Data set sharing transmission instructions for separated dual-side ai / ML-based

Abstract

The present disclosure relates to a method and system for improving the establishment of a bilateral AI / ML-based model implemented in a device (AP1-APX), such as a user equipment (UE), and a device (DE1-DEY), such as a server or gNodeB, gNB, in a wireless network system, by sending configuration information providing instructions regarding at least what is shared for data set sharing between the apparatus AP1-APX and the apparatus DE1-CEY, efficient separate or sequential training of the bilateral model can be produced.

Description

Technical Field

[0001] The present disclosure relates to a method and system for improving the establishment of bilateral AI / ML-based models implemented in devices such as user equipment (UE) and devices such as servers or gNodeBs (gNBs) in a wireless network system. By sending configuration information for dataset sharing between the devices and the devices, the configuration information provides instructions on at least what to share, and efficient separate or sequential training of the bilateral models can be generated. Background Art

[0002] Throughout this specification, any discussion of background art should in no way be considered as an admission that such art is well known or forms part of the common general knowledge in the field.

[0003] In at least the fourth-generation (4G) and fifth-generation technology standard (5G) wireless networks, the emergence of artificial intelligence (AI) / machine learning (ML)-based service implementations has driven the need to research new use cases and propose new potential service requirements for introducing enhanced dataset sharing strategies in the network system, especially for new potential service requirements for performing joint inference across multiple network elements (such as user equipment (UE) and gNodeB (gNB)) using bilateral AI / ML models.

[0004] Therefore, a new research project (SI) RP-213599, "Research on Intelligence (AI) / Machine Learning (ML) for NR Air Interface," was initiated in RAN3#94e, which set the goal to specifically explore the following benefits: leveraging features that enable AI / ML-based algorithms to enhance the air interface to improve performance and / or reduce complexity / overhead. Here, the goal of the SI is to lay the foundation for future air interface use cases leveraging AI / ML technologies, covering, for example, channel state information (CSI) feedback enhancement (such as overhead reduction, accuracy improvement), beam management (such as beam prediction in the time and spatial domains to reduce overhead and latency), or location accuracy enhancement.

[0005] Based on this, in 3GPP TSG RAN WG1 #109e, an implementation of a bilateral AI / ML model was subsequently proposed, mainly for CSI compression as a representative sub-use case, where the bilateral AI / ML model is regarded as at least a paired AI / ML model on which joint inference is performed across multiple network elements such as the UE and gNB. For example, initially, the first part of the inference is performed by the UE, and then the remaining part is performed by the gNB (and vice versa). Therefore, as a starting point, the bilateral model of the research case considers at least the AI / ML-based CSI generation part on the UE side to generate CSI feedback information and send it to a given network element (gNB), and the AI / ML-based CSI reconstruction part on the gNB side for reconstructing CSI from the received CSI feedback information.

[0006] In addition, in 3GPP TSG RAN WG1 #110 and #110-bis-e, different AI / ML model training cooperation methods were specified for the above bilateral AI / ML model, including separate training on both sides of the corresponding model (i.e., where the CSI generation part and the CSI reconstruction part are sequentially trained on their respective sides, for example, starting from the UE side training and then the network side training, or starting from the network side training and then the UE side training, or parallel training). Specifically, for this purpose, it is recommended to approximate the given model training by initially jointly training the CSI generation part and the CSI reconstruction part at a specific first-side network element (such as the UE), and subsequently, after the CSI reconstruction part training is completed, sharing a set of information (such as a dataset) with the second-side network element (such as the gNB), which will use this set of information to be able to train the second-side CSI generation part.

[0007] However, despite the current efforts to improve the stability and adaptability of the said bilateral AI / ML model in wireless network systems, there are still problems. That is, due to the dynamic and proprietary architecture of common wireless networks, the basic information for at least enabling the above bilateral separate training mechanism may be lost or modified, resulting in a significant reduction in the efficiency of the corresponding bilateral model.

[0008] For example, since the above training strategy requires sharing a dataset between at least two different network elements, it is important to know what is being shared as the dataset, because for example, without basic knowledge of the inherent data attributes (such as format), the receiving network element that needs to input the shared dataset for subsequent training may not be able to fully utilize all the data information. Similarly, for example, in the case of bilateral model decoder training across multiple UEs, if the receiving network element plans to consider mixed data across different datasets, sharing datasets of different formats by different network elements may result in troublesome operations. Similarly, due to different ownership or privacy issues within the corresponding network elements, it is also important to know what is considered important at a given network element, because in some cases, a given network element may not disclose certain information about the underlying model, while other network elements may disclose it.

[0009] Therefore, there is a need to propose new methods and apparatuses for AI / ML-based bilateral model training, especially separate model training, which solve some or all of the above problems in an efficient, flexible, and reliable manner. Summary of the Invention

[0010] According to some aspects, the subject matter of the independent claims is provided. Some further aspects are defined in the dependent claims.

[0011] According to a first aspect of the present disclosure, a device in a wireless network may be provided, which includes: one or more processors and a memory storing instructions, which when executed by the one or more processors, cause the device to: receive information for configuring the device to share data that will be used to train a bilateral model, the bilateral model facilitating data communication between the device and another device in the wireless network, the information indicating which data will be shared with the other device; train at least a first side of the bilateral model through a machine learning algorithm based on the training data available at the device; determine, according to the received configuration information, a dataset to be shared from the training data; and send the determined dataset for the other device to separately train at least a second side of the bilateral model.

[0012] According to a second aspect of the present disclosure, the device may also be caused to use the first side of the bilateral model when communicating with the other device, and the configuration information indicates whether the input data and / or output data of the first side of the bilateral model should be shared with the other device during training.

[0013] According to a third aspect of the present disclosure, the device determines a hypothetical second side of the bilateral model during training, and the configuration information indicates whether some data for training the hypothetical second side will be shared.

[0014] According to a fourth aspect of the present disclosure, the configuration information may indicate whether to share a metric regarding the relationship between the first-sided training data and the training data of the hypothesized second side.

[0015] Regarding a fifth aspect of the present disclosure, the configuration information may indicate a pre-configured data set to be shared.

[0016] Regarding a sixth aspect of the present disclosure, the configuration information may indicate when to share the data with the other device, and may indicate at least one of a time instance, a sharing condition, or a sharing trigger.

[0017] Regarding a seventh aspect of the present disclosure, the configuration information may indicate how to share the data with the other device, and may indicate at least one of a format for sharing the data, a communication medium for sharing the data, or a priority of the shared data.

[0018] Regarding an eighth aspect of the present disclosure, the configuration information and / or the shared data set may be transmitted by signaling via an air interface, or transmitted via a remote server between the device and the other device.

[0019] Regarding a ninth aspect of the present disclosure, the configuration information may relate to an initial training of the two-sided model, or to an update of the two-sided model based on new training data.

[0020] Regarding a tenth aspect of the present disclosure, the configuration information may include at least one of the following: sharing information, which includes one or more data set identifiers pointing to a set / batch of data sets to be shared with the other device; format information, which describes a vector including the scale of the data set to be shared; or a reporting configuration, which specifies data to be used for at least a first side of the two-sided model.

[0021] Regarding an eleventh aspect of the present disclosure, the two-sided model may facilitate compression / decompression of channel state information CSI. A first side of the two-sided model may be involved in compression of information about a wireless channel, and a second side of the two-sided model is involved in decompression of the wireless channel information.

[0022] Regarding a twelfth aspect of the present disclosure, the two-sided model may facilitate channel encoding / decoding. A first side of the two-sided model may be involved in encoding data to be transmitted to the other device, and a second side of the two-sided model may be involved in decoding data received by the other device.

[0023] Regarding a thirteenth aspect of the present disclosure, the device may be a user equipment UE in the wireless network.

[0024] Regarding a fourteenth aspect of the present disclosure, a device in a wireless network may also be provided, which includes: one or more processors and a memory storing instructions that, when executed by the one or more processors, cause the device to: send information for configuring a device in the wireless network to share data used for training a bilateral model, the bilateral model facilitating data communication between the device and the device in the wireless network, the information indicating which data the device will share with the device; receive the shared data from the device; and train at least a second side of the bilateral model based on the received data through a machine learning algorithm.

[0025] Regarding a fifteenth aspect of the present disclosure, the device may be a next-generation radio access network NG-RAN node in the wireless network, such as a gNB, and the device may be a user equipment UE in the wireless network.

[0026] Regarding a sixteenth aspect of the present disclosure, the device may train and maintain a UE-specific model, and each model is applied to communication with a single UE.

[0027] Regarding a seventeenth aspect of the present disclosure, the device may receive shared data from multiple UEs and train at least a second side of the bilateral model based on a set of shared data received from different UEs.

[0028] Regarding an eighteenth aspect of the present disclosure, the training of the second side of the bilateral model may be at least based on a reference output of a hypothesized second side determined during the training of the bilateral model by the device, and the reference output may be at least a part of the shared data of the device.

[0029] Regarding a nineteenth aspect of the present disclosure, a first method for training a bilateral model that facilitates data communication between a device and another device in a wireless network may also be provided. The method includes: receiving, by the device, information for configuring the device to share data that will be used for training the bilateral model, the information indicating which data will be shared with the other device; separately training at least a first side of the bilateral model through a machine learning algorithm based on training data available at the device, separately from the second side of the bilateral model; determining, according to the received configuration information, a data set to be shared from the training data; and sending the determined data set for the other device to separately train at least a second side of the bilateral model from the first side.

[0030] Regarding a twentieth aspect of the present disclosure, a second method for training a bilateral model may also be provided, where the bilateral model facilitates data communication between a device in a wireless network and another device. The method includes: sending information for configuring the device in the wireless network to share data that will be used to train the bilateral model, the information indicating which data the device will share with the device; receiving the shared data from the device; and training at least a second side of the bilateral model based on the received data through a machine learning algorithm.

[0031] Regarding a twenty - first aspect of the present disclosure, the configuration information may indicate when to share the data with the other device and / or how to share the data with the other device.

[0032] Regarding a twenty - second aspect of the present disclosure, a computer program may also be provided, which includes instructions for causing the device or the device to execute the above - mentioned methods according to the nineteenth to twenty - first aspects of the present disclosure.

[0033] Regarding a twenty - third aspect of the present disclosure, a memory may also be provided, which stores computer - readable instructions for causing the device or the device to execute the above - mentioned methods according to the nineteenth to twenty - first aspects of the present disclosure.

[0034] Here, the computer program and the memory storing the computer - readable instructions may be directly loaded into the internal memory of the computer and / or sent via a network by means of at least one of uploading, downloading, and pushing processes.

[0035] Therefore, although some exemplary embodiments will be described herein with specific reference to the above applications, it should be understood that the present disclosure is not limited to such fields of use, but is applicable to a broader background.

[0036] It should be noted that it should be understood that the methods according to the present disclosure relate to methods of operating the devices according to the above - mentioned exemplary embodiments and their variants. The corresponding statements regarding the devices equally apply to the corresponding methods, and vice versa. Therefore, for the sake of brevity, similar descriptions may be omitted. In addition, even if not explicitly disclosed, the above aspects can be combined in many ways. Those skilled in the art will understand that these combinations of aspects and features / steps are feasible unless they result in contradictions that are explicitly excluded.

[0037] Embodiments of the disclosed devices may include the use of, but are not limited to, one or more processors, one or more application - specific integrated circuits (ASICs), and / or one or more field - programmable gate arrays (FPGAs). Embodiments of the devices may also include the use of other conventional and / or custom hardware, such as software - programmable processors, such as graphics processing unit (GPU) processors.

[0038] It should be understood that any of the above modifications can be applied, individually or in combination, to the respective aspects to which they relate, unless an alternative is explicitly excluded.

[0039] Thus, according to the present disclosure, the problem of undefined dataset transmission of a bilateral AI / ML-based model between at least one device and equipment implemented in a wireless network can be overcome. In addition, the training dataset can be adjusted according to the specific requirements of network elements, and thus the overall efficiency of bilateral model system integration can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Further details, features, objects, and advantages will become apparent from the following detailed description of the preferred embodiments of the present disclosure in conjunction with the accompanying drawings, in which:

[0041] Figure 1 An example of a bilateral AI / ML model training transmission connection system is shown;

[0042] Figure 2 A sequence related to devices and equipment showing the bilateral AI / ML model training steps, including the first side of the bilateral model at the device, the second side of the bilateral model at the equipment, and the assumed second side at the device;

[0043] Figure 3 Message exchanges between a device defined as a user equipment UE and an equipment defined as a network part such as a gNB are shown, which are used for the implementation of the bilateral AI / ML model between two network elements. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] Hereinafter, a communication network architecture based on 3GPP communication network standards (such as 5G / NR) will be used as an example of a communication network to which the embodiments can be applied to describe different exemplary embodiments, but the embodiments are not limited to such an architecture. It will be apparent to those skilled in the art that the embodiments can also be applied to other types of communication networks, where the mobile communication principle is integrated with D2D (device-to-device) or V2X (vehicle-to-everything) configurations, such as SL (sidelink), for example, Wi-Fi, Worldwide Interoperability for Microwave Access (WiMAX), Personal Communication Service (PCS), Wideband Code Division Multiple Access (WCDMA), systems using Ultra-Wideband (UWB) technology, Mobile Ad Hoc Network (MANET), wired access, etc. In addition, without loss of generality, the description of some examples of the embodiments relates to mobile communication networks, but the principles of the present disclosure can be extended and applied to any other type of communication network, such as a wired communication network.

[0045] The following examples and embodiments should only be understood as illustrative examples. Although the specification may refer to "a", "an" or "some" examples or embodiments in several places, this does not necessarily mean that each such reference relates to the same example or embodiment, nor that the feature only applies to a single example or embodiment. The individual features of different embodiments can also be combined to provide other embodiments. In addition, terms such as "comprising" and "including" should be understood not to limit the described embodiments to only consisting of the features that have been mentioned; such examples and embodiments can also include features, structures, units, modules, etc. that are not specifically mentioned.

[0046] The basic system architecture of some examples of (remote) communication networks to which the embodiments can be applied, including mobile communication systems, can include an architecture with a radio access network subsystem and one or more communication networks of a core network. Such an architecture can include one or more communication network control elements or functions, access network elements, radio access network (RAN) elements, access service network gateways or base station transceivers, such as base stations (BSs), access points (APs), Node Bs (NBs), eNBs or gNBs, distributed units (DUs) or central / central units (CUs), which control the corresponding coverage areas or cells, and one or more communication stations that can communicate via one or more communication beams via one or more channels for transmitting several types of data in multiple access domains. The one or more communication stations are, for example, communication elements or functions, such as user equipment or terminal devices, such as user devices (UEs), or another device with a similar function, such as a modem chipset, chip, module, etc., which can also be a station, element, function or application capable of communicating, such as a UE, element or function that can be used in a machine-to-machine communication architecture, or as a separate element attached to such a communication-capable element, function or application, and so on. In addition, it can include core network elements or network functions, such as gateway network elements / functions, mobility management entities, mobile switching centers, servers, databases, etc.

[0047] The following description can provide further details of alternatives, modifications and variations:

[0048] - "Server", especially gNB, includes, for example, providing NR user plane and control plane protocol termination to the UE and connecting to the 5GC via the NG interface, for example, according to Section 3.2 of 3GPP TS 38.300

[0049] V16.6.0 (2021-06), Nodes Connected to the 5GC. In addition, the following description and the proposed features of the present disclosure are not limited to being applied to the indicated framework, but can also be applied to other generations, such as Long Term Evolution (LTE) technology and eNBs, or other technologies, such as Advanced LTE (LTE-A) technology.

[0050] - A user equipment (UE) may include a wireless or mobile device, a device having a radio interface for interacting with a RAN (Radio Access Network), a smart phone, a vehicle-mounted device, an IoT device, an M2M device, etc. Such a UE or device may include: at least one processor; and at least one memory including computer program code; wherein the at least one memory and the computer program code are configured to use the at least one processor to cause the device to at least perform certain operations, such as for example RRC connection and / or implementation of the above AI / ML model. The UE is configured, for example, to generate a message (e.g., including a cell ID) to be transmitted via radio to, for example, a RAN (e.g., to reach and communicate with a serving cell). The UE may generate, transmit, and receive RRC messages including one or more RRC PDUs (Packet Data Units).

[0051] - In addition, the AI / ML model may be at least defined according to RAN1#109-e as a data-driven algorithm that applies AI / ML techniques to generate a set of outputs based on a set of inputs. Therefore, AI / ML model training is regarded as a process of training the AI / ML model in a data-driven manner (by learning input / output relationships) and obtaining a trained AI / ML model for inference.

[0052] - In addition, AI / ML model inference should be at least defined as a process of using a trained AI / ML model to generate a set of outputs based on a set of inputs. Here, at least a unilateral AI / ML model and a bilateral AI / ML model can be distinguished, where the unilateral AI / ML model is an AI / ML model that performs inference entirely at one network entity such as at the UE or server side (e.g., gNB); the bilateral AI / ML model is a paired model on which joint inference is performed. Here, joint inference can be at least defined as AI / ML inference that is jointly performed across at least two network entities such as UW and gNB, i.e., the first part of the inference can be first performed by the UE and then the remaining part can be performed by the gNB, and vice versa.

[0053] - In addition, "AI / ML model transfer" may at least correspond to delivering a model via an air interface, while parameters of the corresponding model structure known to the receiving end or a new model with parameters may be transmitted. In addition, the delivery may include a complete model or a partial model. In addition, the model may be downloaded,

[0054] i.e., at least transferred from the network to the UE, or the model may be uploaded, i.e., at least transferred from the UE to the network.

[0055] The UE may have different states (e.g., according to Sections 4.2.1 and 4.4 of 3GPP TS 38.331 V16.5.0 (2021-06), which are incorporated herein by reference).

[0056] When the RRC connection is established, the UE is, for example, in the RRC connected (RRC_CONNECTED) state or in the RRC inactive (RRC_INACTIVE) state.

[0057] In the RRC connected state, the UE may:

[0058] ○ Store the AS context;

[0059] ○ Transfer unicast data to / from the UE;

[0060] ○ Monitor the control channel associated with the shared data channel to determine whether data is scheduled for the data channel;

[0061] ○ Provide channel quality and feedback information;

[0062] ○ Perform neighbor cell measurements and measurement reporting.

[0063] The RRC protocol includes, for example, the following main functions:

[0064] ○ RRC connection control;

[0065] ○ Measurement configuration and reporting;

[0066] ○ Establishment / modification / release of measurement configurations (e.g., intra-frequency, inter-frequency, and inter-RAT measurements);

[0067] ○ Setting and releasing of measurement gaps;

[0068] ○ Measurement reporting.

[0069] The general functions and interconnections of the various elements and functions described, which also depend on the actual network type, are known to those skilled in the art and are described in the corresponding specifications. Therefore, for the sake of brevity, their detailed descriptions may be omitted here. However, it should be noted that, in addition to those described in detail below, several additional network elements and signaling links may be employed for communication to or from elements, functions, or applications, such as communication endpoints, communication network control elements (such as servers, gateways, radio network controllers), and other elements of the same or other communication networks.

[0070] The communication network architectures considered in the examples of the various embodiments may also be capable of communicating with other networks such as the public switched telephone network or the Internet. The communication network may also be capable of supporting the use of cloud services for virtual network elements or their functions, where it should be noted that the virtual network part of the telecommunications network may also be provided by non-cloud resources such as internal networks and the like. It should be understood that the network elements and / or corresponding functions of the access system, core network, etc. may be implemented by using any node, host, server, access node, or entity suitable for such purposes. Generally, the network functions may be implemented as network elements on dedicated hardware, as software instances running on dedicated hardware, or as virtualized functions instantiated on a suitable platform such as a cloud infrastructure.

[0071] In addition, network elements, such as communication elements, such as UEs, terminal devices, control elements or functions, such as access network elements, such as base stations / BSs, gNBs, radio network controllers, core network control elements or functions, such as gateway elements, or other network elements or functions described herein, and any other elements, functions or applications can be implemented by software, for example by a computer program product for a computer, and / or by hardware. To perform their respective processing, the devices, nodes, functions or network elements used accordingly may include a number of devices, modules, units, components, etc. (not shown), which are required for control, processing and / or communication / signaling functions. Such devices, modules, units and components may include, for example: one or more processors or processor units, which include one or more processing parts for executing instructions and / or programs and / or for processing data; a storage or memory unit or device, which is used to store instructions, programs and / or data for use as a working area of the processor or processing part, etc. (such as ROM, RAM, EEPROM, etc.); an input or interface device for inputting data and instructions by software (such as a floppy disk, CD-ROM, EEPROM, etc.); a user interface for providing the user with monitoring and manipulation options (such as a screen, keyboard, etc.); other interfaces or devices for establishing links and / or connections under the control of the processor unit or part (such as wired and wireless interface devices, radio interface devices including, for example, antenna units, devices for forming a radio communication part, etc.), etc., where each device forming an interface, such as a radio communication part, may also be located at a remote site (such as a radio head or radio station, etc.). It should be noted that in this specification, a processing part should not be regarded only as a physical part representing one or more processors, but may also be regarded as a logical division of the indicated processing tasks executed by one or more processors. It should be understood that according to some examples, the concept of a so-called "liquid" or flexible network can be adopted, where the operations and functions of network elements, network functions or other entities of a network can be executed in a flexible manner in different entities or functions, such as in nodes, hosts or servers. In other words, the "division of labor" among the network elements, functions or entities involved may vary depending on the situation.

[0072] In addition, when it comes to bilateral AI / ML models, as previously mentioned, the model can be executed and / or trained at at least two network entities, e.g., on the gNB side and / or the UE side. Here, the aforementioned network entities can have one or more available trained or to-be-trained models, which are preferably configured to solve a certain predetermined problem, preferably an optimization problem, which will be further illustrated by examples below. In addition, a given network entity can also have non-ML algorithms implemented internally (e.g., local algorithms in the network entity). Thus, each network can be able to indicate which model the network entity should use at any given time and when to activate this model.

[0073] On the other hand, as pointed out in the above description previously, for bilateral AI / ML model applications, joint inference of different network entity parts such as UE and network side parts is always required, and careful consideration of model training and model updating is needed to ensure that the inference performance remains at a good level (i.e., it is reliable and operates within predetermined limits). Here, as can be discussed, for example, in RAN1#110-bis-e, a separate or sequential training of the bilateral AI / ML model, i.e., a training process where the AI / ML model is initially trained on one model side (e.g., on the UE model side) and then fully trained on the other model side (e.g., on the network model side), can be preferred because such a separate training process allows possible options for simple, accessible, and efficient updates and adaptive changes. However, on the other hand, since for such a training process, the accurate transfer of the training dataset from the first model side to the second model side becomes crucial for achieving efficient model training (and thus inference quality), an adaptive training dataset transfer strategy is needed, especially to adjust the given training data transfer according to the specific attributes of the network entities to which the bilateral model is applied.

[0074] As a result, for example, for such a transfer process, it becomes important to know what to share as the training dataset because specific network entities may be incompatible with certain data attributes such as format. In addition, depending on what is to be shared, some network entities, e.g., gNB, may have different ownership or privacy issues, so not all information transmitted and / or received by a given network entity is accessible. Finally, it can also be reasonably assumed that dataset sharing or updating can be applied offline (e.g., via uploading to a remote server). However, in cases where online transmission (e.g., dataset transmission via the air interface) is also to be used (e.g., because fast model updates can be performed more efficiently in this way), different auxiliary signaling may be needed to control this operation, which leads to the fact that before the actual dataset sharing process, the question of what changes are needed for air interface transmission must also be addressed.

[0075] Therefore, in order to overcome the above problems, the subject matter of the independent claims is proposed. Specifically, it is proposed to provide the above-sought training data transmission strategy by providing, at least to another network entity (further defined as a "device") responsible for training on the first model side, additional configuration information sent by a network entity (further defined as a "facility"), wherein the network entity may preferably be the receiving end of the training data to be shared by the bilateral AI / ML model, and the configuration information may at least include instructions on, for example, which data to share with the facility. As a result, by means of the configuration information, by additionally providing specific transmission information to the sending end of the training data of the bilateral model, a very adaptable and thus efficient bilateral model training can be achieved.

[0076] As a result, by the present invention, there is provided at least a device in a wireless network, comprising: one or more processors and a memory storing instructions, which when executed by the one or more processors, cause the device to: receive information to configure the device to share data for training a bilateral model that facilitates data communication between the device and another facility in the wireless network, the information indicating which data to share with the other facility; train at least a first side of the bilateral model by a machine learning algorithm based on training data available at the device; determine a data set to be shared from the training data according to the received configuration information; and transmit the determined data set for the other facility to separately train at least a second side of the bilateral model from the first site.

[0077] Therefore, through the above disclosure, a mechanism is provided which, by means of specific configuration information, enables an adjusted transmission of the training data of the bilateral model from a first network element (here the above-mentioned device) including the first model side to a second network element (the other facility) including the corresponding second model side of the bilateral model, with the following effect: the corresponding training data to be sent to the other facility can be efficiently adapted to the requirements of the corresponding network element. More specifically, by providing (even before the corresponding bilateral model training on the first and second model sides) specific configuration information that can be used as instructions on what to share from the first network element to the second network element, a given training process can be efficiently customized for the conditions applied between the two network elements, thus resulting in the termination of the above data set transmission problem.

[0078] Here, in some examples, the corresponding device may be one of several devices implemented in the corresponding wireless network. In addition, the other facility may also be only one of several facilities implemented in a given wireless network, and the number of devices and facilities in the present invention may not be limited to a specific quantity. In addition, multiple models may be implemented respectively in the system and its network elements, each model corresponding to a different system configuration / scenario and / or data set to be used.

[0079] In addition, in some examples, it may be preferably made such that the above-described apparatus uses the first side of the bilateral model that communicates with the other device, and the configuration information may indicate whether the input data and / or output data of the first model side during training should be shared with the other device. Here, the input data may, for example, at least refer to the input data at the preprocessing of the first model side of the bilateral model or the input at the first model side of the bilateral model. In addition, the output data may at least refer to the data at the output of the first model side of the bilateral model or the output of the quantizer considered after the first model side (in the case of transmission). Thus, the respective configuration information may include one or more data sets to be shared with the at least other device.

[0080] In another example, the configuration indication may also indicate whether some data for the training of the second model side will be shared and / or whether a measure of the relationship between the first model side training data and the second model side training data will be shared. Thus, in the present disclosure, not only pure input or output information may be included in the corresponding configuration information, but also mathematical assignments such as functions, statistical relationships, or format requirements of the first and / or second model sides will be considered. In addition, in cases where the corresponding data has not yet been included in the data set to be shared with the other device, these measures may preferably also be derived at least based on the above input and / or output data sets of the first model side.

[0081] In addition, in another example, the hypothetical second side of the bilateral model may also be determined during the training of the apparatus, and the configuration information may indicate whether some data for the training of the hypothetical second model side will be shared and / or whether a measure of the relationship between the first model side training data and the training data of the hypothetical second model side will be shared. This particularly has the advantage that additional input parameters, such as the training data generated during the training of the second model side in the apparatus, can also be generated, resulting in supplementary information being used for the construction of the configuration information. Thus, the training data of the hypothetical second model may, for example, preferably be used as feedback for the training in the other device or may equally be implemented in the above measures.

[0082] In addition, the configuration information may also indicate a preconfigured data set to be shared. Here, the preconfigured data may include, for example, data from a device capability report.

[0083] In another example, the configuration information may similarly indicate when to share the data with the other device, and indicate at least one of a time instance, a sharing condition, or a sharing trigger. Specifically, in one example, the time to share the data set from the device to the at least one other device may be defined by completing an initial first model-side training, where, in this case, the data set corresponding to the model training may be shared with the at least one other device.

[0084] In another example, the time for sharing a given data set may also be defined by at least a first model-side update. In particular, the device may share the data set before or after performing a model update. Alternatively, when, for example, the first model side needs to update a part of the device for inference, the data set corresponding to the model update (even if completed by the device before the network agrees) may not be directly shared with the at least one other device, but the device may initially trigger a model update request and then perform data sharing, which results in a more secure and efficient transmission mechanism.

[0085] Furthermore, in yet another specific example for model update, a potential trigger for model update / data set sharing may be determined based on parallel model monitoring activities. In particular, when the first model side and another model side of a bilateral model, such as, for example, the aforementioned hypothetical second model side, observe different performances, a model update (which may also result in a data set update) may be triggered, for example, such that the claimed subject matter includes an automatic model quality assessment strategy while providing the effect of corresponding configuration information.

[0086] In addition, in another example, the configuration information may further indicate how to share the data with the other device, and indicate at least one of a format for sharing the data, a communication medium for sharing the data, or a priority of the shared data. Specifically, the format of the data set may preferably be received and / or defined by the other device, which may preferably be an implementation element of the network, such as a gNB. In addition, the format of the data set may preferably include, for example, the scale or size of the data set (sample size), the ordering of the information carried by the data set, quantization information corresponding to information on whether quantization is applied to data entries, or any other type of information that may allow the other party (specifically the other device) to interpret the data set.

[0087] In addition, the configuration information and / or the determined data set may preferably be sent via signaling over the air interface or via a remote server implemented in the network between the device and the other device.

[0088] Here, the indication of whether the dataset should be reported via one of the above-mentioned network means (or even other means that may be known to the device described above before this indication) can preferably be defined by one of the following aspects.

[0089] If the dataset sharing can be done via the air interface, any configuration defining the reporting resources and / or periodicity and / or the maximum size per transmission batch or report can be defined by the at least another device, which is preferably an implemented network element such as a gNB. Thus, the device can be configured to send the training dataset to be shared with the other device at least in one or more transmission batches based on the received configuration information.

[0090] On the other hand, if the dataset sharing can be done via a remote server, any configuration of the confirmation report after successful upload of the dataset can be initially defined by the other device. Based on this, the device can then be configured to upload the training dataset and send the confirmation of the dataset upload to the other device via higher layers and / or dynamic signaling, such as via the air interface.

[0091] In addition, the configuration information may also include different priorities considered for different data in the dataset. For example, some data may be equipped with a lower priority tag, resulting in the exclusion of this data from the dataset in case the dataset sharing resources may be limited. Thus, the device or the other device can preferably be configured to equip the corresponding data implemented in the shared dataset with a lower priority tag, and in case of detecting low dataset sharing resources within the network (e.g., by additional detection means), an additional analysis entity implemented in the device or the other device can analyze the shared dataset and exclude the low-priority data before the sending process.

[0092] In another example, the configuration information may preferably relate to the initial training of the bilateral model or to the update of the bilateral model based on new training data.

[0093] Thus, the other device may also be configured, for example, to indicate the same or different data set collection settings to a plurality of different devices, and thus consider using the most reliable or higher quality data sets to train / update the second model side of the bilateral model. To this end, the other device may be configured, for example, to compare the data sets received from different devices and train or update the second model side portion of the bilateral model at least based on a predetermined quality identifier. To this end, in one variant, the other device may be configured, for example, to compare the data received from different data sets and identify the quality of the transmitted data by the difference between a first value of a metric reported by the corresponding device and a second value of a metric. Here, this difference in the metric may represent, for example, the relationship between the input used by the first model side of the bilateral model and the corresponding output of the second model side, resulting in the fact that a small detected difference may result in a high-quality identifier and thus a data set preferably used for the training and / or updating of the second model side. Thus, by additional analysis steps, an even more efficient training and updating strategy for at least the second model side can be generated.

[0094] In addition, in another example, the configuration information may further include at least one of the following: sharing information, which includes one or more data set identifiers pointing to a given set or batch of data sets to be shared with the other device; format information, which describes a vector including the scale of the data sets to be shared; or a reporting configuration, which specifies the data to be used for at least the first side of the bilateral model. Specifically, in the sharing information, each data set may be associated with a predetermined ID, which may be unique and serve a specific deployment scenario for a given use case. In addition, the format information may also include quantization information of the corresponding data, i.e., for example, the bit width of the corresponding data samples or the number of training samples used by the device together with the sharing information. Finally, the reporting configuration may preferably be constructed at least based on a combination of the following data:

[0095] - The input at the preprocessing of the first model side of the bilateral model,

[0096] - The input of the first model side of the bilateral model,

[0097] - The output of the first model side of the bilateral model,

[0098] - The output of the quantizer considered after the first model side of the bilateral model,

[0099] - The output of the hypothetical dequantizer considered before the hypothetical second model side of the bilateral model,

[0100] - The output of the hypothetical second model side of the bilateral model,

[0101] - The output of the postprocessing of the hypothetical second model side of the bilateral model.

[0102] In another example, the corresponding bilateral model can also particularly facilitate the compression / decompression of channel state information CSI. The first model side can specifically involve the compression of information about the wireless channel, and the second model side can involve the decompression of wireless channel information. This leads to a direct contribution and utilization of the claimed subject matter to the above RAN standardization, specifically related to the standards of 3GPP TSG RAN WG1 #110 and #110-bis-e. In other examples, it is also possible that the corresponding bilateral model can be used in other network methods, including, for example, facilitating channel coding / decoding, so that the corresponding inventions can be applied to other use cases and even standards (such as 6G).

[0103] Furthermore, in one example, the device can correspond to a user equipment UE in a wireless network, and the other device can correspond to a server, particularly a gNB of the wireless network. Vice versa, in another example, the device can also be defined as a server, particularly a gNB, and the other device can be one or even multiple UEs. Therefore, an improved bilateral AI / ML-based model can be implemented across UEs and servers, which leads to efficient and user-friendly transmission interactions.

[0104] Alternatively, there can also be provided a device in a wireless network, which includes: one or more processors and a memory storing instructions. When the instructions are executed by the one or more processors, the device is caused to: send information to configure a device in the wireless network to share data used for training a bilateral model, the bilateral model facilitating data communication between the device and the other device in the wireless network, the information indicating which data the device is going to share with the other device; receive the shared data from the device; and based on the received data, train at least a second side of the bilateral model through a machine learning algorithm.

[0105] Here, in one example of the second aspect, the device can be a next-generation radio access network NG-RAN node in the wireless network, such as a gNB, and the device is a user equipment UE in the wireless network.

[0106] In addition, in another example, the device may be configured to train and maintain UE-specific models, each model being applied to communication with an individual UE. Specifically, as described above, each UE may include one or more bilateral models that are at least connected to the device through the above-described joint inference. In addition, in order to receive a given data set for separate bilateral model training, the device may receive the corresponding data set at least via downloading from a remote server or network, placement by a corresponding device, and / or via an air interface existing between the device and a given device. In addition, the device may again be configured to indicate the same or different data set collection settings to different UEs, and may be configured to consider training and / or updating the second model side of the bilateral model using at least the most reliable data set(s) or higher quality data set(s), as already described above. As a result, in another example, the device may also thus receive shared data from multiple UEs and train at least the second model side based on a collection of the received shared data from different UEs.

[0107] In another example, the training of the second side of the bilateral model may similarly be based at least on a reference output of a hypothesized second side determined during the training of the bilateral model by the device, where the reference output is at least a part of the shared data of the device. Specifically, since the bilateral model may preferably include joint inference at least across the device and the device, the required reference output data of the hypothesized second model side may be used as feedback data, which is generally included in the shared data set for the training of the second model side. Alternatively, the reference output data may also be shared with the device separately and / or during different training cycles, resulting in the formation of a persistent feedback mechanism within the transmission system.

[0108] In addition, according to another aspect of the present invention, there may also be provided a first method for training a bilateral model that facilitates data communication between a device in a wireless network and another device. The method is preferably implemented in the device and includes: receiving, by the device, information to configure the device to share data that will be used for training the bilateral model, the information indicating which data will be shared with the other device; training at least a first side of the bilateral model separately from the second side of the bilateral model by a machine learning algorithm based on training data available at the device; determining a data set to be shared from the training data according to the received configuration information; and sending the determined data set for the other device to use for training at least a second side of the bilateral model.

[0109] In addition, according to another aspect of the present invention, there can also be provided a second method for training a bilateral model facilitating data communication between a device in a wireless network and another device, which is preferably implemented in the device, including: sending information to configure the device in the wireless network to share data to be used for training the bilateral model, the information indicating which data the device is going to share with the device; receiving the shared data from the device; and training at least a second side of the bilateral model based on the received data through a machine learning algorithm.

[0110] In at least one of the methods mentioned above, by way of example, the configuration information may again indicate when to share data with the other device and / or how to share data with the other device. In addition, one of the other features specified for the above-mentioned device and / or device can also be applied to the above method, so it will not be repeated in this section.

[0111] In addition, according to another aspect of the present invention, a network system can be similarly provided, including: at least one of the devices mentioned above and at least one of the devices mentioned above, wherein the at least one device and the at least one device are connected via a wireless network and are configured to share a data set for training a bilateral model implemented across the at least one device and the at least one device based on the method of sending and / or receiving the foregoing configuration information. In addition, any feature specified for the above-mentioned device, device, and / or method can also be applied to the elements in the system, and no further description will be repeated in this section.

[0112] In addition, according to another aspect of the present invention, there can also be provided: a computer program including instructions for causing the previously described device or device to execute any of the above methods; and a memory storing computer-readable instructions for causing the previously described device or device to execute any of the above methods.

[0113] Now referring to the drawings, Figure 1 A first example of a feasible connection system between a plurality of devices AP1 to APX responsible for training and applying a first side of a bilateral model and a plurality of devices DE1 to DEY responsible for training and applying a second side of the corresponding bilateral model implemented in the devices AP1 to APX is shown. Here, one or more bilateral models can be implemented within the system, and each bilateral model corresponds to a different configuration, scenario, and / or data set to be used. For this purpose, each of the devices AP1 to APX maintaining the corresponding first side of the bilateral model can be respectively connected to the corresponding devices DE1 to DEY maintaining the second side of the given bilateral model.

[0114] Thus, within a given system, joint training (where the first model-side training and the second model-side training are performed in the same training loop, also mentioned in RAN1#110-bis-e) or separate training (i.e., sequential training starting with the training of the first model side and then the second model side, or even parallel training) can be applied to the model training for device-device cooperation, while at the same time, the model training and update can be performed at least offline (i.e., without any model transfer) or online.

[0115] Additionally, at least in the latter case, for the transmission of a specific training dataset from the first-side model implemented in devices AP1 to APX to the second-side model implemented in devices DE1 to DEY, devices AP1 to APX can be configured to share the corresponding training dataset based on a predetermined condition via a remote server / carrier server 4 or a direct air interface connection S101A to S101X. Thus, a given device AP1 to APX can, for example, upload a given training dataset to the remote server / carrier server 4 via a server connection S102, and the remote server / carrier server 4 is configured to store the training dataset in a corresponding data collection and storage unit 2. Subsequently, the corresponding devices DE1 to DEY holding the respective second model sides of the transmitting devices AP1 to APX can be configured to download the stored training dataset for the subsequent second model-side training. Here, the upload and download processes to the remote server / carrier server 4 can preferably be performed by adding a predetermined identifier to the uploaded training dataset. Specifically, by uploading a dataset from one of the devices AP1 to APX, a given training dataset can be tagged with an identifier, for example, stored as metadata in the data collection and storage unit 2, which can be requested by the corresponding devices DE1 to DEY for a specific download process. On the other hand, a given training dataset or update can also be directly shared from the devices AP1 to APX to the given devices DE1 to DEY via a higher layer and / or dynamic signaling, such as, for example, via the air interface connection S101A to S101X.

[0116] Furthermore, Figure 2 shows a sequence of a bilateral AI / ML-based training process according to an exemplary embodiment of a given invention, particularly a separate bilateral AI / ML-based training process, which is also implemented in Figure 1 the shown transmission system.

[0117] Here, it can be done by Figure 2The training on the first model side implemented in the device AP1 shown is used to initiate a corresponding separate bilateral model training process. Specifically, the first model side training can start with a predetermined input X, such as a predefined data set used as the input for the preprocessing S201 of the device part of the corresponding bilateral model. Subsequently, the preprocessed data can be used as the input Y for the actual first (i.e., device AP1) side model training S202, and the output Z of the first side model training is then quantized S203, and the output M' or M" is used for subsequent training data transmission.

[0118] Furthermore, since for the separate bilateral AI / ML model training, the first model side is configured to generate a full bilateral model training process (while at the same time not communicating with the second model side in each training iteration step), the device AP1 may also include an additional so-called "hypothetical second model side HDE1" for also training the complementary second model side part at the device AP1 (mainly to complete each training loop iteration without explicitly transmitting the training data to the network entity holding the second model side, i.e., the device DE1). At this time, the training process of the hypothetical second model side can preferably be equivalent to or at least similar to the training process carried out on the second model side, specifically, so as to allow the accurate output of the training data from the first model side. As a result, the training process of the hypothetical second model side can at least include a first hypothetical dequantization step, in which the quantized output M' from the first model side of the device AP1 is dequantized to conform to the conditions of the hypothetical second model side. Thereafter, the so-generated output Z' of the hypothetical dequantizer S204 can then be input again into the hypothetical second (i.e., device) model side training algorithm S205 for training the hypothetical second model side, where the corresponding output of S205, hereinafter defined by Y', can be further refined through a hypothetical post-processing step S206 before the final post-processing training output X' of the hypothetical second model side can be generated.

[0119] As mentioned above, the hypothetical training steps carried out in this way can be only for the separate training of the first model side and thus are not required for subsequent model inference. In a further preferred embodiment, the so-generated hypothetical outputs Z', Y' and / or X' may not be discarded after generation but are used for feedback to the network system, as will be described below.

[0120] Furthermore, considering the (actual) second model side training carried out at the corresponding device DE1, as mentioned above, the required training process can preferably be equivalent to or at least similar to the processing steps of the hypothetical second model side training. Therefore, after the first model training is completed, a predefined and quantized training data set can be generated as the output M" on the first model side. Subsequently, the output M" can (for example, via Figure 1The described transferable options) are shared / transferred to the device DE1 holding the second model side for second model side training, where in S207, it is initially dequantized and output as the corresponding dequantized training dataset output Z”, and is used for second model side training in the second model side training algorithm process S208. Thereafter, before the final post-processed training output X” can be used for other purposes (such as actual AI / ML model applications, inference generation, storage, etc.), the thus-generated training output Y” can be post-processed again in step S209.

[0121] Now, to address the aforementioned issues and to further improve the transfer / sharing of the training dataset from the first model side (i.e., the aforementioned devices AP1 to APX) to the second model side (i.e., the aforementioned device), the claimed invention may include the following:

[0122] For the devices AP1 to APX that support bilateral AI / ML models, the devices AP1 to APX may, for example, directly receive configuration information from the device DE1 holding the second model side and via a higher layer and / or a dynamic signaling method such as via an air interface, or indirectly (e.g., via the remote server 4), where the configuration information provides instructions on what to share at least for the dataset sharing of the second model side, so as to enable model training and / or update of the device (i.e., the second model) side part of the bilateral model, and where the instructions may be at least one of the following:

[0123] The shared dataset may be constructed based on multiple training samples, where each training sample may be constructed based on at least one or a combination of the following data, also described in the previous Figure 2 as:

[0124] - The input X at the preprocessing of the first model side part of the bilateral model,

[0125] - The input Y of the first model side part of the bilateral model,

[0126] - The output Z of the first model side part of the bilateral model,

[0127] - The output M’ of the quantizer considered after the first model side part of the bilateral model,

[0128] - The output Z’ of the hypothetical dequantizer considered before the hypothetical second model side part of the bilateral model,

[0129] - The output Y’ of the hypothetical second model side part of the bilateral model,

[0130] - The output Z’ of the post-processing of the hypothetical second model side part of the bilateral model.

[0131] In addition, each training sample may further include additional metrics, which are derived based at least on any one or a combination of the above data X, Y, Z, M’, Z’, Y’, Z’. For example, the metric may represent the relationship between the input X used by the first model-side part and the corresponding output Y’ of the assumed second model-side part, so that the metric can be defined as f(Y, Y’). In addition, these additional metrics may also be defined by the specification and / or configured for the functions specific to devices AP1 to APX and / or the device (i.e., the first model side), and may also be reported to the aforementioned device DE1 in the dataset or in other ways.

[0132] Thus, in one variant, the above configuration information indicating the pre-configured data sharing possible options may be known to the receiving device DE1 (e.g., via an earlier device capability reporting process), and the device may select one of the possible data sharing options in devices AP1 to APX, where the dataset sharing options may be created based on one of the data X, Y, Z, M’, Z’, Y’, Z’ or at least one or a combination of the above additional metrics.

[0133] As a result, in one embodiment, where the corresponding sharing of the training dataset may be used for sub-use cases such as CSI compression using a bilateral model (e.g., for an autoencoder), the above parameters for defining the configuration information may be one or more of the following:

[0134] - The input X at the preprocessing of the first model-side part of the bilateral model, which is at least the corresponding channel measurement,

[0135] - The input Y of the first model-side part of the bilateral model, which is at least the corresponding channel feature vectors of devices AP1 to APX,

[0136] - The output Z of the first model-side part of the bilateral model, which is at least the latent space,

[0137] - The output M’ of the quantizer considered after the first model-side part of the bilateral model, which is at least the CSI feedback,

[0138] - The output Z’ of the assumed dequantizer considered before the assumed second model-side part of the bilateral model, which is at least the dequantized latent space,

[0139] - The output Y’ of the assumed second model-side part of the bilateral model, which is at least the reconstructed channel feature vectors at device DE1, and

[0140] - The output Z’ of the post-processing of the assumed second model-side part of the bilateral model, which is at least the reconstructed channel measurement.

[0141] - Additionally, at least one metric representing the relationship between the input used by the representation device (i.e., the first model) side portion and the corresponding output of the assumed device (i.e., the second model) side portion can be transformed. For example, the function f can be the difference in signal-to-interference plus noise ratio (SINR), cosine similarity, or normalized mean squared error (NMSE). Additionally, this metric can also be used as an indication of the need for model update. For example, if devices AP1 to APX adopt different second model side assumptions, they can calculate the difference f(.) metrics, such as f1(V, V1), f2(V, V2), f3(V, V3), where V1, V2, V3 are different network reconstruction options.

[0142] Furthermore, in another embodiment, the corresponding sharing of the training dataset can also be used for another sub-use case, particularly for the use case of channel coding / decoding using a bilateral model (e.g., for an autoencoder). In this case, the above parameters can similarly be one or more of the following parameters:

[0143] - The input X at the preprocessing of the first model side portion of the bilateral model is at least the transport block before codeblock segmentation and cyclic redundancy check attachment (CRC attachment) (these steps can be preprocessing steps).

[0144] - The input Y of the first model side portion of the bilateral model is at least the codeblock with attached CRC.

[0145] - The output Z of the first model side portion of the bilateral model is at least the coded bits after channel coding.

[0146] - The output M’ of the quantizer considered after the first model side portion of the bilateral model is at least the rate matcher, and the transmitted block can be the coded block with rate matching.

[0147] - The output Z’ of the assumed dequantizer considered before the assumed second model side portion of the bilateral model is at least the recovered coded bits.

[0148] - The output Y’ of the assumed second model side portion of the bilateral model is at least the reconstructed codeblock with attached CRC, and

[0149] - The output Z’ of the post-processing of the assumed second model side portion of the bilateral model is at least the recovered transport block.

[0150] - Additionally, at least one metric f(V, V’) representing the relationship between the input used by the representation device (i.e., the first model) side portion and the corresponding output of the assumed device (i.e., the second model) side portion can be, for example, a function corresponding to the block error rate (BLER) or bit error rate (BER), etc.

[0151] In addition, in this embodiment, the objective of each bilateral model can be to align on a code construction sequence, a channel encoder, or any other parameter associated with a given channel coding process.

[0152] In addition to the above instructions regarding at least what to share for the dataset sharing of the second model side to implement the model training and / or update of the device (i.e., the second model) side part of the bilateral model, the instructions of the configuration information can also be provided with instructions regarding when to share the corresponding training data.

[0153] In one embodiment, devices DE1 to DEY can be configured to select relevant parameters and conditions for devices AP1 to APX for dataset sharing, where the parameters can at least define the time instances, conditions, or triggers that allow devices AP1 to APX to share one or more training datasets. Here, the corresponding instructions can at least include the following:

[0154] - In one example, the trigger or time instance can be defined as when devices AP1 to APX complete the initial first model side training of the bilateral model, such that the dataset corresponding to the model training can be shared with devices DE1 to DEY.

[0155] - In another example, the trigger or time instance can be defined as when the first model side needs to update the first model side part of the bilateral model for inference. Here, the dataset corresponding to the model update (even if this has been completed by devices AP1 to APX before the consent of devices DE1 to DEY) may not be directly shared with the network. Instead, devices AP1 to APX can be configured to initially trigger a model update request and then perform dataset sharing.

[0156] - In addition, in an additional variant, especially in the case of model update, potential triggers for model update / dataset sharing can be determined based on parallel monitoring activities. In particular, when the first model (i.e., devices AP1 to APX) side part of the bilateral model and the assumed second model side part (assuming there are multiple devices AP1 to APX) observe different performances, a model update (which subsequently leads to dataset update) can be triggered. For example, as described above, if devices AP1 to APX can calculate different performance metrics, such as fl(V, V1), f2(V, V2), f3(V, V3), where V1, V2, or V3 can be different second model side reconstruction options available at a given device AP1 to APX, then devices AP1 to APX can trigger the model update and / or dataset sharing process according to the values of metrics fl, f2, or f3.

[0157] In addition to the above instructions regarding at least what to share and / or when to share for dataset sharing on the second model side to implement model training and / or updating of the model training and / or updating of the device (i.e., the second model) side of the bilateral model, the instructions for the configuration information may also be provided with instructions on how to share the corresponding training data. Here, the corresponding instructions may at least include one or more of the following:

[0158] - In one embodiment, the instructions may specify the format of the training dataset to be shared. Specifically, the format of the dataset may correspond to the dataset scale or size (e.g., sample size), the sorting of the information carried in the dataset, whether quantization is applied to the data entries, or other information that allows the other party (i.e., devices DE1 to DEY) to interpret the dataset.

[0159] - Additionally, in another embodiment, the instructions may further include an indication of whether the dataset should be reported via the air interface, via the above-mentioned remote server (i.e., Figure 1 the remote server / carrier server 4 mentioned above) or any other means that the given devices AP1 to APX should be aware of before this indication.

[0160] - Specifically, if the dataset sharing can be completed via the air interface, any configuration defining the reporting resources and / or periodicity and / or maximum size per batch / report may be defined by devices DE1 to DEY. Thus, the devices may send the training dataset in one or more batches based on the received configuration.

[0161] - Alternatively, if the dataset sharing can be completed via the remote server / carrier server 4, any configuration that can define the confirmation report after successfully uploading the dataset S102 may be defined by devices DE1 to DEY. Here, devices AP1 to APX may upload the training dataset and send the confirmation of the dataset upload S102 to devices DE1 to DEY via the air interface.

[0162] - Additionally, the same or different priorities may also be considered for the data in the dataset. If the dataset sharing resources may be limited, some data may also be excluded from the dataset.

[0163] As a result, using the corresponding configuration information indicated by the respective devices DE1 to DEY above, the dataset to be shared from devices AP1 to APX to devices DE1 to DEY for bilateral AI / ML-based model training can be efficiently adapted to the requirements and characteristics of each network entity.

[0164] In addition, to implement the given configuration information provided from devices DE1 to DEY to apparatuses AP1 to APX, the configuration information may include at least one or more information elements (IEs), which may preferably be signaled as part of a Radio Resource Control (RRC) message. The RRC message may be defined based on reusing existing messages or may be formatted by defining a new message (e.g., named "dataset configuration"). Here, when using Abstract Syntax Notation to describe the RRC message information elements, the corresponding configuration information message may be defined, for example, as follows:

[0165]

[0166]

[0167]

[0168]

[0169]

[0170] Here, apparatuses AP1 to APX may receive the above message, process its content according to the training sequence described below, and format the output in the form of input-output pairs. The format of the input-output pairs is described using the information element ReportingConfiguration-IE.

[0171] Therefore, Figure 3 shows an example of a corresponding message exchange implemented between apparatuses and devices (such as DE1 to DEY) in a wireless network according to an embodiment of the present invention, which is used to provide separate bilateral AI / ML-based model training. Here, the apparatus may be exemplarily defined as a user equipment UE holding the first model side of the bilateral model, while the device may be defined as a device (i.e., a network entity), especially a gNB, holding the second model side and thus configured to send the corresponding configuration information to the apparatus (i.e., UE). In another embodiment, the apparatus and the device may also be different network entities. As an example, Figure 3 the shown corresponding message exchange may also be mirrored, i.e., the UE may be exchanged by the network entity, while the network entity (and thus its training process) may be performed by a given UE.

[0172] Therefore, Figure 3The sequence shown can be just one of several training processes carried out by the said device and / or the said equipment, and thus can also result in sharing a large number of different training data sets with the said equipment. As a result, a given device can also be configured to collect different data set settings for training and / or updating a bilateral AI / ML-based model, and can preferably select a given data set based on predefined parameters, such as quality or reliability parameters generated within the system.

[0173] Now, referring to Figure 3 the message exchange shown, the following can be done for each separate bilateral AI / ML-based model training process:

[0174] S301: Initially, the network prepares configuration information for training data set sharing. Thus, the corresponding configuration information can set at least three cell fields for preparing a given configuration information package, which includes the configuration information instructions mentioned above. The three cell fields can include:

[0175] - "Sharing information", including one or more data set identifiers pointing to a set and / or a batch of training samples. Specifically, each data set identifier can be unique and can serve a specific deployment scenario for a given use case. In one example, for example, in the use case of CSI compression model training, two different data set IDs will contain training samples for two different types of codebook implementations, which can also be further different based on the size of the training data (e.g., 16 or 32-bit wide). In addition, the network can also instruct the UE to combine all data sets or subsets from each ID, for example, because some samples from one data set ID may not be compatible with another data set ID.

[0176] - Second, it can include "format information". Here, "format" can describe, for example, a vector, which includes the scale of the data, the quantization information of the data, that is, the bit width or the number of training samples of the data samples used by the UE together with the "sharing information".

[0177] - Third, it can include the so-called "report configuration". The report configuration can allow the network to specify which information is expected for each training sample and when the training data set can be shared. Thus, in the report configuration, at least the above-mentioned configuration information instructions on what, when, and how the data set sharing should be carried out can also be included. As a result, due to the corresponding content of the report configuration, each training sample can be interpreted based on at least one or a combination of the following data (while referring again to Figure 2 ) :

[0178] ■ The input X of the preprocessing of each UE part of the bilateral model;

[0179] ■ The input Y of the UE part of the bilateral model;

[0180] ■ The output Z of the UE part of the bilateral model;

[0181] ■ The output M' of the quantizer considered after the UE part of the bilateral model;

[0182] ■ The output Z' of the hypothesis de - quantizer considered before the hypothesis network part of the bilateral model;

[0183] ■ The output Y' of the hypothesis network part of the bilateral model; and

[0184] ■ The output Z' of the post - processing of the hypothesis network part of the bilateral model.

[0185] Meanwhile, it should also be understood that other data, such as the above - mentioned metric f or additional parameters, can equally be used to define a given training dataset sample.

[0186] S302 and S303: Subsequently, the UE prepares a target dataset based on the information obtained in S301 under the guidance of the network. Specifically, the UE can initially receive a shared information set forwarded by the network via the air interface (S302), and then prepare a given dataset for model training and / or update (S303). For example, the UE can combine data from one or more datasets according to the network's suggestion and combine data with the same format specifier as in the "format information". Thus, at the end of this step, the UE has prepared the dataset to start model training of the bilateral AI / ML - based model.

[0187] S304: After that, the UE can start separate bilateral model training according to at least the Figure 2 process steps described.

[0188] S305: Then, the UE can collect different outputs from the first model side and the assumed second model side training process steps (i.e., for example, X, Y, Z, M', Z', Y', X'), and sort them for each training sample as follows:

[0189] - In one embodiment, each training sample can be constructed based on a combination of one or more of the following data:

[0190] ■ The input X at the pre - processing of the UE part of the bilateral model,

[0191] ■ The input Y of the UE part of the bilateral model,

[0192] ■ The output Z of the UE part of the bilateral model,

[0193] ■ The output M' of the quantizer considered after the UE part of the bilateral model,

[0194] ■ The output Z’ of the hypothesis de-quantizer considered before the hypothesis network part of the bilateral model,

[0195] ■ The output Y’ of the hypothesis network part of the bilateral model,

[0196] ■ The output Z’ of the post-processing of the hypothesis network part of the bilateral model.

[0197] - In addition, each training sample may also include additional metrics that can be derived based on the above data, specifically:

[0198] ■ A metric representing the relationship between the input X used by the UE side part and the corresponding output Y’ of the virtual network side part, such as f(Y, Y’), which can be, for example, cosine similarity (dot product).

[0199] ■ Correspondingly, these additional metrics may also be based on or depend on the UE and / or UE-specific functions, which may also be reported to the network in the dataset or in other ways.

[0200] S306: Thereafter, the UE may upload the output from S305 to the remote server / operator server 4. Here, the output of each network dataset hypothesis can be stored separately (e.g., in the storage unit 2 of the remote server 4) so that the network can distinguish between multiple datasets present in the system. As described above, different identifiers can be used for further differentiation.

[0201] S307 - S309: Subsequently, the UE can indicate to the network that it has prepared the (multiple) datasets, and the network can pick up the dataset output from the remote server 4. For this purpose, the UE can initially indicate any training changes (S307) to any system element such as the network via the air interface, and then prepare for transmission to indicate the start of model training on the network side and the availability of the corresponding training dataset (S308). In addition, in S309, the availability of a new available dataset for model training and / or update can preferably be indicated similarly via the air interface.

[0202] S310 - S314: Thereafter, the network can retrieve from the server a data buffer containing the reference output of the UE and use it to train the second model side of the bilateral AI / ML - based model on the network side. Specifically, at S310, the network can initially forward an indication of the available data set received from the corresponding air interface of the network to an internal network unit (defined herein as "model management"). Subsequently, at S311, the network can download the corresponding training data set, and at S312 and S313, before the actual second model side training process (S313) occurs, the network can first interpret and prepare the data set (S312). Additionally, after the second - side model training (thus completing the entire bilateral model training), the network can similarly send an acknowledgement to the system (e.g., also via the air interface) so that at the end of the efficient training, due to the assistance of the UE in earlier steps, the network and the UE can ensure that the AI / ML - based model is correctly trained.

[0203] S315: Thus, thereafter, bilateral model inference can be implemented via the trained bilateral AI / ML - based model.

[0204] Finally, it should still be noted that, as can be understood and recognized by those skilled in the art, although in the above - mentioned exemplary embodiments (with reference to the accompanying drawings), the messages communicated / exchanged between network components / elements may seem to have specific / definite names, according to various embodiments (e.g., the emphasized technology), these messages may have different names and / or communicate / exchange in different forms / formats.

[0205] According to some exemplary embodiments, a corresponding method suitable for being executed by the above - mentioned devices (network elements / components) such as the UE, etc. is also provided.

[0206] Furthermore, it should be noted that the above - mentioned device (or equipment) features correspond to the respective method features, but for the sake of brevity, these method features may not be explicitly described. The disclosure of this document is considered to also extend to such method features. In particular, the present disclosure is understood to relate to a method of operating the above - mentioned devices, and / or a method of providing and / or arranging the corresponding elements of these devices.

[0207] In addition, according to some further exemplary embodiments, a corresponding device (e.g., implementing the UE, network, etc. as described above) is also provided, which includes at least one processing circuit and at least one memory for storing instructions executed by the processing circuit, wherein the at least one memory and the instructions are configured to use the at least one processing circuit to cause the corresponding device to at least execute the corresponding steps as described above.

[0208] In still other exemplary embodiments, a corresponding apparatus (e.g., implementing a UE, network, etc. as described above) is provided that includes corresponding means configured to perform at least the corresponding steps described above.

[0209] It should be noted that the examples of the embodiments of the present disclosure are applicable to various different network configurations. In other words, the examples shown in the above figures, which are the basis of the above examples, are merely illustrative and do not limit the present disclosure in any way. That is, based on the defined principles, other existing and proposed new functions available in the corresponding operating environment can be used in combination with the examples of the embodiments of the present disclosure.

[0210] It should also be noted that the disclosed exemplary embodiments can be implemented in various ways using hardware and / or software configurations. For example, the disclosed embodiments can be implemented using dedicated hardware and / or hardware associated with software executable thereon. The components and / or elements in the figures are merely examples and do not limit the scope of use or functionality of any hardware, software combined with hardware, firmware, embedded logic components, or combinations of two or more such components implementing a particular embodiment of the present disclosure.

[0211] It should also be noted that the specification and the figures only illustrate the principles of the present disclosure. Although not explicitly described or shown herein, those skilled in the art will be able to implement various arrangements that embody the principles of the present disclosure and are included within its spirit and scope. In addition, all examples and embodiments outlined in the present disclosure are mainly for explanatory purposes to help the reader understand the principles of the proposed method. Moreover, all statements of the principles, aspects, and embodiments of the present disclosure provided herein, along with their specific examples, are intended to cover their equivalents.

[0212] List of Abbreviations

[0213] 5G - Fifth Generation

[0214] gNB - 5G / NR Base Station

[0215] NR - New Radio

[0216] Al - Artificial Intelligence

[0217] ML - Machine Learning

[0218] UE - User Equipment

[0219] DCI - Downlink Control Information

[0220] MAC CE - Medium Access Control Control Element

[0221] MIMO - Multiple Input Multiple Output

[0222] NN - Neural Network

[0223] CSI - Channel State Information

[0224] RS - Reference Signal

[0225] FPGA - Field Programmable Gate Array

[0226] GPU - Graphics Processing Unit

[0227] ASIC - Application Specific Integrated Circuit

[0228] BS - Base Station

[0229] AP - Access Point

[0230] NB - Node B

[0231] DU - Distribution Unit

[0232] CU - Centralized / Central Unit

[0233] SINR - Signal to Interference plus Noise Ratio

[0234] NMSE - Normalized Mean Square Error

Claims

1. An apparatus (AP1-APX) in a wireless network, comprising: One or more processors; And A memory storing instructions that, when executed by the one or more processors, cause the apparatus to: (S302) Receive information for configuring the apparatus (AP1-APX) to share data that will be used to train a bilateral model, the bilateral model facilitating data communication between the apparatus (AP1-APX) and another device (DE1-DEY) in the wireless network, the information indicating which data will be shared with the other device (DE1-DEY); (S304) Based on training data available at the apparatus (AP1-APX), train at least a first side of the bilateral model by a machine learning algorithm; Determine a dataset to be shared from the training data according to the received configuration information; (S306) Send the determined dataset for the other device (DE1-DEY) to separately train at least a second side of the bilateral model from the first site.

2. The device (AP1-APX) according to claim 1, wherein, The apparatus (AP1-APX) is further caused to use the first side of the bilateral model when communicating with the other device (DE1-DEY), and the configuration information indicates whether the input data and / or output data of the first side of the bilateral model should be shared with the other device (DE1-DEY) during training.

3. The apparatus (AP1-APX) according to claim 2, wherein, The apparatus (AP1-APX) determines a hypothesized second side (HDE1) of the bilateral model during training, and the configuration information indicates whether some data for training the hypothesized second side will be shared.

4. The apparatus (AP1-APX) according to claim 3, wherein, The configuration information indicates whether a metric (f) regarding the relationship between the first side training data and the training data of the hypothesized second side (HDE1) will be shared.

5. The device (AP1-APX) according to any one of the preceding claims, wherein, The configuration information indicates a pre-configured dataset to be shared.

6. The device (AP1 - APX) according to any one of the preceding claims, wherein, The configuration information indicates when to share the data with the other device (DE1-DEY), and indicates at least one of a time instance, a sharing condition, or a sharing trigger.

7. The apparatus (AP1-APX) according to any one of the preceding claims, wherein, The configuration information indicates how to share the data with the other device (DE1-DEY), and indicates at least one of a format for sharing the data, a communication medium for sharing the data, or a priority of the shared data.

8. The apparatus (AP1-APX) according to any one of the preceding claims, wherein, The configuration information and / or the shared dataset are transmitted via an air interface by signaling (S101A; S101B; S101X), or transmitted via a remote server (4) between the apparatus (AP1-APX) and the other device (DE1-DEY) (S102; S103).

9. The apparatus (AP1-APX) according to any one of the preceding claims, wherein, The configuration information relates to an initial training of the bilateral model, or relates to an update of the bilateral model based on new training data.

10. The apparatus (AP1 - APX) according to any one of the preceding claims, wherein, The configuration information includes at least one of the following: sharing information, which includes one or more dataset identifiers pointing to a set / batch of datasets to be shared with the other device (DE1-DEY); Format information that describes a vector including the scale of the data set to be shared; or a reporting configuration that specifies data to be used for at least a first side of the bilateral model.

11. The apparatus (AP1-APX) according to any one of the preceding claims, wherein, The bilateral model facilitates compression / decompression of channel state information CSI. A first side of the bilateral model relates to compression of information about a wireless channel, and a second side of the bilateral model relates to decompression of the wireless channel information.

12. The apparatus (AP1-APX) according to any one of claims 1 to 11, wherein, The bilateral model facilitates channel encoding / decoding. A first side of the bilateral model relates to encoding of data to be transmitted to the other device (DE1 - DEY), and a second side of the bilateral model relates to decoding of data received by the other device (DE1 - DEY).

13. The device (AP1-APX) according to any one of the preceding claims, wherein, The device (AP1 - APX) is a user equipment UE in the wireless network.

14. A device (DE1 - DEY) in a wireless network, comprising: One or more processors; And A memory storing instructions that, when executed by the one or more processors, cause the device (DE1 - DEY) to: (S301; S302) Send information for configuring a device (AP1 - APX) in the wireless network to share data for training a bilateral model, the bilateral model facilitating data communication between the device (AP1 - APX) and the device (DE1 - DEY) in the wireless network, the information indicating which data the device (AP1 - APX) will share with the device (DE1 - DEY); (S311) Receive the shared data from the device (AP1 - APX); (S313) Based on the received data, train at least a second side of the bilateral model by a machine learning algorithm.

15. The device (DE1-DEY) according to claim 14, wherein, The device (DE1 - DEY) is a next generation radio access network NG - RAN node in the wireless network, such as a gNB, and the device (AP1 - APX) is a user equipment UE in the wireless network.

16. The device (DE1-DEY) according to claim 15, wherein, The device (DE1 - DEY) trains and maintains UE - specific models, each model being applied to communication with a single UE.

17. The device (DE1-DEY) according to claim 15, wherein The device (DE1 - DEY) receives shared data from multiple UEs and trains at least a second side of the bilateral model based on a set of shared data received from different UEs.

18. The device (DE1-DEY) according to claim 14, wherein, Training of the second side of the bilateral model is at least based on a reference output of a hypothesized second side (HDE1) determined during training of the bilateral model by the device (AP1 - APX), the reference output being at least part of the shared data of the device (AP1 - APX).

19. A method for training a bilateral model, the bilateral model facilitating data communication between a device (AP1 - APX) and another device (DE1 - DEY) in a wireless network, the method comprising: (S302) Receive, by the device (AP1 - APX), information for configuring the device (AP1 - APX) to share data to be used for training the bilateral model, the information indicating which data will be shared with the other device (DE1 - DEY); (S304) Based on training data available at the device (AP1-APX), train at least the first side of the bilateral model separately from the second side of the bilateral model by a machine learning algorithm; Determine a data set to be shared from the training data according to the received configuration information; (S306) Send the determined data set for use by the other device (DE1-DEY) to train at least the second side of the bilateral model separately from the first side.

20. A method for training a bilateral model, the bilateral model facilitating data communication between a device (AP1-APX) and another device (DE1-DEY) in a wireless network, the method comprising: (S301; S302) Send information for configuring the device (AP1-APX) in the wireless network to share data to be used for training the bilateral model, the information indicating which data the device (AP1-APX) will share with the device (DE1-DEY); (S311) Receive the shared data from the device (AP1-APX); (S313) Based on the received data, train at least the second side of the bilateral model by a machine learning algorithm.

21. The method according to claim 19 or 20, wherein, The configuration information indicates when to share the data with the other device (DE1-DEY) and / or how to share the data with the other device (DE1-DEY).

22. A computer program comprising instructions for causing a device (AP1-APX) or a device (DE1-DEY) to perform the method according to any one of claims 19 to 21.

23. A memory storing computer-readable instructions for causing a device (AP1-APX) or a device (DE1-DEY) to perform the method according to any one of claims 19 to 21.

Images