
[0158] Fig. 6 schematically illustrates a further another example of a method in accordance with the subject matter described herein. This figure depicts a high level flowchart of a framework updating a side model (located at a device / entity, e.g. UE or NW) of a two-sided model (wherein the two-sided model also comprises another side model located at another device / entity, e.g. NW or UE).
[0159] At step 1 , an update of a side model is performed. Such an update may be done to cope with local context changes at the device / entity where the side model is located (e.g. UE). An example of this step corresponds to block 401 of Fig. 4.
[0160] At step 2, following the updating of the side model, the device / entity (e.g. UE) indicates this update to the other device / entity (e.g. sends such an indication to a gNB) along with the updated side model or a model ID / model description of the updated side model. An example of this step corresponds to the signal / message 402 of Fig. 4.
[0161] At step 3, the other device / entity (gNB in this example) determines the compatibility of the received / indicated updated side model with its own other side of the two-sided model. This may correspond to performing the (part updated) two-sided model together, i.e. performing the updated side model combined with the existing / old (non-updated) other side model residing at the other device / entity.
[0162] At step 4, the result of the compatibility evaluation is compared to a predefined threshold value, th, of a compatibility metric. This may involve determining an accuracy metric value and reliability metric value of the performance of the joint inference of the part updated two- sided model. The accuracy metric and reliability metric can be indicative of a performance gap, i.e. a gap (which may vary) in between an expected performance predicted during training and, when the model is deployed, the actual measured performance. Such a performance gap may be due to the random nature of the transmission channel causing a different performance to be observed than was theoretically anticipated.
[0163] An example of steps 3 and 4 correspond to block 404 of Fig. 4.
[0164] If the compatibility is high enough, for example greater than a threshold value th, then this can be taken as indicating that a performance gap after the occurrence of the update of the side model is not high, and that the other side model (at the gNB in this example) can be expected to be possibly updated without too much overall impact. In which case, at step 5, the other side model may itself then be updated, if needed. An example of step 5 corresponds to block 408 of Fig. 4.
[0165] However, if the performance gap is estimated as high, this would indicate that the updated one side model is too divergent (i.e. the decoder is not capable of decoding the information encoded by the updated encoder), and that it is expected to be costly to update the other side model (e.g. it would require complete retraining and would need a large number of labelled data set). If it were anticipated that the decoder could be updated within 1 microsecond of an update of the encoder, a large (albeit short term / transitory) performance gap would not be a significant issue. On the other hand, if it would take 1 week to update / build a decoder for the updated encoder, then a large performance might cause more a visible performance gap as an updated encoder and (non-updated) decoder pair would need to be used in the meanwhile.
[0166] In which case, at step 6, the other entity (gNB in this example) should discard the side model update. The other device / entity may indicate to the device / entity (UE in this example) to possibly realize a new training round / update, as per step 7. If a new training round / update can be realized, then another round of the side model update is performed, as per step 9. Then the process could loop back to step 1 .
[0167] Otherwise, at step 8, a recovery option is performed. For example, following a network configuration (i.e. the network configuring the UE to perform a recover option), a number of training iterations realized so far may be counted and, based on this count, the device / entity could either fall back to an old / former version of its side mode, or decide that the ML performance is too low and that a switch should be made to legacy non-ML approach. An example of steps 6 - 9 and the control of the use of the updated side model corresponds to block 407 of Fig. 4.
[0168] Fig. 7 schematically illustrates an example of a signaling diagram in accordance with the subject matter described herein. The method of FIG. 7 may be used where the first and second side models are strongly dependent upon one another, e.g. such as in an autoencoder, wherein the first side model is an encoder and the second side model is a decoder. For such strongly dependent side models, the updating / training of one side model requires the use of the other side model. The procedure shown in Fig. 7, is suitable for UE updating / retraining its encoder side model in Type 1 training (i.e. joint training with model transfer).
[0169] Initially, both the UE and the network start by obtaining the same (compatible) versions of the two-sided model, e.g. a two-sided encoder-decoder model. In this regard, the UE obtains and locally stores a copy of the encoder as well as its paired / compatible decoder. Likewise, the NW obtains and locally stores a copy of the encoder as well as the matching / corresponding decoder.
[0170] In some examples, when a ML feature is enabled in an entity (UE or network), the ML side models (e.g. encoder and decoder) may already be deployed and locally stored in that entity. In some examples, an entity could either download a side model from a server or the network, or the side model may be already embedded in the entity, or the entity could construct its own model.
[0171] If the UE detects a need for encoder retaining (which it may do via monitoring its encoder performance, either periodically or upon an event triggering a performance evaluation), then the procedure set out in the Figure may be performed.
[0172] In step 1, the UE sends a request to the NW to request permission to update its side model, the NW then assesses the update request and determine whether to allow the UE to update its side model (this could be based, at least in part, on whether multiple other UEs have tried to do the same / similar update - in which case the network may disregard the request and not permit the UE to update its encoder).
[0173] If the NW approves the update request, it indicates the same to the UE. Then, in step 2, the UE checks with the NW if there have been any updates to the decoder (i.e. whether the NW’s decoder has been updated to a different / new version than the version of the decoder locally stored at the UE.
[0174] If there are no decoder updates (option 1 ), then the NW informs the UE of the same in step 3, e.g. via a Negative Acknowledgement, NACK, message. Then, in step 4, the UE updates / retrains its encoder using the current local version of the decoder that it already has.
[0175] Alternatively, if there is a decoder update (option 2), then the NW informs the UE of the same in step 5, e.g. via an Acknowledgement, ACK, message. In step 6, the UE then requests the network for latest decoder version, responsive to which, in step 7 the NW transfers a copy of the latest decoder version to the UE.
[0176] In step 8, the UE then updates / retrains its encoder using the received latest version of the decoder.
[0177] Following the updating / retraining of the encoder in steps 4 or 8, in step 9, the UE informs the network about the UE’s updated version of the encoder. This allows the network to configure the measurement resources and reports for the updated encoder - for instance to send appropriate CSI-Reference Signals, RS, to the UE in order for the UE to measure CSI.
[0178] The UE then deploys the updated version of the encoder in the transmission of CSI reports to the NW. In this regard, the UE applies / uses the updated version of the encoder in order to encode CSI and, in step 10, sends CSI reports encoded via the updated version of the encoder for decoding by the network’s (non-updated) decoder.
[0179] In step 11 , in response to step 9 and step 10, the network trigger a model performance evaluation of the updated encoder. This is done using the intermediate output from the UE’s updated encoder, namely the encoded CSI reports encoded by the UE’s updated encoder, and the network’s own local decoder to decode the same. The network asses the result of decoding of the encoded CSI report against a ground truth CSI report. Such a ground truth CSI report may be: a) an original / uncompressed CSI report, b) generated using the UE’s local decoder (not available to NW), or c) a compressed CSI report generated by a legacy procedure.
[0180] In step 12, the networks sends an indication of a difference between the ground truth CSI report and the network’s decoded CSI report encoded by the UE’s updated encoder. Such a difference is indicative of a compatibility / performance KPI of the UE’s updated encoder and the network’s (non-updated) decoder. If the compatibility / performance KPI does not meet a required criteria / threshold, then in step 13, the UE continues retraining its encoder. Alternatively, responsive to the compatibility / performance KPI does not meet a required criteria / threshold, then the UE may stop and discard the updated encoder. In this regard, if the compatibility / performance evaluation of the UE’s updated encoder is not good / acceptable (i.e. does not meet compatibility / performance criteria), the network may request / configure the UE to fall back to old encoder version.
[0181] If the compatibility / performance KPI does meet a required criteria / threshold, the UE may use the updated encoder. In this regard, if the compatibility / performance evaluation is good / acceptable (i.e. meets compatibility / performance criteria) then the network may notify the UE to drop the old encoder version and keep using the new version.
[0182] The procedure discussed above pertains to a UE updating / retraining its encoder. It is to be appreciated that a similar procedure may be employed where the network updates / retrains its decoder mutatis mutandis.
[0183] In this regard, if the network detects a need for decoder retaining, then: a. the network checks with at least one UE whether its local encoder version has undergone any version updates. b. if No, the network retrains its decoder using its current local version of the encoder. c. if Yes, network requests the at least one UE to send its latest encoder version model.
[0184] The network may request plural UEs, whose encoder has recently been updated and verified, to send their latest encoder version models, whereupon the network combines knowledge from the received encoders to create an encoder version at its side. The network’s selection of UEs to transfer their encoder may be based on a performance evaluation that is carried out at the network after each encoder retraining session is completed at the respective UE side.
[0185] The network then updates / retrains its decoder using the received (or merged) latest version of the encoder. d. the network triggers a model performance evaluation using an intermediate output received from the at least one UE (i.e. CSI reports encoded by the latest version of the encoder). i. if the performance evaluation is not good, then the network falls back to its old / former decoder version. ii. if the performance evaluation is good, then the network notifies the at least one UE and transfers the new version of the decoder to the at least one UE in case it needs to retrain its local encoder.
[0186] Fig. 8 schematically illustrates an example of a signaling diagram in accordance with the subject matter described herein. The method of FIG. 8 may be used where the first and second side models are not strongly dependent upon one another. The procedure shown in Fig. 8 is suitable for a UE updating / retraining its encoder side model in Type 2 training (i.e. joint training without model transfer). Fig. 9 schematically illustrates an example of a signaling diagram in accordance with the subject matter described herein. The method of FIG. 9 may be used where the first and second side models are not strongly dependent upon one another. The procedure shown in Fig. 9 is suitable for a UE updating / retraining its encoder side model in Type 3 training (i.e. separate training).
[0187] Fig. 10 schematically illustrates an example of a signaling diagram in accordance with the subject matter described herein. The method of FIG. 10 may be used where the first and second side models are not strongly dependent upon one another. The procedure shown in Fig. 10 is suitable for a UE updating / retraining its encoder side model in Type 3 training (i.e. separate training). The procedure in Fig. 10 provides autonomous UE side training (i.e. the UE does not initially seek permission / approval from the network prior to updating its encoder). Here the procedure enables the UE to update its side model on its own. However, in the event the UE is unable to train / retrain its encoder until compatibility / performance criteria is satisfied, or the UE is unable to do so in a timely manner (i.e. within a predetermined time period), a timer is provided. If the UE’s updated encoder cannot meet the compatibility / performance requirements within the time window, the UE reverts to a last working model, or re-synchronizes with a working.
[0188] A similar procedure, with a timer, can be used for enabling the networkto update its decoder model autonomously, mutatis mutandis.
[0189] Each of the above methods for updating a side of a two-sided model are applicable to timedependent two-side models.
[0190] Another alternative to those discussed above is to train new side models for the UE and the network in a server (e.g. Next Generation Radio Access Network, NGRAN; 5G and beyond; Over-The-Top, OTT; RAN Intelligent Controller, RIC) and the model update procedure 'over- the-air’ comprises adding a new model-ID pair to a list of supported model-ID pairs for a given functionality. As part of the procedure, a performance monitoring request may be triggered to check the performance of the new model pair.
[0191] Fig. 11 schematically illustrates an overview of knowledge distillation, KD, where a large model, potentially hosted on a server / network side provides distilled knowledge to a smaller model potentially hosted at a client / UE. This distillation is based on client’s requirements / preferences.
[0192] Building a large-scale ML model to enhance performance and to ensure generalizability is no longer an issue given the availability of processing resources involving multi-processing with multiple GPU(s) and very large amounts of physical memory e.g., in a cloud. However, deploying such model in a UE is still challenging due to several reasons. These challenges include, but are not limited to: limited processing resources being available at the UE; bandwidth between gNB and UE may not be sufficient to transmit extremely large ML models; the UE may require a specific model that is optimal for a particular time period (i.e. a ML model applied at the UE is time dependent); and so on. Particularly, for the latter case, a generalized model at gNB is not sufficient to infer the accurate decision made by the UE. In this situation, delivering the model directly from gNB to UE, using model compression, or fine-tuning will not solve the issue. Rather, a more sophisticated solution, such as knowledge distillation, is required.
[0193] Fig. 11 shows a ML model based on the multi-layer perceptron. Distillation is the process through which a "source” ML model with multiple hidden layers is reduced to a "target” ML model with fewer hidden layers but maintaining the number of input and output layers constant. The resultant ML model shares a subset of the "characteristics” of the source model with respect to performing the inference i.e., on a data set that is more limited than what can be processed using the source model.
[0194] Knowledge distillation is a powerful tool to reduce a large and cumbersome model or ensemble of models and create an efficient and small model fit for embedded devices like a 3GPP 5G UE. The idea behind this is to distil the knowledge / learning from a large model (often denoted as teacher) to a smaller, yet efficient, model (often called as student model). Model pruning (reducing the number of hidden layers), model compression (finding a more compact representation of the ML model), continual learning (evolving the ML model based on incremental training when suitable data is available), weight quantization (reducing the number of bits consumed by the model weights and biases), decomposition (splitting a ML model into two or more components allowing reuse of the components) are similar but different concepts from KD and this technique can be combined to accelerate inference and to reduce memory footprint for the student model, which forms the updated side model.
[0195] The methods illustrated in Figs. 4 - 10 can be considered to illustrate a plurality of methods, in the sense that they can be considered to illustrate actions performed by the UE 110 and actions performed by the NW (i.e. gNB 120 or core network node 130). These Figures can therefore be considered to illustrate a plurality of individual methods performed by each of the UE and NW.
[0196] The flowcharts of Figs. 4 - 10 each represent one possible scenario among others. The order of the blocks shown is not absolutely required, so in principle, the various blocks can be performed out of order. Not all the blocks are essential. In certain examples one or more blocks can be performed in a different order or overlapping in time, in series or in parallel. One or more blocks can be omitted or added or changed in some combination of ways.
[0197] The blocks illustrated in Figs. 4 - 10 can represent actions in a method, functionality performed by an apparatus, and / or sections of instructions / code in a computer program. The component blocks illustrated in Figs. 4 - 10 are functional and the functions described can be performed by a single physical entity (such as is described with reference to Fig. 12). The functions described can also be implemented by a computer program (such as is described with reference to Fig. 13).
[0198] It will be understood that each block and combinations of blocks illustrated in Figs. 4 - 10, as well as the further functions described above, can be implemented by various means, such as hardware, firmware, and / or software including one or more computer program instructions. For example, one or more of the functions described below can be performed by a duly configured apparatus (such as an apparatus, UE, or NW node comprising means for performing the below described functions). One or more of the functions described below can be embodied by a duly configured computer program (such as a computer program comprising computer program instructions which embody the functions described below and which can be stored by a memory storage device and performed by a processor).
[0199] As will be appreciated, any such computer program instructions can be loaded onto a computer or other programmable apparatus (i.e. hardware) to produce a machine, such that the instructions when performed on the programmable apparatus create means for implementing the functions specified in the blocks. These computer program instructions can also be stored in a computer-readable medium that can direct a programmable apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the blocks. The computer program instructions can also be loaded onto a programmable apparatus to cause a series of operational actions to be performed on the programmable apparatus to produce a computer-implemented process such that the instructions which are performed on the programmable apparatus provide actions for implementing the functions specified in the blocks.
[0200] Various, but not necessarily all, examples of the present disclosure can take the form of a method, an apparatus or a computer program. Accordingly, various, but not necessarily all, examples can be implemented in hardware, software or a combination of hardware and software.
[0201] Various, but not necessarily all, examples of the present disclosure are described using flowchart illustrations and schematic block diagrams. It will be understood that each block (of the flowchart illustrations and block diagrams), and combinations of blocks, can be implemented by computer program instructions of a computer program. These program instructions can be provided to one or more processor(s), processing circuitry or controller(s) such that the instructions which execute on the same create means for causing implementing the functions specified in the block or blocks, i.e. such that the method can be computer implemented. The computer program instructions can be executed by the processor(s) to cause a series of operational block / steps / actions to be performed by the processor(s) to produce a computer implemented process such that the instructions which execute on the processor(s) provide block / steps for implementing the functions specified in the block or blocks.
[0202] Accordingly, the blocks support: combinations of means for performing the specified functions; combinations of actions for performing the specified functions; and computer program instructions / algorithm for performing the specified functions. It will also be understood that each block, and combinations of blocks, can be implemented by special purpose hardware-based systems which perform the specified functions or actions, or combinations of special purpose hardware and computer program instructions.
[0203] Fig. 12 schematically illustrates a block diagram of an apparatus 10 for performing the methods, processes, procedures and signalling described in the present disclosure and illustrated in Figs. 4 - 10, in this regard the apparatus can perform the roles of either a UE 110, or a network node 120,130 in the illustrated and described methods.
[0204] The apparatus comprises a controller 11 , which could be provided within a device such as a UE 110, or a network node 120, 130.
[0205] The controller 11 can be embodied by a computing device, not least such as those mentioned above. In some, but not necessarily all examples, the apparatus can be embodied as a chip, chip set, circuitry or module, i.e. for use in any of the foregoing. As used here 'module’ refers to a unit or apparatus that excludes certain parts / components that would be added by an end manufacturer or a user.
[0206] Implementation of the controller 11 can be as controller circuitry. The controller 11 can be implemented in hardware alone, have certain aspects in software including firmware alone or can be a combination of hardware and software (including firmware).
[0207] The controller 11 can be implemented using instructions that enable hardware functionality, for example, by using executable instructions of a computer program 14 in a general- purpose or special-purpose processor 12 that can be stored on a computer readable storage medium 13, for example memory, or disk etc, to be executed by such a processor 12.
[0208] The processor 12 is configured to read from and write to the memory 13. The processor 12 can also comprise an output interface via which data and / or commands are output by the processor 12 and an input interface via which data and / or commands are input to the processor 12. The apparatus can be coupled to or comprise one or more other components 15 (not least for example: a radio transceiver, sensors, input / output user interface elements and / or other modules / devices / components for inputting and outputting data / commands).
[0209] The memory 13 stores a computer program 14 comprising instructions (computer program instructions / code) that controls the operation of the apparatus 10 when loaded into the processor 12. The instructions of the computer program 14, provide the logic and routines that enables the apparatus to perform the methods, processes and procedures described in the present disclosure and illustrated in Figs. 4 - 10. The processor 12 by reading the memory 13 is able to load and execute the computer program 14.
[0210] The computer program instructions may be comprised in a computer program, a non- transitory computer readable medium, a computer program product, a machine-readable medium. The term “non-transitory,” as used herein, is a limitation of the medium itself (i.e. tangible, not a signal) as opposed to a limitation on data storage persistency (e.g. RAM vs. ROM). In some but not necessarily all examples, the computer program instructions may be distributed over more than one computer program.
[0211] Although the memory 13 is illustrated as a single component / circuitry it can be implemented as one or more separate components / circuitry some or all of which can be integrated / removable and / or can provide permanent / semi-permanent / dynamic / cached storage.
[0212] Although the processor 12 is illustrated as a single component / circuitry it can be implemented as one or more separate components / circuitry some or all of which can be integrated / removable. The processor 12 can be a single core or multi-core processor.
[0213] The apparatus can include one or more components for effecting the methods, processes and procedures described in the present disclosure and illustrated in Figs. 4 - 10. It is contemplated that the functions of these components can be combined in one or more components or performed by other components of equivalent functionality. The description of a function should additionally be considered to also disclose any means suitable for performing that function.
[0214] Where a structural feature has been described, it can be replaced by means for performing one or more of the functions of the structural feature whether that function or those functions are explicitly or implicitly described.
[0215] Although examples of the apparatus have been described above in terms of comprising various components, it should be understood that the components can be embodied as or otherwise controlled by a corresponding controller or circuitry such as one or more processing elements or processors of the apparatus. In this regard, each of the components described above can be one or more of any device, means or circuitry embodied in hardware, software or a combination of hardware and software that is configured to perform the corresponding functions of the respective components as described above.
[0216] The apparatus can, for example, be a client device, a server device, a mobile cellular telephone, a base station in a mobile cellular telecommunication system, a wireless communications device, a hand-portable electronic device, a location / position tag, a hyper tag etc. The apparatus can be embodied by a computing device, not least such as those mentioned above. However, in some examples, the apparatus can be embodied as a chip, chip set, circuitry or module, i.e. for use in any of the foregoing.
[0217] In one example, the apparatus is embodied on a hand held portable electronic device, such as a mobile telephone, mobile communication device, wearable computing device or personal digital assistant, that can additionally provide one or more audio / text / video communication functions (for example tele-communication, video-communication, and / or text transmission (Short Message Service (SMS) / Multimedia Message Service (MMS) / emailing) functions), interactive / non-interactive viewing functions (for example web-browsing, navigation, TV / program viewing functions), music recording / playing functions (for example Moving Picture Experts Group-1 Audio Layer 3 (MP3) or other format and / or (frequency modulation / amplitude modulation) radio broadcast recording / playing), downloading / sending of data functions, image capture function (for example using a (for example in-built) digital camera), and gaming functions, or any combination thereof.
[0218] In examples where the apparatus is provided within a UE 110, the apparatus comprises: at least one processor 12; and at least one memory 13 storing instructions that, when executed by the at least one processor 12, cause the apparatus at least to: update a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; send, to the network node, first side model update information for enabling the network node to determine the updated first side model; receive, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and control use of the updated first side model based at least in part on the compatibility information.
[0219] In examples where the apparatus is provided within a network node 120, 130 the apparatus comprises: at least one processor 12; and at least one memory 13 storing instruction that, when executed by the at least one processor 12, cause the apparatus at least to: update a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receive first side model information for enabling the network node to determine the first side model of the two-sided model; determine compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and control use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0220] According to some examples of the present disclosure, there is provided a system (for example at least one UE 110 and a network node 120, 130).
[0221] The above described examples find application as enabling components of: tracking systems, automotive systems; telecommunication systems; electronic systems including consumer electronic products; distributed computing systems; media systems for generating or rendering media content including audio, visual and audio visual content and mixed, mediated, virtual and / or augmented reality; personal systems including personal health systems or personal fitness systems; navigation systems; user interfaces also known as human machine interfaces; networks including cellular, non-cellular, and optical networks; ad-hoc networks; the internet; the internet of things (IOT); Vehicle-to-everything (V2X), virtualized networks; and related software and services.
[0222] The apparatus can be provided in an electronic device, for example, a mobile terminal, according to an example of the present disclosure. It should be understood, however, that a mobile terminal is merely illustrative of an electronic device that would benefit from examples of implementations of the present disclosure and, therefore, should not be taken to limit the scope of the present disclosure to the same. While in certain implementation examples, the apparatus can be provided in a mobile terminal, other types of electronic devices, such as, but not limited to: mobile communication devices, hand portable electronic devices, wearable computing devices, portable digital assistants (PDAs), pagers, mobile computers, desktop computers, televisions, gaming devices, laptop computers, cameras, video recorders, GPS devices and other types of electronic systems, can readily employ examples of the present disclosure. Furthermore, devices can readily employ examples of the present disclosure regardless of their intent to provide mobility.
[0223] Fig. 13, illustrates a computer program 14 which may be conveyed via a delivery mechanism 20. The delivery mechanism 20 can be any suitable delivery mechanism, for example, a machine readable medium, a computer-readable medium, a non-transitory computer- readable storage medium, a computer program product, a memory device, a solid-state memory, a record medium such as a Compact Disc Read-Only Memory (CD-ROM) or a Digital Versatile Disc (DVD) or an article of manufacture that comprises or tangibly embodies the computer program 14. The delivery mechanism can be a signal configured to reliably transfer the computer program. An apparatus can receive, propagate or transmit the computer program as a computer data signal. In certain examples of the present disclosure, there is provided a computer program comprising instructions, which when executed by an apparatus (UE 110), cause the apparatus to perform at least the following or for causing performing at least the following: update a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at a UE and a second side model of the two-sided model is located at a node of a network; send, to the network node, first side model update information for enabling the network node to determine the updated first side model; receive, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and control use of the updated first side model based at least in part on the compatibility information.
[0224] In certain examples of the present disclosure, there is provided computer program comprising instructions, which when executed by an apparatus (network node 120, 130), cause the apparatus to perform at least the following or for causing performing at least the following: update a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at a network node and a first side model of the two-sided model is located at a User Equipment, UE; receive first side model information for enabling the network node to determine the first side model of the two-sided model; determine compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and control use of the updated second side of the two-sided model based at least in part on the compatibility information.
[0225] References to 'computer program’, 'computer-readable storage medium’, 'computer program product’, 'tangibly embodied computer program’ etc. or a 'controller’, 'computer’, 'processor’ etc. should be understood to encompass not only computers having different architectures such as single / multi- processor architectures and sequential (Von Neumann) / parallel architectures but also specialized circuits such as field-programmable gate arrays (FPGA), application specific circuits (ASIC), signal processing devices and other devices. 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.
[0226] As used in this application, the term 'circuitry’ can refer to one or more or all of the following: (a) hardware-only circuitry implementations (such as implementations in only analog and / or digital circuitry) and
[0227] (b) combinations of hardware circuits and software, such as (as applicable):
[0228] (i) a combination of analog and / or digital hardware circuit(s) with software / firmware and
[0229] (ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions and
[0230] (c) hardware circuit(s) and / or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (for example firmware) for operation, but the software may not be present when it is not needed for operation.
[0231] This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit for a mobile device or a similar integrated circuit in a server, a cellular network device, or other computing or network device.
[0232] Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.
[0233] Features described in the preceding description can be used in combinations other than the combinations explicitly described.
[0234] Although functions have been described with reference to certain features, those functions can be performable by other features whether described or not.
[0235] Although features have been described with reference to certain examples, those features can also be present in other examples whether described or not. Accordingly, features described in relation to one example / aspect of the disclosure can include any or all of the features described in relation to another example / aspect of the disclosure, and vice versa, to the extent that they are not mutually inconsistent.
[0236] Although various examples of the present disclosure have been described in the preceding paragraphs, it should be appreciated that modifications to the examples given can be made without departing from the scope of the invention as set out in the claims. For example, whilst the first and second models have been described with respect to being encoders and decoders, it is to be appreciated that examples of the present disclosure are not limited to encoder and decoder side model, but that side models of other types, functionalises in inferences may be used. Also, whilst examples have been described with respect to the input information / data being channel measurements and CSI reports, it is to be appreciated that examples of the present disclosure are not limited to the same but that other types of input information / data may be used.
[0237] The term 'comprise’ is used in this document with an inclusive not an exclusive meaning. That is any reference to X comprising Y indicates that X can comprise only one Y or can comprise more than one Y. If it is intended to use 'comprise’ with an exclusive meaning then it will be made clear in the context by referring to "comprising only one ...” or by using "consisting”.
[0238] In this description, the wording 'connect’, 'couple’ and 'communication’ and their derivatives mean operationally connected / coupled / in communication. It should be appreciated that any number or combination of intervening components can exist (including no intervening components), i.e. so as to provide direct or indirect connection / coupling / communication. Any such intervening components can include hardware and / or software components.
[0239] As used herein, the term "determine / determining" (and grammatical variants thereof) can include, not least: calculating, computing, processing, deriving, measuring, investigating, identifying, looking up (for example, looking up in a table, a database or another data structure), ascertaining and the like. Also, "determining" can include receiving (for example, receiving information), retrieving / accessing (for example, retrieving / accessing data in a memory), obtaining and the like. Also, " determine / determining" can include resolving, selecting, choosing, establishing, and the like.
[0240] "As used herein, a description of an action should also be considered to disclose enabling, and / or causing, and / or controlling that action. For example, a description of transmitting information should also be considered to disclose enabling, and / or causing, and / or controlling transmitting information. Similarly, for example, a description of an apparatus transmitting information should also be considered to disclose at least one means or controller of the apparatus enabling, and / or causing, and / or controlling the apparatus to transmit the information.”
[0241] References to a parameter (for example compatibility or performance) or value of a parameter, should be understood to refer to "data indicative of”, "data defining” or "data representative of” the relevant parameter / parameter value if not explicitly stated (unless the context demands otherwise). The data may be in any way indicative of the relevant parameter / parameter value, and may be directly or indirectly indicative thereof.
[0242] In this description, reference has been made to various examples. The description of features or functions in relation to an example indicates that those features or functions are present in that example. The use of the term ’example’ or 'for example’, 'can’ or 'may’ in the text denotes, whether explicitly stated or not, that such features or functions are present in at least the described example, whether described as an example or not, and that they can be, but are not necessarily, present in some or all other examples. Thus 'example’, 'for example’, 'can’ or 'may’ refers to a particular instance in a class of examples. A property of the instance can be a property of only that instance or a property of the class or a property of a sub-class of the class that includes some but not all of the instances in the class.
[0243] In this description, references to "a / an / the” [feature, element, component, means ...] are used with an inclusive not an exclusive meaning and are to be interpreted as "at least one” [feature, element, component, means ...] unless explicitly stated otherwise. That is any reference to X comprising a / the Y indicates that X can comprise only one Y or can comprise more than one Y unless the context clearly indicates the contrary. If it is intended to use 'a’ or 'the’ with an exclusive meaning then it will be made clear in the context. In some circumstances the use of 'at least one’ or 'one or more’ can be used to emphasise an inclusive meaning but the absence of these terms should not be taken to infer any exclusive meaning. As used herein, “at least one of the following: ” and “at least one of ” and similar wording, where the list of two or more elements are joined by “and” or “or”, mean at least any one of the elements, or at least any two or more of the elements, or at least all the elements.
[0244] The presence of a feature (or combination of features) in a claim is a reference to that feature (or combination of features) itself and also to features that achieve substantially the same technical effect (equivalent features). The equivalent features include, for example, features that are variants and achieve substantially the same result in substantially the same way. The equivalent features include, for example, features that perform substantially the same function, in substantially the same way to achieve substantially the same result.
[0245] In this description, reference has been made to various examples using adjectives or adjectival phrases to describe characteristics of the examples. Such a description of a characteristic in relation to an example indicates that the characteristic is present in some examples exactly as described and is present in other examples substantially as described. In the above description, the apparatus described can alternatively or in addition comprise an apparatus which in some other examples comprises a distributed system of apparatus, for example, a client / server apparatus system. In examples where an apparatus provided forms (or a method is implemented as) a distributed system, each apparatus forming a component and / or part of the system provides (or implements) one or more features which collectively implement an example of the present disclosure. In some examples, an apparatus is re-configured by an entity other than its initial manufacturer to implement an example of the present disclosure by being provided with additional software, for example by a user downloading such software, which when executed causes the apparatus to implement an example of the present disclosure (such implementation being either entirely by the apparatus or as part of a system of apparatus as mentioned hereinabove).
[0246] The above description describes some examples of the present disclosure however those of ordinary skill in the art will be aware of possible alternative structures and method features which offer equivalent functionality to the specific examples of such structures and features described herein above and which for the sake of brevity and clarity have been omitted from the above description. Nonetheless, the above description should be read as implicitly including reference to such alternative structures and method features which provide equivalent functionality unless such alternative structures or method features are explicitly excluded in the above description of the examples of the present disclosure.
[0247] Whilst endeavouring in the foregoing specification to draw attention to those features of examples of the present disclosure believed to be of particular importance it should be understood that the applicant claims protection in respect of any patentable feature or combination of features hereinbefore referred to and / or shown in the drawings whether or not particular emphasis has been placed thereon.
[0248] The examples of the present disclosure and the accompanying claims can be suitably combined in any manner apparent to one of ordinary skill in the art. Separate references to an "example”, "in some examples” and / or the like in the description do not necessarily refer to the same example and are also not mutually exclusive unless so stated and / or except as will be readily apparent to those skilled in the art from the description. For instance, a feature, structure, process, block, step, action, or the like described in one example may also be included in other examples, but is not necessarily included.
[0249] Each and every claim is incorporated as further disclosure into the specification and the claims are embodiment(s) of the present disclosure. Further, while the claims herein are provided as comprising specific dependencies, it is contemplated that any claims can depend from any other claims and that to the extent that any alternative embodiments can result from combining, integrating, and / or omitting features of the various claims and / or changing dependencies of claims, any such alternative embodiments and their equivalents are also within the scope of the disclosure.
Claims
CLAIMSWe claim:
1. A User Equipment, UE, comprising: means for updating a first side model of a two-sided model to form an updated first side model, wherein the first side model is located at the UE and a second side model of the two-sided model is located at a node of a network; means for sending, to the network node, first side model update information for enabling the network node to determine the updated first side model; means for receiving, from the network node, compatibility information indicative of whether and to what extent the updated first side model is compatible with the second side model; and means for controlling use of the updated first side model based at least in part on the compatibility information.
2. The UE of any of claim 1 , wherein the two-sided model comprises at least one selected from the group of: a machine learning, ML, model, a pair of side models comprising a first side model and a second side model jointly configured to perform joint inference, wherein an output from the first model is input into the second model, an autoencoder, and the first and second side models being configured to perform encoding and decoding of channel measurements.
3. The UE of any previous claim, wherein the first side model update information comprises at least one selected from the group of: the updated first side model; an identifier of the updated first side model, a description of the changes made to the first side model to form the updated first side model, information indicative of the configuration, structure or parameters of the updated first side model, an indication that an update of the first side model has occurred, and information for enabling the network node to update a local copy of the first side model to form a local copy of the updated first side model.
4. The UE of any previous claim, wherein the compatibility information comprises: an indication of whether a type and / or format of an output from the updated first side model is compatible with a type and / or format of an input of the second side model; and / or an indication of a configuration of the second side model.
5. The UE of any previous claim, further comprising means for sending, to the network node: information input to the updated first side model, and information output from the updated first side model; and wherein the compatibility information is based at least in part on: the input information and the output information.
6. The UE of any previous claim, wherein the first side model is an encoder configured to encode information and the second side model is a decoder configured to decode the encoded information, and wherein the compatibility information is indicative of whether and to what extent the second side model decoder is able to recover the information encoded by the updated first side model encoder.
7. The UE of any previous claim, wherein the compatibility information is indicative of whether and to what extent the joint performance of the updated first side model and the second side model meets one or more performance criteria; and optionally, wherein the one or more performance criteria comprises at least one selected from the group of: an accuracy of the joint performance of the updated first side model and the second side model; a reliability of the joint performance of the updated first side model and the second side model; an amount of processing resources required in the joint performance of the updated first side model and the second side model; and an amount of network resources required in the joint performance of the updated first side model and the second side model.
8. The UE of any previous claim, wherein the control of the use of the updated first side model comprises at least one of: discarding, retaining, applying, or further updating the updated first side model.
9. The UE of any previous claim, wherein the update of the first side model comprises performing knowledge distillation from a teacher first side model to a student first side model to form the updated first side model, and wherein the UE comprises means for sending knowledge distillation information from the knowledge distillation process to the network node; and optionally wherein the knowledge distillation information comprises at least one selected from the group of: knowledge distillation specific performance key performance indicators; distillation loss; latent space metrics; correlation metrics; configuration of the updated first side model; information indicative of a model structure of the updated first side model; andinformation indicative of model parameters of the updated first side model.
10. The UE of claim 9, wherein the knowledge distillation information is sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the updated first side model and the second side model.
11. A network node comprising: means for updating a second side model of a two-sided model to form an updated second side model of the two-sided model, wherein the second side model of the two-sided model is located at the network node and a first side model of the two-sided model is located at a User Equipment, UE; means for receiving first side model information for enabling the network node to determine the first side model of the two-sided model; means for determining compatibility information indicative of whether and to what extent the updated second side model of the two-sided model is compatible with the first side model of the two-sided model; and means for controlling use of the updated second side of the two-sided model based at least in part on the compatibility information.
12. The network node of claim 11 , wherein the two-sided model comprises at least one selected from the group of: a machine learning, ML, model, a pair of side models comprising a first side model and a second side model jointly configured to perform joint inference, wherein an output from the first model is input into the second model, an autoencoder, and the first and second side models being configured to perform encoding and decoding of channel measurements.
13. The network node of any of claims 11 to 12, wherein the first side model information comprises at least one selected from the group of: the first side model; an identifier of the first side model, and information indicative of the configuration, structure or parameters of the first side model.
14. The network node of any of claims 11 to 13, wherein the compatibility information comprises: an indication of whether a type and / or format of an output from the first side model is compatible with a type and / or format of an input of the updated second side model; and / or an indication of a configuration of the updated second side model.
15. The network node of any of claims 11 to 14, further comprising means for receiving, from the network node: information input to the first side model, and information output from the first side model; and wherein the compatibility information is determined based at least in part on: the input information and the output information.
16. The network node of any of claims 11 to 15, wherein the first side model is an encoder configured to encode information and the second side model is a decoder configured to decode the encoded information, and wherein the compatibility information is indicative of whether and to what extent the updated second side model decoder is able to recover the information encoded by the side model encoder.
17. The network node of any of claims 11 to 16, further comprising means for determining one or more performance criteria of the joint performance of the first side model and the updated second side model; and wherein the compatibility information is indicative of whether and to what extent the joint performance of the first side model and the updated second side model meets the one or more performance criteria; and optionally, wherein the one or more performance criteria comprises at least one selected from the group of: an accuracy of the joint performance of the first side model and the updated second side model; a reliability of the joint performance of the first side model and the updated second side model; an amount of processing resources required in the joint performance of the first side model and the updated second side model; and an amount of network resources required in the joint performance of the first side model and the updated second side model.
18. The network node of any of claims 11 to 17, wherein the control of the use of the updated second side model comprises at least one of: discarding, retaining, applying, or further updating the updated second side model.
19. The network node of any of claims 11 to 18, wherein the update of the second side model comprises performing knowledge distillation from a teacher second side model to a student second side model to form the updated second side model, and wherein the network node comprises means for sending knowledge distillation information from the knowledge distillation process to the UE; and optionally, wherein the knowledge distillation information comprises at least one selected from the group of: knowledge distillation specific performance key performance indicators; distillation loss; latent space metrics; correlation metrics; configuration of the updated first side model;information indicative of a model structure of the updated second side model; and information indicative of model parameters of the updated second side model.
20. The network node of claim 19, wherein the knowledge distillation information is sent based at least in part on the compatibility information and / or an evaluation of a joint performance of the first side model and the updated second side model.
Citation Information
Cited By
Data transmission method and data transmission equipment
CN121814273A
Model pairing for backward compatible models
WO2026137375A1
