Semi-supervised learning with data enhancement

Abstract

Example embodiments of the present disclosure provide a signaling framework for semi-supervised learning (SSL) with data enhancement. In an example method, a first device receives configuration information related to data enhancement-based SSL training from a second device. The first device trains an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unmarked data at the first device, and transmits at least one model parameter of the trained AI / ML model to the second device. In this manner, a distributed arrangement for training on different entities involving individual pipes using data enhancement in a network is presented, where the performance of the AI / ML model may be improved.

Description

Technical Field

[0001] The exemplary embodiments of this disclosure generally relate to the field of communications, and particularly to devices, methods, apparatuses, and computer-readable media for semi-supervised learning (SSL) with data augmentation, such as federated SSL with data augmentation for enhanced AI / ML-based UE positioning. Background Technology

[0002] 3GPP Radio Access Network (RAN) 1 has initiated a research project on AI / ML for the air interface, and one of the key use cases being considered is leveraging AI / ML to enhance positioning accuracy. There is a consensus that the availability of baseline truth labels should be evaluated to assess the performance impact of AI / ML positioning cases.

[0003] However, in practical deployments, benchmark ground truth labels are difficult or even impossible to obtain. Therefore, it is a necessary and timely task to study the standardization impact of AI / ML localization methods based on semi-supervised learning (SSL) and the corresponding signaling. Summary of the Invention

[0004] In general, the exemplary embodiments of this disclosure provide a solution for semi-supervised learning (SSL) with data augmentation, such as federated SSL with data augmentation for enhanced AI / ML-based UE localization.

[0005] In a first aspect, a first device is provided. The first device includes at least one processor and at least one memory storing instructions, which, when executed by the at least one processor, cause the first device to at least: receive configuration information from a second device related to data-augmented semi-supervised learning (SSL) training; train an artificial intelligence / machine learning (AI / ML) model deployed on the first device based on the received configuration information and unlabeled data on the first device; and send at least one model parameter of the trained AI / ML model to the second device.

[0006] In a second aspect, a second device is provided. The second device includes at least one processor and at least one memory storing instructions, which, when executed by the at least one processor, cause the second device to at least: send first configuration information to a first device related to data-augmented semi-supervised learning (SSL) training; receive from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; train a second AI / ML model deployed at the second device based on second configuration information related to data-augmented supervised learning (SL) training and labeled data at the second device; and aggregate at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0007] In a third aspect, a method is provided. The method includes: receiving, at a first device, configuration information related to data-augmented semi-supervised learning (SSL) training from a second device; training an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unlabeled data at the first device; and sending at least one model parameter of the trained AI / ML model to the second device.

[0008] In a fourth aspect, a method is provided. The method includes: at a second device, sending first configuration information related to data-augmented semi-supervised learning (SSL) training to a first device; receiving from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; training a second AI / ML model deployed at the second device based on second configuration information related to data-augmented supervised learning (SL) training and labeled data at the second device; and aggregating at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0009] In a fifth aspect, an apparatus is provided. The apparatus includes: components for receiving configuration information related to data-augmented semi-supervised learning (SSL) training from a second device at a first device; components for training an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unlabeled data at the first device; and components for transmitting at least one model parameter of the trained AI / ML model to the second device.

[0010] In a sixth aspect, an apparatus is provided. The apparatus includes: components for transmitting first configuration information from a second device to a first device in relation to data-augmented semi-supervised learning (SSL) training; components for receiving from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the transmitted first configuration information and unlabeled data at the first device; components for training a second AI / ML model deployed at the second device based on second configuration information in relation to data-augmented supervised learning (SL) training and labeled data at the second device; and components for aggregating at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0011] In a seventh aspect, a non-transient computer-readable storage medium is provided, comprising program instructions. When executed by a device, the program instructions cause the device to perform at least the following: receive configuration information from a second device related to data-augmented semi-supervised learning (SSL) training; train an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unlabeled data at the first device; and send at least one model parameter of the trained AI / ML model to the second device.

[0012] In an eighth aspect, a non-transient computer-readable storage medium is provided, comprising program instructions. When executed by a device, the program instructions cause the device to perform at least the following: sending first configuration information to a first device related to data-augmented semi-supervised learning (SSL) training; receiving from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; training a second AI / ML model deployed at the second device based on second configuration information related to data-augmented supervised learning (SL) training and labeled data at the second device; and aggregating at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0013] In a ninth aspect, a computer program including instructions is provided that, when executed by a device, causes the device to at least: receive configuration information from a second device related to data-augmented semi-supervised learning (SSL) training; train an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unlabeled data at the first device; and send at least one model parameter of the trained AI / ML model to the second device.

[0014] In a tenth aspect, a computer program including instructions is provided that, when executed by a device, causes the device to at least: send first configuration information to a first device in relation to data-augmented semi-supervised learning (SSL) training; receive from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; train a second AI / ML model deployed at the second device based on second configuration information in relation to data-augmented supervised learning (SL) training and labeled data at the second device; and aggregate at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0015] In an eleventh aspect, a first device is provided. The first device includes: a receiving circuit system configured to receive, at the first device, configuration information related to data-augmented semi-supervised learning (SSL) training from a second device; a training circuit system configured to train an artificial intelligence / machine learning (AI / ML) model deployed at the first device based on the received configuration information and unlabeled data at the first device; and a transmitting circuit system configured to transmit at least one model parameter of the trained AI / ML model to the second device.

[0016] In a twelfth aspect, a second device is provided. The second device includes: a transmitting circuit system configured to transmit, from the second device to a first device, first configuration information related to data-augmented semi-supervised learning (SSL) training; a receiving circuit system configured to receive from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the transmitted first configuration information and unlabeled data at the first device; a training circuit system configured to train a second AI / ML model deployed at the second device based on second configuration information related to data-augmented supervised learning (SL) training and labeled data at the second device; and an aggregation circuit system configured to aggregate at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0017] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0018] Some exemplary embodiments will now be described with reference to the accompanying drawings, in which:

[0019] Figure 1 An example communication network in which embodiments of the present disclosure can be implemented is illustrated;

[0020] Figure 2 Examples of process flows according to some exemplary embodiments of this disclosure are illustrated;

[0021] Figure 3 The illustration shows an example framework for collaborative semi-supervised learning (SSL) for AI / ML-based localization according to some example embodiments of the present disclosure;

[0022] Figure 4 The illustration shows an example process for collaborative / federated semi-supervised learning (SSL) training according to some example embodiments of the present disclosure;

[0023] Figure 5 Another example of a process flow according to some exemplary embodiments of the present disclosure is illustrated;

[0024] Figure 6 The diagram illustrates an overview of the neural network model used for training.

[0025] Figure 7 The diagram illustrates the evaluation results between different options;

[0026] Figure 8 The illustration shows a flowchart of an example method implemented at a first device according to some other embodiments of the present disclosure;

[0027] Figure 9 The illustration shows a flowchart of an example method 900 implemented at a first device according to some other embodiments of the present disclosure;

[0028] Figure 10 The illustration shows a simplified block diagram of a device suitable for implementing some example embodiments of the present disclosure; and

[0029] Figure 11 A block diagram illustrating an example of a computer-readable medium according to some exemplary embodiments of the present disclosure is shown.

[0030] In all the accompanying drawings, the same or similar reference numerals denote the same or similar elements. Detailed Implementation

[0031] The principles of this disclosure will now be described with reference to some exemplary embodiments. It should be understood that these embodiments are described merely for illustrative purposes and to assist those skilled in the art in understanding and implementing this disclosure, and are not intended to limit the scope of this disclosure in any way. This disclosure described herein can be implemented in various ways different from those described below.

[0032] In the following description and claims, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0033] References to "an embodiment," "embodiment," "example embodiment," etc., in this disclosure indicate that the described embodiment may include a particular feature, structure, or characteristic, but not every embodiment must include that particular feature, structure, or characteristic. Furthermore, such phrases do not necessarily refer to the same embodiment. Additionally, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is within the knowledge of those skilled in the art to consider how it may affect such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not).

[0034] It should be understood that while the terms “first” and “second” may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.

[0035] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. As used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It will also be understood that the terms “comprises,” “comprising,” “has,” “having,” “includes,” and / or “including” as used herein specify the presence of the stated features, elements, and / or components, but do not exclude the presence or addition of one or more other features, elements, components, and / or combinations thereof. As used herein, “at least one of the following: ” and “at least one of ” and similar wording (where the list of two or more elements is connected by “and” or “or”) means at least any one of these elements, or at least any two or more of these elements, or at least all of these elements.

[0036] As used in this application, the term "circuit system" may refer to one or more or all of the following: (a) Hardware circuit implementation only (such as implementation only in analog and / or digital circuit systems) and (b) A combination of hardware circuitry and software, such as (if applicable): (i) A combination of (multiple) analog and / or digital hardware circuits with software / firmware; and (ii) Any part of a hardware processor(s) having software (including digital signal processor(s)), software, and memory(s) working together to enable a device such as a mobile phone or server to perform various functions; and (c) (multiple) hardware circuits and / or (multiple) processors, such as (multiple) microprocessors or a portion thereof, which require software (e.g., firmware) to operate, but may be absent when operation is not required.

[0037] The definition of "circuit system" applies to all uses of the term in this application (including in any claim). As another example, as used herein, the term "circuit system" also covers implementations of hardware circuitry or processors (or processors in general) and a portion thereof, along with their accompanying software and / or firmware. The term "circuit system" also covers (e.g., and if applicable to a particular claim element) baseband integrated circuits or processor integrated circuits for mobile devices or similar integrated circuits in servers, cellular network devices, or other computing or networking devices.

[0038] As used herein, the terms “network,” “communication network,” or “data network” refer to a network that conforms to any suitable communication standard, such as Long Term Evolution (LTE), LTE-A, Wideband Code Division Multiple Access (WCDMA), High-Speed ​​Packet Access (HSPA), Narrowband Internet of Things (NB-IoT), Wi-Fi, etc. Furthermore, communication between terminal devices and network devices / elements in a communication network can be performed according to any suitable generation of communication protocol, including but not limited to fourth-generation (4G), 4.5G, future fifth-generation (5G), the IEEE 802.11 communication protocol, and / or any other currently known or future-developed protocols. Embodiments of this disclosure can be applied to various communication systems. Given the rapid development of communications, there will naturally be communication technologies and systems that embody future types of this disclosure. The scope of this disclosure should not be limited to the systems described above.

[0039] As used herein, the term "network device" refers to a node in a communications network through which terminal devices receive services (e.g., location services). Network devices can refer to core network equipment or access network equipment, such as base stations (BS) or access points (APs) or transmit and receive points (TRPs), for example, Node B (NodeB or NB), evolved Node B (eNodeB or eNB), NR NB (also known as gNB), remote radio unit (RRU), radio headend (RH), remote radio headend (RRH), WiFi device, relay, low-power nodes such as femtoseconds, picoseconds, etc., depending on the terminology and technology applied. In the following description, the terms "network device," "AP device," "AP," and "access point" may be used interchangeably.

[0040] The term "terminal device" refers to any terminal device capable of wireless communication. As an example and not a limitation, a terminal device may also be referred to as a communication device, user equipment (UE), subscriber station (SS), portable subscriber station, mobile station (MS), station (STA), or station equipment, or access terminal (AT). Terminal devices may include, but are not limited to, mobile phones, cellular phones, smartphones, Voice over IP (VoIP) phones, wireless local loop phones, tablets, wearable terminal devices, personal digital assistants (PDAs), portable computers, desktop computers, image capture terminal devices such as digital cameras, gaming terminal devices, music storage and playback devices, in-vehicle wireless terminal devices, wireless endpoints, mobile stations, laptop embedded devices (LEE), laptop mounted devices (LME), USB dongles, smart devices, wireless customer premises equipment (CPE), Internet of Things (IoT) devices, watches or other wearable devices, head-mounted displays (HMDs), vehicles, drones, medical devices and applications (e.g., remote surgery), industrial devices and applications (e.g., robots and / or other wireless devices operating in industrial and / or automated processing chain environments), consumer electronics devices, devices operating on commercial and / or industrial wireless networks, etc. In the following description, the terms “station”, “station equipment”, “STA”, “terminal equipment”, “communication equipment”, “terminal”, “user equipment” and “UE” may be used interchangeably.

[0041] As used herein, the term "Location Management Function" (LMF) refers to an entity defined in a network (e.g., a 5G network) that provides location functionality by determining the geographic location of an end device through downlink and uplink location radio measurements. The LMF can be provided in network equipment in the core or access network, or in the end device moving within the communication environment. The LMF can receive measurement and assistance information from the Next Generation Radio Access Network (NG-RAN) and the end device via, for example, Access and Mobility Management Functions (AMFs) to calculate the location of the end device.

[0042] As used herein, the term "Location Reference Unit" (PRU) refers to an entity that provides accurate location information as a reference truth to other entities in a communication network, such as terminal devices and LMFs. A PRU can be geographically fixed or mobile in the environment.

[0043] 3GPP RAN1 has initiated a research project on AI / ML for the air interface, with one of the key use cases being considered being AI / ML-enhanced positioning accuracy. There is a consensus that the availability of ground truth labels should be evaluated to assess the performance impact of AI / ML positioning cases. In practical deployments, obtaining ground truth labels is costly or even impossible. Therefore, researching the standardization impact of semi-supervised learning-based AI / ML positioning methods and the corresponding signaling is a necessary and timely task.

[0044] Supervised learning, self-supervised learning, and semi-supervised learning for AI / ML models are researched and applied in various fields. When labeled data is plentiful, fitting a model to the dataset is easier. However, in many cases, access to labeled data is limited—meaning that learning a well-performing model is more difficult and carries a high risk of overfitting a limited amount of data. Therefore, in such situations, effectively utilizing unlabeled data is crucial.

[0045] A common characteristic of semi-supervised learning is that the dataset consists of a small set of labeled data and a large amount of unlabeled data. The expectation is that using semi-supervised learning in scenarios with equally small labeled datasets will yield gains compared to using supervised learning with small labeled datasets.

[0046] UE location can be estimated using AI / ML methods. Typically, for direct positioning, supervised learning (SL)-based methods can be applied to infer UE location based on pre-selected inputs (e.g., wireless measurements). Similarly, for assisted positioning, AI / ML with supervised learning methods can be used to estimate intermediate features (e.g., line-of-sight (LOS) / non-line-of-sight (NLOS) markers), and then these intermediate features can be used to estimate UE location.

[0047] A UE may have abundant positioning-related measurements but may lack location estimates (due to the unavailability or shutdown of the Global Navigation Satellite System (GNSS)). This means that most UEs have a large amount of unlabeled data. On the other hand, a relatively small number of target UEs also have their location estimates. Furthermore, the LMF may know the location of a small number of UEs, and these UEs may forward some raw measurement samples to the LMF. This means that a relatively small number of target UEs may have labeled data, and the LMF itself may have a relatively small amount of labeled data. Therefore, one possible scenario is that most UEs will have a large amount of unlabeled data, while the LMF and a small number of UEs may have labeled data.

[0048] In order to perform inference using a supervised learning model, a training phase is necessary. This involves estimating the model parameters using a set of labeled (benchmark) data (the input and the model's corresponding expected output). The performance of the trained model can be highly dependent on the labeled data used / available.

[0049] In practice, collected labeled data (e.g., wireless measurements using UE location as a label) can be costly, and there is likely to be more unlabeled data available (e.g., wireless measurements without UE location). In this regard, semi-supervised learning (SSL) is a good choice. Specifically, for the training phase, unlabeled data can be utilized in addition to labeled data to enhance model training performance. Therefore, by utilizing unlabeled data, the overhead of data collection can be significantly reduced. In particular, to enable semi-supervised learning, different entities with their capabilities (PRU and regular UE) can improve the data collection phase. PRUs can be used to provide labeled data, while regular UEs collect unlabeled data samples by default.

[0050] Semi-supervised learning leverages the advantages of unlabeled data in addition to labeled data to further enhance the performance of trained models. However, the way in which such semi-supervised methods are enabled for ML-based localization needs to be explicitly defined to ensure the desired performance.

[0051] Furthermore, model training is typically performed at a single entity (such as LMF or UE). However, when using large training datasets, the training process can be cumbersome, and collaboration between different entities can help optimize the training process (in terms of resources used and latency).

[0052] Embodiments of this disclosure provide a framework for enabling collaboration among different entities to train robust AI / ML models for UE localization. Embodiments of this disclosure provide a distribution for training across different entities in a network, involving separate pipelines for labeled and unlabeled data and describing support for data augmentation. Each network entity utilizes data augmentation and processes a local training process using either labeled or unlabeled data; the local training outputs are then aggregated at a central node to combine their results. A detailed description of a signaling framework tailored for federated SSL operations and a description of training a federated / cooperative SSL model using labeled and unlabeled datasets are provided below.

[0053] For illustrative purposes, please refer to the following: Figures 1 to 11 The principles and exemplary embodiments of this disclosure are described. However, it should be noted that these embodiments are given to enable those skilled in the art to understand the inventive concept of this disclosure and to implement the solutions presented herein, and are not intended to limit the scope of this application in any way.

[0054] Figure 1 The illustration shows an example of an application scenario 100 in which some exemplary embodiments of the present disclosure can be implemented. Application scenario 100 is part of a communication network, including terminal devices and network devices.

[0055] In the description of the exemplary embodiments of this disclosure, network environment 100 may also be referred to as communication system 100 (e.g., part of a communication network). For illustrative purposes only, the aspects of the exemplary embodiments will be described in the context of one or more terminal devices and network devices communicating with each other. However, it should be understood that the description herein can be applied to other types of devices or other similar devices referred to using other terms.

[0056] The first device 110 can receive services (e.g., location services) from the second device 120, and the first device 110 and the second device 120 can communicate data and control information with each other via network 102. In some embodiments, the first device 110 and the second device 120 can communicate using a direct link / channel. The first device 110 can be a mobile terminal device or UE in network environment 100. The second device can be a network device located in the access network or core network. In some embodiments, the second device can be or includes an LMF (Local Management Function).

[0057] In communication system 100, the link from second device 120 to first device 110 is called the downlink (DL), and the link from first device 110 to second device 120 is called the uplink (UL). In the downlink, second device 120 is a transmitting (TX) device (or transmitter), and first device 110 is a receiving (RX) device (or receiver). In the uplink, first device 110 is a transmitting (TX) device (or transmitter), and second device 120 is an RX device (or receiver).

[0058] Network 102 can be implemented according to any suitable wireless or wired communication protocol, including but not limited to cellular communication protocols and core network communication protocols such as fourth-generation (4G) and fifth-generation (5G), wireless local area network communication protocols such as IEEE 802.11, and / or any other currently known or to be developed in the future. Furthermore, communication can utilize any suitable wireless communication technology, including but not limited to: Code Division Multiple Access (CDMA), Frequency Division Multiple Access (FDMA), Time Division Multiple Access (TDMA), Frequency Division Duplex (FDD), Time Division Duplex (TDD), Multiple-Input Multiple-Output (MIMO), Orthogonal Frequency Division Multiplexing (OFDM), Discrete Fourier Transform Extended OFDM (DFT-s-OFDM), and / or any other currently known or to be developed in the future.

[0059] It should be understood that, Figure 1 The number of devices shown, as well as their connections and types, are for illustrative purposes and do not imply any limitation. Communication system 100 may include any suitable number of devices appropriate for implementing embodiments of this disclosure.

[0060] Figure 2 An example of a process flow 200 according to some exemplary embodiments of the present disclosure is illustrated. For ease of understanding, reference will be made to... Figure 1 To describe process flow 200. It should be understood that, although referenced... Figure 1 Application scenario 100 describes process flow 200, but process flow 200 can also be applied to other similar communication scenarios.

[0061] The second device 120 sends (201) first configuration information 202 related to data augmentation-based semi-supervised learning (SSL) training. In some embodiments, the first configuration information may be determined based on the capabilities of the first device 110. In some embodiments, the first configuration information may include at least one of the following: a data augmentation function for data augmentation; at least one parameter related to SSL training; at least one initial model parameter for an artificial intelligence / machine learning (AI / ML) model; or at least one convex combination weight.

[0062] Alternatively or additionally, the second device 120 may have second configuration information related to data augmentation-based supervised learning (SL) training, wherein the second configuration may include at least one of the following: the same data augmentation function for data augmentation; at least one parameter related to SL training; the same initial model parameters for the AI / ML model; or at least one convex combination weight.

[0063] First device 110 receives (203) first configuration information 202 related to data-augmented semi-supervised learning (SSL) training and performs local training. First device 110 trains (204) a first AI / ML model deployed on first device 110 based on the received first configuration information 202 and unlabeled data at first device 110. Then first device 110 sends (205) at least one model parameter of the trained first AI / ML model 206 to second device 120.

[0064] The second device 120 receives (207) at least one model parameter of the trained first AI / ML model 206. The second device 120 trains (208) a second AI / ML model deployed at the second device 120 based on second configuration information related to data augmentation-based supervised learning (SL) training and labeled data at the second device 120.

[0065] The second device 120 aggregates (209) at least one model parameter of the trained first AI / ML model 206 and at least one model parameter of the trained second AI / ML model. In some embodiments, the second device 120 may aggregate at least one model parameter of the first AI / ML model 206 and at least one model parameter of the second AI / ML model based on model parameter weights to obtain at least one aggregated model parameter. In some embodiments, the model parameter weights may include a first model parameter weight for at least one model parameter of the first AI / ML model and a second model parameter weight for at least one model parameter of the second AI / ML model, wherein the first model parameter weight and the second model parameter weight are non-negative and sum to one.

[0066] In some embodiments, the first device 110 may receive at least one aggregated model parameter from the second device 120 and may obtain an updated first AI / ML model having at least one aggregated model parameter. In response to satisfying an iteration condition, the first device 110 may train an updated first AI / ML model based on unlabeled data. Furthermore, in some embodiments, the second device 120 may obtain an updated second AI / ML model having at least one aggregated model parameter. In response to satisfying an iteration condition, the second device 120 may train an updated second AI / ML model based on labeled data.

[0067] In view of the foregoing, embodiments of this disclosure provide a signaling framework enabling collaborative semi-supervised learning (SSL) for AI / ML-based localization. This signaling framework enables distributed training across different entities in the network, involving separate pipelines for labeled and unlabeled data and supporting data augmentation. The local training outputs are then aggregated at a central node to combine their results and provide a better trained model.

[0068] Figure 3 An example framework 300 for collaborative semi-supervised learning (SSL) for AI / ML-based localization, according to some example embodiments of this disclosure, is illustrated. Figure 3 The document describes the distribution for training on different entities within the network, involving separate pipelines for labeled and unlabeled data and supporting data augmentation. Each network entity utilizes data augmentation and processes its local training process using either labeled or unlabeled data; the local training outputs are then aggregated at a central node to combine their results. Framework 300 can be implemented through interoperability between the UE and the LMF, which can be respectively... Figure 1 and Figure 2 Examples of the first device 110 and the second device 120.

[0069] like Figure 3 As shown, in the pipeline with unlabeled data, local training 313 can be implemented by the UE using unlabeled data 321 and a data augmentation configuration shared by the LMF, where the augmentation configuration is based on its UE capabilities. In the pipeline with unlabeled data, local training 323 can be implemented by the gNB / UE using labeled data 311 and a data augmentation configuration shared by the LMF. Based on the augmentation configuration, data augmentations 312 and 322 can be performed on both labeled data 311 and unlabeled data 321 for better training. Local training 313 and 323 are then combined at the LMF to form aggregate training 330.

[0070] Figure 4 An example process 400 for collaborative / federated semi-supervised learning (SSL) training according to some example embodiments of this disclosure is illustrated. Figure 4 The text describes each local pipeline and the aggregation performed at a central entity (such as the LMF), where local training is implemented at different entities / locations in the network using either labeled or unlabeled datasets. Process 400 can be implemented through interoperability between the UE and the LMF, which can be respectively... Figure 1 and Figure 2 Examples of the first device 110 and the second device 120.

[0071] like Figure 4 As shown, in the pipeline with unlabeled data, the unlabeled data 321 may include a collected labeled dataset (x, y), where x, for example, refers to a location reference signal measurement, such as channel impulse response (CIR) or reference signal received power (RSRP), and This could refer to corresponding label measurements, such as geographic location. The selected data augmentation implementation is then applied to the data. Data augmentation implementations can be selected from, but are not limited to, the following options.

[0072] One option is phase rotation. For example, each element of the original Channel State Information (CSI) vector is rotated by the same phase. The chosen phase is small in amplitude and... The angle is randomly selected from the range where θ is a small angle.

[0073] Another option is amplitude amplification or amplitude reduction. For example, the amplitude of each element of the original CSI vector is changed by the same amount. This amount is randomly chosen between [-α, α], where α is a small positive number and represents a fraction of the average amplitude of all elements of the original CSI vector.

[0074] Another option is Gaussian noise. For example, a Gaussian noise vector can be added to the original CSI vector, and the power of this noise vector is a fraction of the average power of the original CSI vector.

[0075] Another option is filtering / smoothing. For example, if the CSI is in the frequency domain, the amplitude across subcarriers can be smoothed using a low-pass filter.

[0076] Another option is neural network-based augmentation. This might involve using a separate neural network trained to find appropriate augmentations for the input CSI vector. The above options are exemplary; other options may also be used.

[0077] LMF can provide the configuration for the selected data augmentation implementation (e.g., offset angle values ​​for phase rotation implementation). Figure 3 In, function g This refers to data augmentation functions. This refers to the newly enhanced data, while It is a neural network function (e.g., or The neural network (NN) model (also simply called a neural network) will be described as an example of an AI / ML model. This local training is performed to minimize the error between the predicted locations for both the measurement samples and the augmented samples and the ground truth locations.

[0078] like Figure 4As shown, in the pipeline with untagged data, the tagged data 311 may include collected PRS wireless measurements, such as CSI and RSRP. Figure 4 In the context of data augmentation functions g This was also applied to this data. Subsequently, local training was performed to estimate the local neural network. The parameters. For this purpose, the entity feeds in an unlabeled input x to obtain... It also uses data augmentation functions. g Send the input to obtain g(x), and then feed g(x) into the neural network. Get Loss of 403 was taken as The mean square error (MSE) between ȳ and (i.e., as Figure 4 The MSE loss shown ).

[0079] Subsequently, the two locally trained neural network functions are shared at the central entity and aggregated to minimize the total loss 404 and obtain a robust ML model. Data augmentation is used for both labeled and unlabeled data to make the ML model robust to generalization errors.

[0080] In the case where supervised learning (SL) training occurs at the LMF and (SSL) training occurs at the UE, the LMF uses its local neural network (NN) in each iteration. (with weight) To calculate the loss 401 between the labels on a batch of labeled training samples and the predicted output (i.e., as shown in the original text) Figure 4 The y-coordinate of SL training at the LMF shown in the figure MSE loss between L 1 ), and the loss 402 between the same label and the predicted output for the same batch of input training samples that have undergone data transformation (i.e., as Figure 4 The MSE loss between y and ȳ during SL training at the LMF shown is illustrated. L 2 Similarly, the UE uses its local NN. (with weight) ) to calculate the loss 403 between the predicted output of a batch of unlabeled samples and the predicted output of the same batch of unlabeled samples that has undergone data transformation (i.e., as Figure 4 SSL training at the UE shown MSE loss between ȳ L 3 ).

[0081] The UE will suffer this loss L3 Send to LMF, LMF aggregates all losses (using by) Given possible convex combinations, where the combination weights are... It is non-negative, among which And execute Local model updates (e.g., through gradient descent or any other optimization technique). Similarly, LMF also sends the loss to the UE. Then UE aggregates all losses (For example, by using the same weight combination) to perform its local NN The model is updated. Therefore, in each iteration, the LMF and UE exchange their respective local losses and perform local model updates separately. With perfect training, the two local models are expected to... They are equal after each iteration because they are updated using the same aggregate loss and optimizer.

[0082] LMF calculates the loss as described in the previous case. And execute in multiple iterations The local model is updated. Similarly, the UE also calculates the loss. And it updates its local NN locally in multiple iterations. After a predefined number of iterations, the LMF and UE transfer their respective local NN weights. and They send to each other. Then each entity updates its local NN weights using a convex combination of predefined individual weights, given as... ,in as well as It is non-negative and sums to one. With perfect training, two local models are expected. They are synchronized after multiple iterations because they are updated using the same weight convex combination. Alternatively or additionally, the LMF does not send its local NN weights to the UE, and updates are performed only at the LMF based on the received local weights calculated by the UE (or possibly multiple UEs).

[0083] Figure 5 Another example of a process flow 500 according to some exemplary embodiments of the present disclosure is illustrated. Figure 5 The document describes the signaling between UE 51 and LMF 52, where UE 51 and LMF 52 can be respectively... Figure 1 and Figure 2 Examples of the first device 110 and the second device 120.

[0084] Figure 5The description describes a process where a first local training with unlabeled data is performed at the UE side, a second local training with labeled data is performed at the LMF, and a final aggregation is performed at the LMF side. Alternatively or additionally, the second training with labeled data can be performed at the gNB, and the final aggregation can be performed at the LMF; or the second training with locally collected labeled data can be performed at a second UE, and the final aggregation can be performed at either the gNB or the LMF.

[0085] In the signaling stream used for federated semi-supervised learning (SSL) training, UE 51 can perform SSL based on unlabeled data, and LMF 52 can perform supervised learning (SL) training using labeled data. LMF 52 then iteratively executes different local neural network weights. and The aggregation / combination continues until the final NN model is obtained.

[0086] At 501, initial steps for UE localization can be performed, including initializing the localization process and requesting UE capabilities. At 502, after the initialization process, LMF 52 can decide, based on UE capabilities, to continue utilizing the federated SSL method with data augmentation to obtain a better ML-based model in terms of accuracy, robustness, and generalization ability. In some embodiments, LMF 52 can determine configuration information related to data-augmented SSL training based on UE 51's capabilities. At 503, LMF 52 can send a request to UE 51 for SSL training with data augmentation options.

[0087] At point 503, LMF 52 can send a request to UE 51 for SSL training with data augmentation options. Once approval to participate in this federated SSL training is obtained from UE 51 at point 504, LMF 52 can provide UE 51 at point 505 with the necessary configuration and parameters for SSL training with data augmentation, including data augmentation functions (e.g., Figure 4 In g And SSL-related parameters such as the size of unlabeled data, NN function f Parameters (e.g., initial model parameters for the AI / ML model deployed at UE 51), convex combination weights, etc. In some embodiments, LMF 52 may send configuration information determined at 502 to UE 51.

[0088] The iterative training process begins at 506. Conditions for determining whether to continue iteration may include, but are not limited to, a predefined number of training iterations or conditions for aggregating the accuracy of the ML model. At 507, UE 51 may perform SSL training with data augmentation following the LMF directive (i.e., configuration information from the decision of LMF 52). Once this SSL training is performed, the local NN weights are adjusted at 508. The model parameters of the trained AI / ML model deployed at UE 51 can be sent to LMF 52.

[0089] Furthermore, at point 509, LMF 52 can perform SL training using available labeled datasets, which allows for the acquisition of local NN weights after training. At position 510, LMF 52 can adjust the weights of two neural networks. and Perform aggregation (such as a convex combination of predefined individual weights, given as) ,in and (It is non-negative and sums to one). At 511, LMF 52 may optionally share the aggregated NN weights and return to step 7, where UE 51 may continue SSL training on the unlabeled data, but using the updated NN weights received from LMF.

[0090] Collaborative / federated SSL training with local training can be performed at different entities / locations within the network, implying labeled or unlabeled datasets. Alternatively or additionally, aggregation can be performed at a central entity (such as the LMF). Furthermore, SL training can be completed in each UE participating in the collaborative / federated approach, and SSL training can be completed on the network side (LMF).

[0091] Alternatively or additionally, when the AI / ML model is deployed on the gNB side, SL training can occur on the gNB side, while SSL training is completed on the LMF side. Each gNB participating in the collaboration / federation approach can utilize the LMF as an aggregation entity.

[0092] In RAN1, there are 5 sub-use cases, which primarily focus on entities (UE, gNB, LMF) that deploy AI / ML models. Alternatively or additionally, based on use cases and entity capabilities, and the availability of label data, the LMF can determine which entity will be used for SL and SSL. This indication can be provided by the LMF to the relevant nodes (UE, PRU, gNB) before the training phase for collaborative / federated learning begins.

[0093] Alternatively or additionally, the LMF can select an appropriate data enhancer, such as a set of enhancement functions. The LMF and UE can exchange messages, whereby the LMF can send requests to the UE and adjust the scope of data enhancements via the Link-to-Point Protocol (LPP).

[0094] The following discussion presents simulation results demonstrating the performance of SSL with data augmentation, where the dataset is used to illustrate the contribution of 3GPP RAN1 to ML-based localization. For the simulation dataset, an InF-DH scenario with 40% factory clutter density, 2-meter factory clutter height, 2-meter factory clutter size, and a 10-meter factory ceiling height was considered. From this dataset, 5000 samples were selected as labeled data, and 10000 samples were selected as test data. The remaining 23384 samples were selected as unlabeled data (i.e., no labels were used).

[0095] The following schemes were trained and compared. Baseline-1 is supervised learning (SL) with only 5000 labeled data samples (no data augmentation (DA)). Baseline-2 is supervised learning (SL) with data augmentation (DA) with 5000 labeled data samples. This corresponds to... Figure 4 This only involves a portion of the LMF process. According to some example embodiments of this disclosure, the scheme involves performing semi-supervised learning (SSL with DA) with data augmentation on 5000 labeled data samples and 23384 unlabeled data samples. This corresponds to... Figure 4 The entire process is described in the text.

[0096] In all cases, the same 5000 samples are used as labeled data, and the same 10000 samples are used as test data. Mean squared error loss is used during training. Figure 6 The diagram illustrates an outline of the neural network model used for training. For data augmentation, an appropriate Gaussian noise generator is used.

[0097] Figure 7 The diagram illustrates the evaluation results between different options. Figure 7 The paper describes the empirical cumulative distribution function (CDF) of the prediction error for three cases. For the test sample... i The prediction error is caused by Given, where the baseline truth value is And the prediction is .

[0098] like Figure 7As shown, evaluation result 701 corresponds to semi-supervised learning with data augmentation (SSL with DA), evaluation result 702 corresponds to supervised learning with data augmentation (SL with DA), and evaluation result 703 corresponds to supervised learning without data augmentation (SSL with DA). Figure 7 It can be observed that data augmentation helps to significantly improve accuracy, while semi-supervised learning using unlabeled data also improves prediction quality.

[0099] Figure 8 A flowchart illustrating an example method 800 implemented at a first device according to some other embodiments of the present disclosure is shown. For ease of understanding, reference will be made to... Figure 1 Method 800 is described from the perspective of the first device 110.

[0100] At box 810, the first device 110 receives configuration information from the second device 120 related to data augmentation-based semi-supervised learning (SSL) training. In some embodiments, the configuration information may be determined based on the capabilities of the first device 110. In some embodiments, the configuration information includes at least one of the following: a data augmentation function for data augmentation; at least one parameter related to SSL training; at least one initial model parameter for the AI / ML model; or at least one convex combination weight.

[0101] In some embodiments, before receiving configuration information, the first device 110 may receive a request for SSL training from the second device 120. Based on receiving the request, the first device 110 sends an approval instruction for SSL training to the second device 120.

[0102] At box 820, the first device 110 trains an artificial intelligence / machine learning (AI / ML) model deployed on the first device 110 based on received configuration information and unlabeled data at the first device 110. In some embodiments, the first device 110 may input unlabeled data into the AI / ML model to obtain predicted output data. The first device 110 may transform unlabeled data into enhanced unlabeled data based on a data augmentation function for data augmentation, and may input the enhanced unlabeled data into the AI / ML model to obtain enhanced predicted output data. In some embodiments, the transformation of unlabeled data may be based on at least one of the following: phase rotation; amplitude amplification or amplitude reduction; Gaussian noise vector; filtering or smoothing; or neural network-based augmentation.

[0103] In some embodiments, the first device 110 may calculate a first loss for the AI / ML model based on the difference between the predicted output data and the enhanced predicted output data, and may send the first loss to the second device 120. In some embodiments, the first device 110 may receive an aggregate loss from the second device 120, wherein the aggregate loss is determined based on the first loss of the AI / ML model. The first device 110 may update at least one model parameter of the AI / ML model based on the aggregate loss.

[0104] At block 830, the first device 110 sends at least one model parameter of a trained AI / ML model to the second device 120. In some embodiments, the first device 110 may receive at least one aggregated model parameter from the second device 120, wherein the at least one aggregated model parameter is determined based on at least one model parameter of the AI / ML model. The first device 110 may acquire an updated AI / ML model with at least one aggregated model parameter, and in response to satisfying an iteration condition, may train an updated AI / ML model based on unlabeled data. In some embodiments, the first device 110 may include a user equipment (UE), or the second device 120 may include a location management function (LMF).

[0105] Figure 9 A flowchart illustrating an example method 900 implemented at a first device according to some other embodiments of the present disclosure is shown. For ease of understanding, reference will be made to... Figure 1 Method 900 is described from the perspective of the first device 110.

[0106] At block 910, the second device 120 sends first configuration information related to data augmentation-based semi-supervised learning (SSL) training to the first device 110. In some embodiments, the second device 120 may determine the first configuration information based on the capabilities of the first device 110 before sending it. The second device 120 may send a request for SSL training to the first device 110 and may receive an indication of approval for SSL training from the first device 110. In some embodiments, the configuration information includes at least one of the following: a data augmentation function for data augmentation; at least one parameter related to SSL training; at least one initial model parameter of the AI / ML model; or at least one convex combination weight.

[0107] At box 920, the second device 120 receives from the first device 110 at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the first configuration information sent and unlabeled data at the first device 110.

[0108] At box 930, the second device 120 trains a second AI / ML model deployed on the second device 120 based on second configuration information related to data augmentation-based supervised learning (SL) training and labeled data on the second device 120. In some embodiments, the labeled data may include labeled input data and labeled output data. The second device 120 may feed the labeled input data into the second AI / ML model to obtain predicted output data. The second device 120 may transform the labeled input data into augmented labeled input data based on a data augmentation function, and may feed the augmented labeled input data into the second AI / ML model to obtain augmented predicted output data. In some embodiments, the transformation of the labeled input data may be based on at least one of the following: phase rotation; amplitude amplification or amplitude reduction; Gaussian noise vector; filtering or smoothing; or neural network-based augmentation.

[0109] In some embodiments, the second device 120 may calculate a second loss of the second AI / ML model based on the difference between the predicted output data and the labeled output data, and may calculate a third loss of the second AI / ML model based on the difference between the enhanced predicted output data and the labeled output data. The second device 120 may receive a first loss of the first AI / ML model from the first device 110.

[0110] In some embodiments, the second device 120 can aggregate a first loss, a second loss, and a third loss based on loss weights to obtain an aggregated loss, and can update at least one model parameter of the second AI / ML model based on the aggregated loss. In some embodiments, the loss weights may include a first loss weight for the first loss, a second loss weight for the second loss, and a third loss weight for the third loss, wherein the first loss weight, the second loss weight, and the third loss weight are non-negative and their sum is one.

[0111] At box 940, the second device 120 aggregates at least one model parameter of a trained first AI / ML model and at least one model parameter of a trained second AI / ML model. In some embodiments, the second device 120 may aggregate at least one model parameter of the first AI / ML model and at least one model parameter of the second AI / ML model based on model parameter weights to obtain at least one aggregated model parameter. In some embodiments, the second device 120 may obtain an updated second AI / ML model with at least one aggregated model parameter, and in response to satisfying an iteration condition, may train the updated second AI / ML model based on labeled data. In some embodiments, the model parameter weights may include: a first model parameter weight for at least one model parameter of the first AI / ML model, and a second model parameter weight for at least one model parameter of the second AI / ML model, wherein the first model parameter weight and the second model parameter weight are non-negative and sum to one.

[0112] In some embodiments, in response to the fulfillment of an iteration condition, the second device 120 may send at least one aggregated model parameter to the first device 110, and may also send an additional request for SSL training to the first device 110. In some embodiments, the first device may include a user equipment (UE); or the second device 120 may include a location management function (LMF).

[0113] In some embodiments, the apparatus capable of performing method 800 (e.g., the first device 110) may include components for performing the corresponding steps of method 800. The components may be implemented in any suitable form. For example, the components may be implemented in a circuit system or a software module.

[0114] In some example embodiments, the apparatus includes: components for receiving configuration information related to data-augmented semi-supervised learning (SSL) training from a second device; components for training an artificial intelligence / machine learning (AI / ML) model deployed on the first device based on the received configuration information and unlabeled data at the first device; and components for sending at least one model parameter of the trained AI / ML model to the second device.

[0115] In some embodiments, the configuration information may be determined based on the capabilities of the first device 110. In some embodiments, the configuration information includes at least one of the following: a data augmentation function for data augmentation; at least one parameter related to SSL training; at least one initial model parameter for the AI / ML model; or at least one convex combination weight.

[0116] In some example embodiments, the apparatus may further include: a component for receiving a request for SSL training from a second device before receiving configuration information; and a component for sending an indication of approval for SSL training to the second device based on receiving the request.

[0117] In some embodiments, the components for training an AI / ML model may include: components for inputting unlabeled data into the AI / ML model to obtain predicted output data; components for transforming the unlabeled data into enhanced unlabeled data based on a data augmentation function for data augmentation; and components for inputting the enhanced unlabeled data into the AI / ML model to obtain enhanced predicted output data. In some embodiments, the transformation of the unlabeled data may be based on at least one of the following: phase rotation; amplitude amplification or amplitude reduction; Gaussian noise vector; filtering or smoothing; or neural network-based augmentation.

[0118] In some embodiments, the components for training an AI / ML model may include: components for calculating a first loss of the AI / ML model based on the difference between predicted output data and enhanced predicted output data; and components for sending the first loss to a second device. In some embodiments, the components for training an AI / ML model may include: components for receiving an aggregate loss from the second device, wherein the aggregate loss is determined based on the first loss of the AI / ML model; and components for updating at least one model parameter of the AI / ML model based on the aggregate loss.

[0119] In some example embodiments, the apparatus may further include: components for receiving at least one aggregated model parameter from a second device, wherein the at least one aggregated model parameter is determined based on at least one model parameter of an AI / ML model; components for obtaining an updated AI / ML model having at least one aggregated model parameter; and components for training an updated AI / ML model based on unlabeled data in response to satisfying an iteration condition. In some example embodiments, the first device may include a user equipment (UE); or the second device may include a location management function (LMF).

[0120] In some embodiments, the apparatus further includes components for performing other steps in some embodiments of method 800. In some embodiments, the components include at least one processor and at least one memory including computer program code, the at least one memory and the computer program code being configured, together with the at least one processor, to enable the execution of the apparatus.

[0121] In some embodiments, the apparatus capable of performing method 900 (e.g., the first device 120) may include components for performing the corresponding steps of method 900. The components may be implemented in any suitable form. For example, the components may be implemented in a circuit system or a software module.

[0122] In some example embodiments, the apparatus includes: components for sending first configuration information to a first device in connection with data-augmented semi-supervised learning (SSL) training; components for receiving from the first device at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; components for training a second AI / ML model deployed at the second device based on second configuration information in connection with data-augmented supervised learning (SL) training and labeled data at the second device; and components for aggregating at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

[0123] In some example embodiments, the apparatus may further include: components for determining the first configuration information based on the capabilities of the first device before sending the first configuration information; components for sending a request for SSL training to the first device; and components for receiving an indication of approval for SSL training from the first device. In some embodiments, the configuration information includes at least one of the following: a data augmentation function for data augmentation; at least one parameter related to SSL training; at least one initial model parameter of the AI / ML model; or at least one convex combination weight.

[0124] In some example embodiments, the labeled data may include labeled input data and labeled output data. Components for training the second AI / ML model may include: components for inputting the labeled input data into the second AI / ML model to obtain predicted output data; components for transforming the labeled input data into enhanced labeled input data based on a data augmentation function; and components for inputting the enhanced labeled input data into the second AI / ML model to obtain enhanced predicted output data. In some example embodiments, the transformation of the labeled input data may be based on at least one of the following: phase rotation; amplitude amplification or amplitude reduction; Gaussian noise vector; filtering or smoothing; or neural network-based augmentation.

[0125] In some example embodiments, the components for training the second AI / ML model may include: components for calculating a second loss of the second AI / ML model based on the difference between the predicted output data and the labeled output data; components for calculating a third loss of the second AI / ML model based on the difference between the enhanced predicted output data and the labeled output data; and components for receiving a first loss of the first AI / ML model from a first device. In some example embodiments, the components for training the second AI / ML model may include: components for aggregating the first loss, the second loss, and the third loss based on loss weights to obtain an aggregated loss; and components for updating at least one model parameter of the second AI / ML model based on the aggregated loss. In some example embodiments, the loss weights may include: a first loss weight for the first loss, a second loss weight for the second loss, and a third loss weight for the third loss, wherein the first loss weight, the second loss weight, and the third loss weight are non-negative and sum to one.

[0126] In some example embodiments, the component for aggregating at least one model parameter of a trained first AI / ML model and at least one model parameter of a trained second AI / ML model may include: a component for aggregating at least one model parameter of the first AI / ML model and at least one model parameter of the second AI / ML model based on model parameter weights to obtain at least one aggregated model parameter, and wherein the second device further includes: a component for obtaining an updated second AI / ML model having at least one aggregated model parameter; and a component for training an updated second AI / ML model based on labeled data in response to satisfying an iteration condition. In some example embodiments, the model parameter weights include a first model parameter weight for at least one model parameter of the first AI / ML model and a second model parameter weight for at least one model parameter of the second AI / ML model, wherein the first model parameter weight and the second model parameter weight are non-negative and sum to one.

[0127] In some example embodiments, the apparatus may further include: a component for sending at least one aggregated model parameter to a first device in response to satisfying an iteration condition; and a component for sending an additional request for SSL training to the first device in response to satisfying an iteration condition. In some embodiments, the first device may include a user equipment (UE); or the second device 120 may include a location management function (LMF).

[0128] In some embodiments, the apparatus further includes components for performing other steps in some embodiments of method 900. In some embodiments, the components include at least one processor and at least one memory including computer program code, the at least one memory and the computer program code being configured, together with the at least one processor, to enable the execution of the apparatus.

[0129] Figure 10 A simplified block diagram of a device 1000 suitable for implementing some example embodiments of the present disclosure is illustrated. Device 1000 may be provided to implement a communication device, such as... Figure 1 The first device 110 or the second device 120 shown. As shown, device 1000 includes: one or more processors 1010, one or more memories 1020 coupled to processors 1010, and one or more communication modules 1040 coupled to processors 1010.

[0130] The communication module 1040 is used for bidirectional communication. The communication module 1040 has at least one antenna to facilitate communication. The communication interface can represent any interface necessary for communication with other network elements.

[0131] Processor 1010 can be of any type suitable for a local technology network, and by way of non-limiting example, can include one or more of the following: general-purpose computer, special-purpose computer, microprocessor, digital signal processor (DSP), and processor based on a multi-core processor architecture. Device 1000 can have multiple processors, such as application-specific integrated circuit chips that are time-dependent on a clock synchronized with the main processor.

[0132] Memory 1020 may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, read-only memory (ROM) 1024, electrically programmable read-only memory (EPROM), flash memory, hard disk, optical disc (CD), digital video disc (DVD), and other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, random access memory (RAM) 1022 and other volatile memories that do not persist during power outages.

[0133] Computer program 1030 includes computer-executable instructions that are executed by the associated processor 1010. Program 1030 may be stored in ROM 1024. Processor 1010 may perform any suitable actions and processes by loading program 1030 into RAM 1022.

[0134] The embodiments of this disclosure can be implemented by program 1030, enabling device 1000 to execute as described in the reference. Figure 5 and Figure 6Any process discussed in this disclosure. Embodiments of this disclosure may also be implemented by hardware or by a combination of software and hardware.

[0135] In some example embodiments, program 1030 may be tangibly contained in a computer-readable medium, which may be included in device 1000 (such as in memory 1020) or in other storage devices accessible by device 1000. Device 1000 may load program 1030 from the computer-readable medium into RAM 1022 for execution. The computer-readable medium may include any type of tangible non-volatile storage, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0136] Figure 11 A block diagram illustrating an example of a computer-readable medium 1100 according to some exemplary embodiments of the present disclosure is shown. The computer-readable medium 1100 has a program 1030 stored thereon. Note that although in Figure 11 The computer-readable medium 1100 is depicted in the form of a CD or DVD, but the computer-readable medium 1100 may be any other form suitable for carrying or storing the program 1030.

[0137] Generally, the various embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, as non-limiting examples, the blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0138] This disclosure also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in a program module and executable in a device on a target real or virtual processor, to perform the functions described above. Figure 8 or Figure 9 Method 800 or 900 is described. Typically, a program module includes routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. The functionality of a program module can be combined or split among program modules as needed in various embodiments. The machine-executable instructions for a program module can be executed on a local or distributed device. In a distributed device, a program module can reside on both local and remote storage media.

[0139] Program code used to perform the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that, when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a stand-alone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0140] In the context of this disclosure, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, etc.

[0141] Computer-readable media can be computer-readable signal media or computer-readable storage media. Computer-readable media can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination thereof. More specific examples of computer-readable storage media will include electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable optical disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. As used herein, the term “non-transient” is a limitation on the medium itself (i.e., tangible, not signaling), and not on the persistence of data storage (e.g., RAM and ROM).

[0142] Furthermore, although the operations are described in a specific order, this should not be construed as requiring such operations, or all operations shown, to be performed in the specific order or sequential order shown to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the foregoing discussion, these details should not be construed as limiting the scope of this disclosure, but rather as descriptions of features specific to particular embodiments. Certain features described in the context of individual embodiments may also be implemented in combinations of individual embodiments. Conversely, various features described in the context of individual embodiments may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0143] Although this disclosure has been described in language specific to structural features and / or method actions, it should be understood that this disclosure as defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are disclosed as examples of implementing the claims.

Claims

1. A first device comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the first device to at least: receive, from a second device, configuration information related to semi-supervised learning (SSL) training based on data augmentation; train, based on the received configuration information and unlabeled data at the first device, an artificial intelligence / machine learning (AI / ML) model deployed at the first device; and send, to the second device, at least one model parameter of the trained AI / ML model.

2. The first device of claim 1, wherein the first device is further caused to: receive, from the second device, a request for the SSL training prior to receiving the configuration information; and based on receiving the request, send, to the second device, an indication of an approval for the SSL training.

3. The first device of any of claims 1 or 2, wherein the configuration information is determined based on capabilities of the first device.

4. The first device of any of claims 1-3, wherein the configuration information comprises at least one of: a data augmentation function for the data augmentation; at least one parameter related to the SSL training; at least one initial model parameter of the AI / ML model; or at least one convex combination weight.

5. The first device of any of claims 1-4, wherein the first device is caused to train the AI / ML model by: inputting the unlabeled data to the AI / ML model to obtain predicted output data; transforming the unlabeled data into augmented unlabeled data based on a data augmentation function for the data augmentation; and inputting the augmented unlabeled data to the AI / ML model to obtain augmented predicted output data.

6. The first device of claim 5, wherein the transformation of the unlabeled data is based on at least one of: a phase rotation; an amplitude upscaling or amplitude downscaling; a Gaussian noise vector; a filtering or smoothing; or a neural network based augmentation.

7. The first device of any of claims 1-6, wherein the first device is caused to train the AI / ML model by: calculating a first loss for the AI / ML model based on a difference between the predicted output data and the augmented predicted output data; and sending, to the second device, the first loss.

8. The first device of claim 7, wherein the first device is caused to train the AI / ML model by: receiving, from the second device, an aggregated loss, wherein the aggregated loss is determined based on the first loss for the AI / ML model; and updating, based on the aggregated loss, the at least one model parameter of the AI / ML model.

9. The first device of any of claims 1-7, wherein the first device is further caused to: ​ ​ receive, from the second device, at least one aggregated model parameter, wherein the at least one aggregated model parameter is determined based on the at least one model parameter of the AI / ML model; obtain an updated AI / ML model with the at least one aggregated model parameter; and in response to satisfying an iteration condition, train the updated AI / ML model based on the unlabeled data.

10. The first device of any one of claims 1-9, wherein at least one of: the first device is a user equipment, UE; or the second device is a location management function, LMF.

11. A second device, comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the second device at least to: send, to a first device, first configuration information related to semi-supervised learning, SSL, training based on data augmentation; receive, from the first device, at least one model parameter of a first artificial intelligence / machine learning, AI / ML, model trained based on the sent first configuration information and unlabeled data at the first device; train a second AI / ML model deployed at the second device based on second configuration information related to supervised learning, SL, training based on data augmentation and labeled data at the second device; and aggregate the at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

12. The second device of claim 11, wherein the second device is further caused to: determine the first configuration information based on capabilities of the first device prior to sending the first configuration information; send, to the first device, a request for the SSL training; and receive, from the first device, an indication of an approval for the SSL training.

13. The second device of claim 11 or 12, wherein the first configuration information comprises at least one of: a data augmentation function for the data augmentation; at least one parameter related to the SSL training; at least one initial model parameter of the AI / ML model; or at least one convex combination weight.

14. The second device of any one of claims 11-13, wherein the labeled data comprises labeled input data and labeled output data, and the second device is caused to train the second AI / ML model by: inputting the labeled input data to the second AI / ML model to obtain predicted output data; transforming the labeled input data into augmented labeled input data based on the data augmentation function; and inputting the augmented labeled input data to the second AI / ML model to obtain augmented predicted output data.

15. The second device of claim 14, wherein the transforming of the labeled input data is based on at least one of: a phase rotation; an amplitude upscaling or amplitude downscaling; a Gaussian noise vector; ​ ​ filtering or smoothing; or neural network based augmentation.

16. The second device of any one of claims 11 to 15, wherein the second device is caused to train the second AI / ML model by: computing a second loss of the second AI / ML model based on a difference between the predicted output data and the labeled output data; computing a third loss of the second AI / ML model based on a difference between the augmented predicted output data and the labeled output data; and receiving, from the first device, a first loss of the first AI / ML model.

17. The second device of claim 16, wherein the second device is caused to train the second AI / ML model by: aggregating the first loss, the second loss, and the third loss based on a loss weight to obtain an aggregated loss; and updating at least one model parameter of the second AI / ML model based on the aggregated loss.

18. The second device of claim 17, wherein the loss weight comprises: a first loss weight for the first loss; a second loss weight for the second loss; and a third loss weight for the third loss, and the first loss weight, the second loss weight, and the third loss weight are non-negative and sum to one.

19. The second device of any one of claims 12 to 18, wherein the second device is caused to aggregate the at least one model parameter of the trained first AI / ML model and the at least one model parameter of the trained second AI / ML model by: aggregating the at least one model parameter of the first AI / ML model and the at least one model parameter of the second AI / ML model based on a model parameter weight to obtain at least one aggregated model parameter, and wherein the second device is further caused to: obtain an updated second AI / ML model having the at least one aggregated model parameter; and in response to satisfying an iteration condition, train the updated second AI / ML model based on the labeled data.

20. The second device of claim 19, wherein the model parameter weight comprises: a first model parameter weight for the at least one model parameter of the first AI / ML model; and a second model parameter weight for the at least one model parameter of the second AI / ML model, and the first model parameter weight and the second model parameter weight are non-negative and sum to one.

21. The second device of any one of claims 12 to 20, wherein the second device is further caused to: in response to satisfying an iteration condition, send the at least one aggregated model parameter to the first device; and send, to the first device, a further request for the SSL training.

22. The first device of any one of claims 12 to 21, wherein at least one of: the first device is a user equipment (UE); or the second device is a location management function (LMF).

23. A method comprising: receiving, at a first device from a second device, configuration information related to data augmentation based semi-supervised learning (SSL) training; training, based on the received configuration information and unlabeled data at the first device, an artificial intelligence / machine learning (AI / ML) model deployed at the first device; and sending, to the second device, at least one model parameter of the trained AI / ML model.

24. A method comprising: sending, at a second device to a first device, first configuration information related to data augmentation based semi-supervised learning (SSL) training; receiving, from the first device, at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; training, based on second configuration information related to data augmentation based supervised learning (SL) training and labeled data at the second device, a second AI / ML model deployed at the second device; and aggregating the at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

25. An apparatus comprising: means for receiving, at a first device from a second device, configuration information related to data augmentation based semi-supervised learning (SSL) training; means for training, based on the received configuration information and unlabeled data at the first device, an artificial intelligence / machine learning (AI / ML) model deployed at the first device; and means for sending, to the second device, at least one model parameter of the trained AI / ML model.

26. An apparatus comprising: means for sending, at a second device to a first device, first configuration information related to data augmentation based semi-supervised learning (SSL) training; means for receiving, from the first device, at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; means for training, based on second configuration information related to data augmentation based supervised learning (SL) training and labeled data at the second device, a second AI / ML model deployed at the second device; and means for aggregating the at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model.

27. A non-transitory computer-readable medium comprising program instructions that, when executed by an apparatus, cause the apparatus to perform at least: receiving, at a first device from a second device, configuration information related to data augmentation based semi-supervised learning (SSL) training; training, based on the received configuration information and unlabeled data at the first device, an artificial intelligence / machine learning (AI / ML) model deployed at the first device; and ​ sending, to the second device, at least one model parameter of the trained AI / ML model.

28. A non-transitory computer-readable medium comprising program instructions that, when executed by an apparatus, cause the apparatus to perform at least: at a second device, sending, to a first device, first configuration information related to semi-supervised learning (SSL) training based on data augmentation; receiving, from the first device, at least one model parameter of a first artificial intelligence / machine learning (AI / ML) model trained based on the sent first configuration information and unlabeled data at the first device; training a second AI / ML model deployed at the second device based on second configuration information related to supervised learning (SL) training based on data augmentation and labeled data at the second device; and aggregating the at least one model parameter of the trained first AI / ML model and at least one model parameter of the trained second AI / ML model. ​

Images