Enhancement of distributed machine learning models in wireless communication systems

Abstract

A user equipment (UE) may receive one or more reference signals from a network node and perform one or more measurements using the one or more reference signals. The UE may compress the one or more measurements into one or more measurement results, and send the one or more measurement results to a server. The UE may request at least one of one or more identifiers (IDs) or one or more models associated with the one or more IDs from a server. The UE may receive at least one of one or more IDs or one or more models from the server, wherein the one or more IDs are provided based on the one or more measurements. The UE may send an ordered list of one or more IDs to the network node, receive a response from the network node indicating a selection of an ID of the one or more IDs, and communicate with the network node using the ID.

Description

Technical Field

[0001] This application relates to wireless communication and, more particularly, to systems, apparatuses, and methods for enhanced distributed machine learning model maintenance in a wireless communication system.

[0002] Description of Related Technologies

[0003] The use of wireless communication systems is growing rapidly. In recent years, wireless devices such as smart phones and tablet computers have become increasingly sophisticated. In addition to supporting telephone calls, many mobile devices (i.e., user equipment or UE) now also provide access to the Internet, email, text messaging, and navigation using the Global Positioning System (GPS), and are capable of operating sophisticated applications that utilize these features. Additionally, there are many different wireless communication technologies and wireless communication standards. Some examples of wireless communication standards include UMTS (e.g., associated with the WCDMA or TD-SCDMA air interface), LTE, LTE-Advanced (LTE-A), NR, HSPA, IEEE 802.11 (WLAN or Wi-Fi), Bluetooth TM etc.

[0004] The introduction of an increasing number of features and functionality in wireless communication devices has also created a continuing need for improved wireless communication and improved wireless communication devices. Specifically, it is important to ensure the accuracy of signals transmitted and received by user equipment (UE) devices (e.g., via wireless devices such as cellular phones, base stations, and relay stations used in wireless cellular communication). Additionally, increasing the functionality of UE devices can place a significant strain on the battery life of UE devices. Therefore, it is also very important to reduce the power requirements in UE device designs while allowing UE devices to maintain good transmit and receive capabilities to improve communication. Accordingly, improvements in this area are desired. Summary of the Invention

[0005] Embodiments of apparatuses, systems, and methods for enhanced distributed machine learning model maintenance in a wireless communication system are presented herein.

[0006] In some embodiments, a user equipment (UE) may receive one or more reference signals from a network node and use the one or more reference signals to perform one or more measurements. The UE may compress the one or more measurements into one or more measurement results and send the one or more measurement results to a server. The UE may request from the server at least one of one or more identifiers (IDs) or one or more models associated with the one or more IDs. The UE may receive from the server the one or more IDs or the at least one of the one or more models, wherein the one or more IDs are provided based on the one or more measurement results. The UE may send an ordered list of the one or more IDs to the network node, receive from the network node a response indicating a selection of an ID among the one or more IDs, and communicate with the network node using the ID.

[0007] According to some embodiments, the one or more reference signals may be channel state information-reference signals (CSI-RS). Additionally or alternatively, the UE may be configured to send one or more measurement results to a first server when connected to external power, when connected to Wi-Fi, when operating under high-fidelity signal conditions, or during a pause in application activity. In some embodiments, the one or more measurement results may include metadata indicating at least one of the following: the training status of one or more IDs, the functionality of one or more IDs, an object, an input or output, the latency benchmark of one or more IDs, memory requirements or accuracy, the compression status of one or more IDs, the inference or operating conditions of one or more IDs, or the preprocessing and postprocessing information of one or more measurements.

[0008] According to further embodiments, the one or more measurements may include at least one of one or more channel state information (CSI) measurements or one or more beam scanning measurements. Additionally or alternatively, the first server may be a machine learning model trainer and the second server may be a machine learning model server. In some embodiments, the ordered list of one or more IDs may be arranged in a priority order based on one or more affinity metrics associated with the one or more IDs.

[0009] In some embodiments, the network node may be configured to receive from the UE a request for one or more training resources. The network node may send the one or more training resources to the UE and also receive from the UE an ordered list of one or more model identifiers (IDs). The network node may then request and receive from the server one or more training samples associated with the one or more model IDs. Additionally, the network node may select a model ID among the one or more model IDs and send to the UE a response indicating the selection of the model ID. The network may then communicate with the UE using the model ID.

[0010] According to some embodiments, one or more training samples may include metadata information corresponding to at least one of the following: date, time, or location of capture, network identity, cell identity, beam configuration and identity, device model and software version, or an assessment of the UE's operating environment based on local measurements and sensors of the UE. Additionally or alternatively, one or more model IDs may include information corresponding to at least one of the following: network provider identity, UE provider identity, Public Land Mobile Network (PLMN) ID, use case ID, or the number of neural networks for one or more use cases.

[0011] In some instances, the UE and the network node may operate as an encoder-decoder pair, respectively. Additionally or alternatively, the network node may be configured to associate at least one of a label or a hash value with one or more training resources. In some embodiments, at least one of the label or the hash value may indicate the measurement conditions of one or more training samples. According to some scenarios, the network node may be configured to request from the UE one or more training samples associated with one or more model IDs.

[0012] According to further embodiments, the network node may receive from the UE a request for one or more reference signals. The network node may send to the UE one or more reference signals and receive from the UE one or more compressed measurement results. Additionally, the network node may send one or more compressed measurement results to a first server. Then, the network node may receive from the UE an ordered list of one or more model IDs and request from a second server one or more training samples corresponding to the one or more model IDs. Additionally or alternatively, the network node may receive from the second server one or more training samples and select a model ID from the one or more model IDs based on the one or more training samples. Then, the network node may send to the UE a response indicating the selection of the model ID and communicate with the UE using the model ID.

[0013] Note that the techniques described herein may be implemented in and / or used with several different types of devices, including but not limited to base stations, access points, cellular phones, portable media players, tablet computers, wearable devices, unmanned aerial vehicles, unmanned flight controllers, automobiles, and / or motor vehicles, and various other computing devices.

[0014] The present invention content aims to provide a brief overview of some of the topics described in this document. Therefore, it should be understood that the above features are only examples and should not be construed as narrowing the scope or essence of the topics described herein in any way. Other features, aspects, and advantages of the topics described herein will become apparent from the following detailed description, drawings, and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] A better understanding of the subject matter can be obtained when considering the following detailed description of various embodiments in conjunction with the following drawings, in which:

[0016] Figure 1 An exemplary (and simplified) wireless communication system according to some embodiments is illustrated;

[0017] Figure 2 An exemplary base station communicating with an exemplary wireless user equipment (UE) device according to some embodiments is illustrated;

[0018] Figure 3 An exemplary block diagram of a UE according to some embodiments is illustrated;

[0019] Figure 4 An exemplary block diagram of a base station according to some embodiments is illustrated;

[0020] Figure 5 An example wireless communication system according to some embodiments including possible distributed entities for machine learning model maintenance is illustrated;

[0021] Figure 6 is a communication flowchart illustrating aspects of an example possible method for performing enhanced distributed machine learning model maintenance in a wireless communication system; and

[0022] Figures 7 to 9 Examples of aspects of various possible methods for performing enhanced distributed machine learning model maintenance in a wireless communication system are illustrated.

[0023] Although the features described herein are susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and described in detail herein. However, it should be understood that the drawings and their detailed description are not intended to be limiting to the specific forms disclosed, but rather are intended to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the subject matter as defined by the appended claims. DETAILED DESCRIPTION

[0024] Acronyms

[0025] Various acronyms are used throughout this disclosure. Definitions of the most commonly used acronyms that may appear throughout this disclosure are provided below:

[0026] ● UE: User Equipment

[0027] ● RF: Radio Frequency

[0028] ● BS: Base Station

[0029] ● GSM: Global System for Mobile Communications

[0030] ● UMTS: Universal Mobile Telecommunications System

[0031] ● LTE: Long Term Evolution

[0032] ● NR: New Radio

[0033] ● TX: Transmit

[0034] ● RX: Receive

[0035] ● RAT: Radio Access Technology

[0036] ● TRP: Transmission and Reception Point

[0037] ● DCI: Downlink Control Information

[0038] ● AI: Artificial Intelligence

[0039] ● NN: Neural Network

[0040] ● CSI: Channel State Information

[0041] ● CSI-RS: Channel State Information Reference Signal

[0042] ● SSB: Synchronization Signal Block

[0043] ● CQI: Channel Quality Indicator

[0044] ● PMI: Precoding Matrix Indicator

[0045] ● RI: Rank Indicator

[0046] ● FR: Frequency Range

[0047] ● gNB: Next Generation Node B

[0048] ● ML: Machine Learning

[0049] ● ID: Identifier

[0050] ● URL: Uniform Resource Locator

[0051] ● PLMN: Public Land Mobile Network

[0052] ●NW: Network

[0053] Terms

[0054] The following is a glossary of terms that will appear in this disclosure:

[0055] Memory medium - Any of various types of non-transitory memory devices or storage devices. The term "memory medium" is intended to include installation media such as CD-ROMs, floppy disks, or magnetic tape devices; computer system memory or random access memory such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory such as flash memory, magnetic media such as hard disk drives or optical storage devices; registers, or other similar types of memory elements, etc. The memory medium may also include other types of non-transitory memory or combinations thereof. In addition, the memory medium may be located in a first computer system that executes a program, or may be located in a different second computer system that is connected to the first computer system via a network such as the Internet. In the latter instance, the second computer system may provide program instructions to the first computer system for execution. The term "memory medium" may include two or more memory media, which may reside in different locations in different computer systems connected, for example, via a network. The memory medium may store program instructions (e.g., embodied as a computer program) executable by one or more processors.

[0056] Carrier medium - The memory medium as described above, and physical transmission media such as buses, networks, and / or other physical transmission media that convey signals such as electrical, electromagnetic, or digital signals.

[0057] Computer system (or computer) - Any of various types of computing systems or processing systems, including personal computer systems (PCs), mainframe computer systems, workstations, network appliances, Internet appliances, personal digital assistants (PDAs), television systems, grid computing systems, or other devices or combinations of devices. Generally speaking, the term "computer system" can be broadly defined as any device (or combination of devices) that includes at least one processor that executes instructions from a memory medium.

[0058] User equipment (UE) (or "UE device") - Any of various types of computer systems or devices that are mobile or portable and perform wireless communication. Examples of UE devices include mobile phones or smartphones (e.g., iPhone TM , Android TM -based phones), tablet computers (e.g., iPad TM, Samsung Galaxy TM ), a portable gaming device (e.g., Nintendo DS TM , PlayStation Portable TM , Gameboy Advance TM , iPhone TM ), wearable devices (e.g., smartwatches, smart glasses), laptop computers, PDAs, portable Internet devices, music players, data storage devices, other handheld devices, automobiles and / or motor vehicles, unmanned aerial vehicles (UAVs) (e.g., drones), UAV controllers (UACs), etc. Generally speaking, the term "UE" or "UE device" can be broadly defined to cover any electronic device, computing device, and / or telecommunications device (or combination of these devices) that is easily transportable by a user and capable of wireless communication.

[0059] Wireless device - Any one of various types of computer systems or devices that perform wireless communication. A wireless device can be portable (or mobile), or it can be stationary or fixed in a location. A UE is an example of a wireless device.

[0060] Communication device - Any one of various types of computer systems or devices that perform communication, where the communication can be wired or wireless. A communication device can be portable (or mobile), or it can be stationary or fixed in a location. A wireless device is an example of a communication device. A UE is another example of a communication device.

[0061] Base station (BS) - The term "base station" has the full scope of its ordinary meaning and includes at least a wireless communication station that is installed in a fixed location and used for communication as part of a wireless telephone system or radio system.

[0062] Processing element (or processor) - Refers to various elements or combinations of elements that are capable of performing functions in a device (e.g., user equipment apparatus or cellular network device). Processing elements can include, for example: a processor and associated memory, portions of or circuits for individual processor cores, entire processor cores, processor arrays, circuits such as ASICs (application - specific integrated circuits), programmable hardware elements such as field - programmable gate arrays (FPGAs), and any one of the various combinations above.

[0063] Wi-Fi - The term "Wi-Fi" has the full scope of its ordinary meaning and includes at least a wireless communication network or RAT that is served by wireless LAN (WLAN) access points and provides connectivity to the Internet through these access points. Most modern Wi-Fi networks (or WLAN networks) are based on the IEEE 802.11 standard and are marketed under the name "Wi-Fi". Wi-Fi (WLAN) networks are different from cellular networks.

[0064] Automatically - means that an action or operation is performed by a computer system (e.g., software executed by a computer system) or a device (e.g., a circuit, a programmable hardware element, an ASIC, etc.) without the action or operation being directly specified or performed through user input. Thus, the term "automatically" contrasts with a user manually performing or specifying an operation, where the user provides input to directly perform the operation. An automatic process can be initiated by user-provided input, but the subsequent actions performed "automatically" are not specified by the user, i.e., are not performed "manually", where the user specifies each action to be performed. For example, a user filling out a spreadsheet by selecting each field and providing input to specify information (e.g., by typing information, selecting checkboxes, radio selections, etc.) is filling out the form manually, even though the computer system must update the form in response to the user's actions. The form can be filled out automatically by a computer system, where the computer system (e.g., software executed on the computer system) analyzes the fields of the form and fills out the form without any user input specifying the answers to the fields. As indicated above, the user can invoke the automatic filling of the form but does not participate in the actual filling of the form (e.g., the user does not manually specify the answers to the fields but they are completed automatically). This specification provides various examples of operations that are performed automatically in response to actions that the user has taken.

[0065] Configured to - Various components can be described as "configured to" perform one or more tasks. In such contexts, "configured to" is a broad statement generally meaning "having" the "structure" to perform one or more tasks during operation. Thus, even when a component is not currently performing a task, the component can be configured to perform the task (e.g., a collection of electrical conductors can be configured to electrically connect a module to another module, even when the two modules are not connected). In some contexts, "configured to" can be a broad statement generally meaning "having" the "circuitry" to perform one or more tasks during operation. Thus, even when a component is not currently powered on, the component can be configured to perform a task. Generally, the circuitry that forms the structure corresponding to "configured to" can include hardware circuitry.

[0066] For ease of description, various components may be described as performing one or more tasks. Such descriptions should be interpreted to include the phrase "configured to". A component described as being configured to perform one or more tasks is expressly intended not to invoke the sixth paragraph of 35 U.S.C. § 112 for that component.

[0067] Figure 1 and Figure 2 - Exemplary communication system

[0068] Figure 1 Illustrated is an exemplary (and simplified) wireless communication system that can implement various aspects of the present disclosure according to some embodiments. Note that Figure 1 the system is only one example of possible systems, and the embodiment can be implemented in any of various systems as needed.

[0069] As shown, the exemplary wireless communication system includes a base station 102 that communicates with one or more (e.g., any number of) user equipment 106A, 106B, etc. up to 106N via a transmission medium. Each user equipment may be referred to herein as a "user equipment" (UE) or a UE device. Thus, user equipment 106 is referred to as a UE or a UE device.

[0070] Base station 102 may be a transceiver base station (BTS) or a cell site, and may include hardware and / or software for implementing wireless communication with UEs 106A to 106N. If base station 102 is implemented in the context of LTE, it may be referred to as an "eNodeB" or "eNB". If base station 102 is implemented in the context of 5G NR, it may alternatively be referred to as a "gNodeB" or "gNB". Base station 102 may also be equipped to communicate with network 100 (e.g., the core network of a cellular service provider, a telecommunications network such as the public switched telephone network (PSTN), and / or the Internet, and various possible networks). Thus, base station 102 can facilitate communication between user equipment and / or between user equipment and network 100. The communication area (or coverage area) of a base station may be referred to as a "cell". Also as used herein, with respect to a UE, a base station may sometimes be considered to represent the network when considering the uplink and downlink communications of the UE. Thus, a UE that communicates with one or more base stations in a network may also be understood to be a UE that communicates with the network.

[0071] Base station 102 and user equipment may be configured to communicate via a transmission medium using any of various radio access technologies (RATs), which are also referred to as wireless communication technologies or telecommunications standards, such as UMTS (WCDMA), LTE, advanced LTE (LTE-A), LAA / LTE-U, 5G NR, Wi-Fi, etc.

[0072] Base station 102 and other similar base stations operating according to the same or different cellular communication standards can thus provide a network of one or more cells that can provide continuous or near - continuous overlapping services to UE 106 and similar devices over a geographical area via one or more cellular communication standards.

[0073] Note that UE 106 may be capable of communicating using multiple wireless communication standards. For example, UE 106 may be configured to communicate using either or both of the 3GPP cellular communication standard or the 3GPP2 cellular communication standard. In some embodiments, UE 106 may be configured to perform techniques for model ID selection using machine learning assistance in a wireless communication system, such as according to the various methods described herein. UE 106 may also be configured or alternatively configured to communicate using WLAN, Bluetooth TM , one or more Global Navigation Satellite Systems (GNSS, such as GPS or GLONASS), one and / or more mobile television broadcast standards (e.g., ATSC - M / H), etc. Other combinations of wireless communication standards (including more than two wireless communication standards) are also possible.

[0074] Figure 2 Exemplary user equipment 106 (e.g., one of devices 106A to 106N) communicating with base station 102 according to some embodiments is illustrated. UE 106 can be a device with wireless network connectivity, such as a mobile phone, a handheld device, a wearable device, a computer or tablet computer, an unmanned aerial vehicle (UAV), an unmanned aircraft controller (UAC), a vehicle, or almost any type of wireless device. UE 106 may include a processor (processing element) configured to execute program instructions stored in a memory. UE 106 can execute any of the method embodiments described herein by executing such stored instructions. Alternatively or in addition, UE 106 may include programmable hardware elements, such as a Field Programmable Gate Array (FPGA), an integrated circuit, and / or any of various other possible hardware components configured to (e.g., individually or in combination) execute any of the method embodiments described herein or any part of any of the method embodiments described herein. UE 106 can be configured to communicate using any of multiple wireless communication protocols. For example, UE 106 can be configured to communicate using two or more of LTE, LTE - A, 5G NR, WLAN, or GNSS. Other combinations of wireless communication standards are also possible.

[0075] UE 106 may include one or more antennas that communicate using one or more wireless communication protocols according to one or more RAT standards. In some embodiments, UE 106 may share one or more portions of the receive chain and / or the transmit chain among multiple wireless communication standards. The shared radio components may include a single antenna or may include multiple antennas (e.g., for a multiple-input, multiple-output or “MIMO” antenna system) for performing wireless communication. Generally, the radio components may include any combination of a baseband processor, analog RF signal processing circuitry (e.g., including filters, mixers, oscillators, amplifiers, etc.) or digital processing circuitry (e.g., for digital modulation and other digital processing). Similarly, the radio components may use the foregoing hardware to implement one or more receive chains and transmit chains. For example, UE 106 may share one or more portions of the receive chain and / or the transmit chain among multiple wireless communication technologies (such as those discussed above).

[0076] In some embodiments, UE 106 may include any number of antennas and may be configured to transmit and / or receive directional wireless signals (e.g., beams) using the antennas. Similarly, BS102 may also include any number of antennas and may be configured to transmit and / or receive directional wireless signals (e.g., beams) using the antennas. To receive and / or transmit such directional signals, the antennas of UE 106 and / or BS102 may be configured to apply different “weights” to different antennas. The process of applying these different weights may be referred to as “precoding”.

[0077] In some embodiments, UE 106 may include independent transmit chains and / or receive chains (e.g., including independent antennas and other radio components) for each wireless communication protocol that it is configured to communicate with. As an additional possibility, UE 106 may include one or more radio components shared among multiple wireless communication protocols, and one or more radio components uniquely used by a single wireless communication protocol. For example, UE 106 may include shared radio components for communicating using either LTE or NR, and independent radio components for communicating using each of Wi-Fi and Bluetooth TM among others. Other configurations are possible.

[0078] Figure 3 - Block diagram of an exemplary UE device

[0079] Figure 3A block diagram of an exemplary UE 106 is illustrated in accordance with some embodiments. As shown, the UE 106 may include a system-on-a-chip (SOC) 300, which may include portions for various purposes. For example, as shown, the SOC 300 may include a processor 302 that may execute program instructions for the UE 106, and a display circuit 304 that may perform graphics processing and provide a display signal to a display 360. The SOC 300 may also include a sensor circuit 370, which may include components for sensing or measuring any of a variety of possible characteristics or parameters of the UE 106. For example, the sensor circuit 370 may include a motion sensing circuit configured to detect motion of the UE 106 using, for example, a gyroscope, an accelerometer, and / or any of a variety of other motion sensing components. As another possibility, the sensor circuit 370 may include one or more temperature sensing components for measuring the temperature of each of one or more antenna panels and / or other components of the UE 106. Any of a variety of other possible types of sensor circuits may also or alternatively be included in the UE 106 as needed. The processor 302 may also be coupled to a memory management unit (MMU) 340, which may be configured to receive addresses from the processor 302 and translate those addresses into locations in a memory (such as memory 306, read-only memory (ROM) 350, NAND flash memory 310) and / or other circuitry or devices, such as the display circuit 304, radio components 330, connector I / F 320, and / or the display 360. The MMU 340 may be configured to perform memory protection and page table translation or setup. In some embodiments, the MMU 340 may be included as part of the processor 302.

[0080] As shown, the SOC 300 may be coupled to various other circuits of the UE 106. For example, the UE 106 may include various types of memory (e.g., including NAND flash 310), a connector interface 320 (e.g., for coupling to a computer system, docking station, charging station, etc.), a display 360, and wireless communication circuitry 330 (e.g., for LTE, LTE-A, NR, Bluetooth TM, Wi-Fi, GPS, etc.). The UE device 106 may include or be coupled to at least one antenna (e.g., 335a), and may include multiple antennas (e.g., as illustrated by antennas 335a and 335b) for performing wireless communication with the base station and / or other devices. Antennas 335a and 335b are shown by way of example, and the UE device 106 may include fewer or more antennas. Generally speaking, one or more antennas are collectively referred to as antenna 335. For example, the UE device 106 may use antenna 335 via radio circuitry 330 to perform wireless communication. The communication circuitry may include multiple receive chains and / or multiple transmit chains for receiving and / or transmitting multiple spatial streams, such as in a multiple-input multiple-output (MIMO) configuration. As mentioned above, in some embodiments, the UE may be configured to perform wireless communication using multiple wireless communication standards.

[0081] The UE 106 may include hardware and software components for implementing methods for the UE 106 to perform techniques for model ID selection using machine learning assistance in a wireless communication system, such as further described hereinbelow. The processor 302 of the UE device 106 may be configured to implement part or all of the methods described herein, for example, by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). In other embodiments, the processor 302 may be configured as a programmable hardware element, such as an FPGA (field-programmable gate array) or as an ASIC (application-specific integrated circuit). Additionally, the processor 302 may be coupled to other components and / or may interoperate with other components, as Figure 3 shown, to perform techniques for model ID selection using machine learning assistance in a wireless communication system in accordance with various embodiments disclosed herein. The processor 302 may also implement various other applications and / or end-user applications running on the UE 106.

[0082] In some embodiments, the radio component 330 may include independent controllers dedicated to controlling communication for various respective RAT standards. For example, as Figure 3 shown, the radio component 330 may include a Wi-Fi controller 352, a cellular controller (e.g., an LTE and / or LTE-A controller) 354, and a Bluetooth TM controller 356, and in at least some embodiments, one or more or all of these controllers may be implemented as respective integrated circuits (simply referred to as ICs or chips), which communicate with each other and with the SOC 300 (and more specifically, with the processor 302). For example, the Wi-Fi controller 352 may communicate with the cellular controller 354 via a cell-ISM link or a WCI interface, and / or Bluetooth TMThe controller 356 can communicate with the cellular controller 354 via a cell-ISM link or the like. Although three separate controllers are illustrated within the radio component 330, other implementations with fewer or more similar controllers for various different RATs can be implemented in the UE device 106.

[0083] Additionally, implementations are envisioned in which the controller can implement functionality associated with multiple radio access technologies. For example, according to some implementations, in addition to the hardware and / or software components for performing cellular communication, the cellular controller 354 can also include hardware and / or software components for performing one or more activities associated with Wi-Fi, such as Wi-Fi preamble detection, and / or generation and transmission of Wi-Fi physical layer preamble signals.

[0084] Figure 4 -Block diagram of an exemplary base station

[0085] Figure 4 A block diagram of an exemplary base station 102 according to some implementations is illustrated. Note that Figure 4 the base station is only one example of a possible base station. As shown, the base station 102 can include a processor 404 that can execute program instructions for the base station 102. The processor 404 can also be coupled to a memory management unit (MMU) 440 or other circuitry or device, which can be configured to receive addresses from the processor 404 and translate those addresses into locations in a memory (e.g., memory 460 and read-only memory (ROM) 450).

[0086] The base station 102 can include at least one network port 470. The network port 470 can be configured to couple to a telephone network and provide access to multiple devices such as the UE device 106 to the telephone network as described above in Figure 1 and Figure 2 . The network port 470 (or an additional network port) can also be configured or alternatively can be configured to couple to a cellular network, such as the core network of a cellular service provider. The core network can provide mobility-related services and / or other services to multiple devices (such as the UE device 106). In some cases, the network port 470 can be coupled to the telephone network via the core network, and / or the core network can provide the telephone network (e.g., in other UE devices served by the cellular service provider).

[0087] In some embodiments, the base station 102 may be a next-generation base station, e.g., a 5G New Radio (5G NR) base station, or a "gNB". In such embodiments, the base station 102 may be connected to a traditional Evolved Packet Core (EPC) network and / or connected to a NR Core (NRC) network. Additionally, the base station 102 may be regarded as a 5G NR cell and may include one or more Transmission and Reception Points (TRP). Further, a UE capable of operating according to 5G NR may be connected to one or more TRP within one or more gNBs.

[0088] The base station 102 may include at least one antenna 434 and possibly multiple antennas. The antenna 434 may be configured to operate as a wireless transceiver and may also be configured to communicate with the UE device 106 via the radio component 430. The antenna 434 communicates with the radio component 430 via the communication link 432. The communication link 432 may be a receive link, a transmit link, or both. The radio component 430 may be designed to communicate via various radio communication standards, which include but are not limited to 5G NR, 5G NR SAT, LTE, LTE-A, UMTS, Wi-Fi, etc.

[0089] The base station 102 may be configured to perform wireless communication using multiple radio communication standards. In some instances, the base station 102 may include multiple radio components, which may enable the base station 102 to communicate according to multiple radio communication technologies. For example, as a possibility, the base station 102 may include an LTE radio component for performing communication according to LTE and a 5G NR radio component for performing communication according to 5G NR. In this case, the base station 102 may be capable of operating as both an LTE base station and a 5G NR base station. As another possibility, the base station 102 may include a multi-mode radio component capable of performing communication according to any one of multiple radio communication technologies (e.g., 5G NR and Wi-Fi, 5G NR SAT and Wi-Fi, LTE and Wi-Fi, LTE and UMTS, LTE and UMTS and GSM, etc.).

[0090] As described further below in this document, BS102 may include hardware and software components for implementing or supporting the specific implementation of the features described in this document. The processor 404 of the base station 102 may be configured to implement and / or support the specific implementation of part or all of the methods described in this document, for example, by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). Alternatively, the processor 404 may be configured as a programmable hardware element such as an FPGA (Field Programmable Gate Array), or as an ASIC (Application Specific Integrated Circuit), or a combination thereof. In the case of certain RATs (e.g., Wi-Fi), the base station 102 may be designed as an access point (AP), in which case the network port 470 may be implemented to provide access to a wide area network and / or one or more local area networks. For example, it may include at least one Ethernet port, and the radio component 430 may be designed to communicate according to the Wi-Fi standard.

[0091] In addition, as described in this document, the processor 404 may include one or more processing elements. Therefore, the processor 404 may include one or more integrated circuits (ICs) configured to perform the functions of the processor 404. In addition, each integrated circuit may include circuits (e.g., a first circuit, a second circuit, etc.) configured to perform one or more functions of the processor 404.

[0092] In addition, as described in this document, the radio component 430 may include one or more processing elements. Therefore, the radio component 430 may include one or more integrated circuits (ICs) configured to perform the functions of the radio component 430. In addition, each integrated circuit may include circuits (e.g., a first circuit, a second circuit, etc.) configured to perform the functions of the radio component 430.

[0093] Reference Signal

[0094] A wireless device, such as a user equipment, may be configured to perform various tasks including using reference signals (RS) provided by one or more cellular base stations. For example, initial access and beam measurements of a wireless device may be performed at least in part based on synchronization signal blocks (SSB) provided by one or more cells provided by one or more cellular base stations within the communication range of the wireless device. Another type of reference signal commonly provided in a cellular communication system may include channel state information (CSI) RS. Various types of CSI-RS may be provided for tracking (e.g., for time and frequency offset tracking), beam management (e.g., CSI-RS is configured to have repetitions to help determine one or more beams for uplink and / or downlink communication), and / or channel measurements (e.g., CSI-RS configured in a resource set to measure the quality of a downlink channel and report information related to the quality measurement to a base station), and various possibilities. For example, in the case where CSI-RS is used for CSI acquisition, the UE may periodically perform channel measurements and transmit channel state information (CSI) to the BS. The base station may then receive and use the channel state information during communication with the wireless device to determine adjustments to various parameters. Specifically, the BS may use the received channel state information to adjust the coding of its downlink transmissions to improve downlink channel quality.

[0095] In many cellular communication systems, a base station may periodically transmit some or all such reference signals (or pilot signals), such as SSB and / or CSI-RS. In some cases, non-periodic reference signals may also be provided or alternatively (e.g., non-periodic reference signals for non-periodic CSI reporting).

[0096] As a detailed example, at least according to some embodiments, in the 3GPP NR cellular communication standard, the channel state information based on CSI-RS feedback from a UE for CSI acquisition may include a channel quality indicator (CQI), a precoding matrix indicator (PMI), a rank indicator (RI), a CSI-RS resource indicator (CRI), an SSBRI (SS / PBCH resource block indicator and one or more of a layer indicator (LI).

[0097] Channel quality information may be provided to a base station for link adaptation, e.g., for providing guidance on which modulation and coding scheme (MCS) the base station should use when transmitting data. For example, when the communication quality of the downlink channel between the base station and the UE is determined to be high, the UE may feedback a high CQI value, which may cause the base station to transmit data using a relatively high modulation order and / or a low channel coding rate. Another example, when the communication quality of the downlink channel between the base station and the UE is determined to be low, the UE may feedback a low CQI value, which may cause the base station to transmit data using a relatively low modulation order and / or a high channel coding rate.

[0098] PMI feedback may include preferred precoding matrix information and may be provided to the base station to indicate which MIMO precoding scheme the base station should use. In other words, the UE may measure the quality of the downlink MIMO channel between the base station and the UE based on the pilot signals received on the channel and may recommend which MIMO precoding the desired base station should apply through PMI feedback. In some cellular systems, the PMI configuration is represented in matrix form, which provides linear MIMO precoding. The base station and the UE may share a codebook consisting of multiple precoding matrices, where each MIMO precoding matrix in the codebook may have a unique index. Thus, as part of the channel state information fed back by the UE, the PMI may include the index (or possibly multiple indexes) corresponding to the most preferred MIMO precoding matrix (or matrices) in the codebook. This may enable the UE to minimize the amount of feedback information. Thus, at least according to some embodiments, the PMI may indicate which precoding matrix from the codebook should be used for transmission to the UE.

[0099] For example, when the base station and the UE have multiple antennas, the rank indicator information (RI feedback) may indicate the number of transmission layers that the UE determines can be supported by the channel, which can be achieved through spatial multiplexing for multi-layer transmission. The RI and the PMI may jointly allow the base station to know which precoding needs to be applied to which layer, for example, depending on the number of transmission layers.

[0100] In some cellular systems, the PMI codebook is defined according to the number of transmission layers. In other words, for R-layer transmission, N N t ×R matrices may be defined (e.g., where R represents the number of layers, N t represents the number of transmitter antenna ports, and N represents the size of the codebook). In such a scenario, the number of transmission layers (R) may match the rank value of the precoding matrix (N t ×R matrix), and thus R may be referred to as the "rank indicator (RI)" in this context.

[0101] Accordingly, the channel state information may include an allocated rank (e.g., rank indicator or RI). For example, a MIMO-capable UE communicating with a BS may include four receiver chains, e.g., may include four antennas. The BS may also include four or more antennas to enable MIMO communication (e.g., 4×4 MIMO). Accordingly, the UE is capable of receiving up to four (or more) signals (e.g., layers) from the BS simultaneously. A layer-to-antenna mapping may be applied, e.g., each layer may be mapped to any number of antenna ports (e.g., antennas). Each antenna port may transmit and / or receive information associated with one or more layers. The rank may include multiple bits and may indicate the number of signals that the BS may transmit to the UE in an upcoming time period (e.g., during an upcoming transmission time interval or TTI). For example, an indication of rank 4 may indicate that the BS will transmit 4 signals to the UE. As a possibility, the length of the RI may be two bits (e.g., since two bits are sufficient to distinguish 4 different rank values). Note that according to various embodiments, other numbers and / or configurations of antennas (e.g., at either or both of the UE or the BS) and / or other numbers of data layers are possible.

[0102] Enhancement of Distributed Machine Learning Models in Wireless Communication Systems

[0103] Interest in the use of artificial intelligence and machine learning-based algorithms and tools is increasing. It is possible to utilize such tools in any of the various possible areas of cellular communication. One such area may include channel state feedback and beam measurement feedback, which are used to train models of encoder / decoder pairs (e.g., UE / BS pairs) to improve link budget and / or other wireless communication system characteristics. For example, while an encoder-decoder pair may be well-specified and efficient enough given appropriate training samples and an associated model, it may be beneficial to allow both the encoder and decoder models to independently evolve (e.g., update) over time and as the operating environment may change. More specifically, the encoder (e.g., the UE) may be able to utilize uplink compression techniques when providing information (e.g., training samples) for a machine learning model, while the base station (e.g., the decoder) may perform decoding or decompression of the training samples in order to be able to identify or select a compatible machine learning model for communication.

[0104] For example, there may be a wireless communication scenario where a mobile device may need to transmit a summary of its observations to one or more entities such as a base station (e.g., a next-generation node B (gNB)), a network-side server, and / or a UE-side server. Additionally, some examples of the observations may include channel state feedback and beam measurement feedback. Additionally, it may be desirable to minimize the number of bits required to transmit the observations in order to reduce transmit power, extend battery life, and minimize network and air interface uplink resource consumption.

[0105] In some embodiments, one approach can be to compress the observations (or associated signaling) to minimize the number of bits required to transmit them. Compression techniques generally require a compressor (e.g., an encoder such as a UE) and a matching decompressor (e.g., a decoder such as a base station). Depending on some scenarios, the encoder should generally be tuned to the statistical characteristics of the observations and the acceptable distortion level. These characteristics can vary with the device itself (including software version), its operating environment (channel characteristics), network configuration, and various other factors. In other words, it may be beneficial for the encoder (e.g., a wireless device) to perform occasional or scheduled measurements of its operating environment to understand the state of the channel in terms of how efficient its wireless communication is. Thus, when reporting its measurements, it may also be desirable to minimize the number of bits required to transmit this information. Therefore, prior to transmission, the encoder can benefit from compressing these measurements or measurement results (e.g., compressing into a reduced number of bits) in order to achieve reduced transmission power (and thus potentially extend battery life) and minimize uplink resource consumption. Similarly, the decoder communicating with the encoder can also benefit from the reduced number of bits of these measurements (e.g., compression), as decompressing a smaller number of bits may require less processing power.

[0106] In addition, it may be desirable to allow the network and device software to evolve, update, or change according to their own schedules, and to select any implementation of the encoder and decoder, as long as the two remain compatible. The network may have additional considerations, such as implementing a single decoder that is compatible with the encoders of various device types from different device manufacturers. In addition, device and network vendors may prefer to complete the development of the encoder and decoder while protecting user privacy (e.g., identity, location, operating environment) and minimizing the disclosure of proprietary information about device or network capabilities and configurations.

[0107] According to some embodiments, an encoder-decoder pair can be associated with a model ID. For example, the model ID can be further associated with the observation statistics and a compatibility field (e.g., NW vendor identifier, UE vendor identifier, etc.), or include the observation statistics and a compatibility field. Thus, the field corresponding to compatibility can determine whether the model ID can be used for communication between the encoder (e.g., compressor) and the decoder (e.g., decompressor). Therefore, it may be beneficial to develop model identifiers (IDs) for the encoder and decoder (and the encoder / decoder pair) using model learning techniques while maintaining compatibility.

[0108] For example, according to some embodiments, by associating a model ID with a set of observed statistics (e.g., channel characteristics, hardware (HW) or software (SW) version, network configuration, etc.), when an encoder / decoder pair (e.g., as an example, a UE / BS pair) operates under different conditions (e.g., operates under different channel characteristics, different HW / SW versions, etc.), different model IDs can be appropriately selected to achieve more efficient communication corresponding to the current operating conditions of the pair. Additionally, by updating the observed statistics associated with the model ID, the model can also be effectively updated through the association of the model with the model ID. In other words, the updated model used by the encoder / decoder pair can reflect or include the updated observed statistics (e.g., channel characteristics, HW / SW version, etc.) through the association of the model ID with the model. Therefore, more efficient communication between the pair can be achieved by continuously or semi-persistently training the model based on the observed conditions and subsequently selecting a compatible and most efficient model ID.

[0109] In addition, beam measurement techniques for beamforming are widely used in wireless communication systems, typically as a technique for improving the link budget. Beamforming can be implemented, for example, in both a cellular base station (e.g., gNB, eNB, etc.) and a wireless device (e.g., UE) in a cellular communication system. At least in some instances, a properly matched and efficient beam pair can contribute to improving system performance.

[0110] For a BS-UE beam pair, it is possible that the BS transmits multiple downlink reference signals, where different BS beams can be applied to different reference signals so that the UE can measure the quality of each beam. The UE can also use different receive beams to receive different instances of a reference signal, e.g., to identify the best UE beam for each BS beam. In some embodiments, the downlink reference signals provided by the BS can include a synchronization signal block (SSB) or a channel state information reference signal (CSI-RS). Therefore, to identify the BS-UE beam pair, the UE may need to perform measurements on several BS beams using a UE beam scanning operation.

[0111] However, it is possible to use machine learning techniques to avoid the need for the UE to perform such extensive beam measurements. Such machine learning techniques can be used, for example, to help identify the best BS beam without directly measuring the BS beam, such that the UE can potentially identify the UE beam to adapt to the best BS beam faster and / or with less overhead than otherwise.

[0112] In one possible scenario, aspects of such machine learning techniques can be implemented on the BS side. Alternatively, in another possible scenario, machine learning can be implemented on the UE side. As another possibility, machine learning can be implemented in part by each of the BS side and the UE side. For example, in one scenario, training (e.g., machine learning) can be implemented on the BS side, while inference can be implemented on the UE side, and in another scenario, training can be implemented on the UE side, while inference can be implemented on the BS side. Inference can be based on metadata or operating conditions associated with the model ID. It is possible that the use of which of such scenarios can be configured by the BS potentially at least in part based on the UE's ability to support one or more such scenarios (e.g., as may be indicated by the UE in the capability information provided by the UE to the BS).

[0113] To support the use of such techniques, it may be important to provide a framework according to which wireless devices and cellular networks can exchange information to determine whether such techniques are mutually supportive and potentially negotiate or agree on the characteristics and parameters according to which artificial intelligence / machine learning model maintenance is performed, and / or exchange information for supporting the operation of artificial intelligence / machine learning used in performing cellular communications, and provide techniques for using machine learning in performing cellular communications.

[0114] Figure 5 - Distributed entities for machine learning

[0115] Figure 5 An example wireless communication system is illustrated that includes possible distributed entities for machine learning model maintenance according to some embodiments. More specifically, Figure 5 it is illustrated how a user equipment 512 (which can be an example of the UE 106), a network node 514 (which can be an example of the BS 102), and machine learning distributed entities operating in a network 516 can select and use a compatible and efficient model ID for communication.

[0116] For example, Figure 5 a network 516 is illustrated in which a user equipment 512 can communicate with a network node 514. Additionally, according to some embodiments, the network 516 can include various distributed machine learning entities, such as a model trainer 510, a model server 504 (and a model repository 502), and a training data server 508 (and a training data repository 506), which can allow collection of training samples from devices, training and distribution of models, and distribution of training samples to the network.

[0117] According to some embodiments, the model trainer 510 may receive encoder or decoder training samples from a device such as the user device 512 and generate an encoder model based on the training samples. For example, the user device 512 may collect or aggregate training samples (e.g., measurement results) when receiving training resources (e.g., reference signals) from the network node 514. Thus, the model trainer 510 may be able to use or further analyze the training samples to generate / update / train a machine learning model and form an affinity metric for the model.

[0118] In addition, the model server 504 may maintain or store the trained models associated with the model IDs in a repository and provide them to the devices upon request. For example, the model server may include a repository in the model repository 502 for storing the models, the associated affinity metrics, and / or the model IDs. Thus, once the model trainer 510 has trained a model based on the affinity metric, the model trainer 510 may send the trained model and the affinity metric to the model server 504 accessible by the user device 512.

[0119] According to some embodiments, the user device may access or request model relationships from the model server. For example, the user device 512 may request an encoder model and / or the associated affinity metric or model ID from the model server 504. In some embodiments, the user device 512 may request a specific model ID from the model server 504, which may also be provided to the device during an operating system (OS) update or discovered during a negotiation of model IDs with one or more networks. Additionally or alternatively, the user device 512 may provide metadata to the model server and search for a suitable model or set of models. In some embodiments, the user device 512 may also query the model server 504 for an upgraded model based on the existing model IDs it has access to and potentially request decoder compatibility. Thus, according to some embodiments, the model server 504 may use its knowledge of the model relationships to provide a suitable model for the device.

[0120] Additionally or alternatively, the model trainer 510 may also send training samples with model IDs for the decoder to the training data server 508. The training data server 508 may store the decoder training samples associated with the model IDs in a repository such as the training data repository 506 and provide them to the network upon request. For example, according to some embodiments, the network node 514 may request training samples associated with a model ID from the training data server 508. Additionally or alternatively, according to some embodiments, the network may also query the training data server for the latest set of training samples evolved from a specific model ID.

[0121] Thus, once the user equipment 512 has received the model IDs it has requested access to (based on the affinity metric) and the network node 514 has received the training samples it has requested, the user equipment 512 can transmit an ordered list of model IDs to the network node 514 as part of the model ID negotiation. Once the network node 514 has received the ordered list, it can select or pick the model ID for which it has a suitable decoder from among these models and further communicate that selection or pick to the device.

[0122] Figure 6 - Method for Enhanced Distributed Machine Learning Model Maintenance in a Wireless Communication System

[0123] Thus, it may be beneficial to specify techniques for supporting machine learning-based model ID selection. To illustrate a set of such possible techniques, Figure 6 is a communication flow diagram illustrating a method for enhanced distributed machine learning model maintenance in a wireless communication system according to at least some embodiments.

[0124] Figure 6 Aspects of the method may be implemented by a wireless device and / or a cellular base station (such as the UE 106 and / or BS 102 illustrated and described herein with respect to the various figures), or more generally, in conjunction with any of the computer circuits, systems, devices, elements, or components etc. shown in the above figures as needed. For example, in some embodiments, it may be the case that aspects of the method are implemented by a wireless device, while in other embodiments, it may be the case that aspects of the method are implemented by a cellular base station. For example, the processor (and / or other hardware) of such a device may be configured to cause the device to perform any combination of the illustrated method elements and / or other method elements.

[0125] Note that although at least some elements of the method are described in a manner that involves the use of communication technologies and / or features associated with 3GPP and / or NR specification documents, this description is not intended to limit the present disclosure, and aspects of the method may be used in any suitable wireless communication system as needed. Figure 6 In various embodiments, some of the elements of the illustrated method may be executed simultaneously in a different order than shown, may be replaced by other method elements, or may be omitted. Additional method elements may also be executed as needed. As shown, Figure 6 the method may operate as follows. Figure 6 the method may operate as follows.

[0126] In 610, a wireless device 604 (e.g., UE 106 or user equipment 512) and a network node 606 (e.g., cellular base station 102 or network node 514) may establish a wireless link. According to some embodiments, the wireless link may include a cellular link according to 5G NR. For example, the wireless device may establish a session with an access and mobility management function (AMF) entity of the cellular network via one or more gNBs that provide radio access to the cellular network. As another possibility, the wireless link may include a cellular link according to LTE. For example, the wireless device may establish a session with a mobility management entity (MME) of the cellular network via an eNB that provides radio access to the cellular network. According to various embodiments, other types of cellular links are also possible, and the cellular network may also or alternatively operate according to another cellular communication technology (e.g., UMTS, etc.).

[0127] At least according to some embodiments, establishing the wireless link may include establishing an RRC connection between the wireless device 604 and the serving cellular base station. Establishing the first RRC connection may include configuring various parameters for communication between the wireless device and the cellular base station, establishing environmental information of the wireless device 604, and / or any of various other possible features, e.g., relating to establishing an air interface of the wireless device for cellular communication with the cellular network associated with the cellular base station. After establishing the RRC connection, the wireless device may operate in the RRC connected state. In some instances, the RRC connection may also be released (e.g., after a certain period of inactivity with respect to data communication), in which case the wireless device 604 may operate in the RRC idle state or the RRC inactive state. In some instances, e.g., due to wireless device mobility, change in wireless medium conditions, and / or any of various other possible reasons, the wireless device 604 may perform a handover (e.g., when in the RRC connected mode) or cell reselection (e.g., when in the RRC idle mode or the RRC inactive mode) to a new serving cell.

[0128] According to at least some embodiments, the wireless device 604 may establish multiple wireless links with multiple TRPs of a cellular network, for example, according to a multi-TRP configuration. In this case, the wireless device may be configured to have one or more transmission control indicators (TCIs) (e.g., via RRC signaling), for example, the one or more transmission control indicators may correspond to various beams that can be used for communicating with the TRP. Additionally, it is possible that one or more of the configured TCI states may be activated by a medium access control (MAC) control element (CE) of the wireless device 604 at a particular time. In some instances, the cellular connection between the wireless device 604 and the cellular network may include links to multiple cells operating in different frequency ranges. For example, as a possibility, the wireless device 604 may be attached to at least one cell operating in 3GPP frequency range 1 (FR1) and at least one cell operating in 3GPP frequency range 2 (FR2).

[0129] At least in some instances, establishing a wireless link may include the wireless device 604 providing capability information of the wireless device 604. Such capability information may include information related to any one of various types of wireless device capabilities. The capability information may also or alternatively be provided from the wireless device to the cellular base station (or vice versa) at any one of various other times and / or in any one of various ways. At least as a possibility, the capability information may include an indication of the machine learning model capabilities of the wireless device from the wireless device 604 to the cellular base station / network node 606, for example, indicating what machine learning model parameters (such as the maximum number of hidden layers and the maximum number of nodes per layer for a neural network type artificial intelligence model) are supported by the wireless device 604.

[0130] In 612, according to some embodiments, the wireless device 604 may request training resources from the network node 606. For example, once the wireless device 604 (e.g., UE) has established a connection with the network node 606 (e.g., base station), the wireless device 604 may need resources from the network to observe or measure the channel for channel estimation purposes or for training a machine learning model. For example, observing the channel state may use measurements of channel state information-reference signals (CSI-RS) sent by the network node 606. In other words, the training resources may include CSI-RS or resources associated with CSI-RS. Additionally, the training resources may include RSs for beam management, selection, etc.

[0131] Thus, in 614, network node 606 may transmit resources (e.g., CSI-RS) such that wireless device 604 can use the CSI-RS to perform channel measurements. In some scenarios, network node 606 may need to reduce transmissions on some resources for wireless device 604 to observe. Thus, according to some embodiments, network node 606 may provide training resources to wireless device 604 upon request. In other words, the process of collecting training samples (e.g., performing measurements) may be initiated by wireless device 604 or by network node 606, which may configure appropriate measurement resources for wireless device 604.

[0132] In 616, upon receiving training resources (e.g., CSI-RS) from network node 606, wireless device 604 may perform its observations or measurements of the channel. For example, CSI-RS may be configured in a resource set for measuring the quality of the downlink channel, and the UE may perform channel measurements periodically and transmit channel state information (CSI) to the BS (e.g., network node 606). Thus, according to some embodiments, the base station may be able to use the CSI to determine adjustments to various parameters during communication with the wireless device. In some embodiments, the measurements may include at least one of CSI measurements or beam scanning measurements. For example, measurements may be performed using one or more receive beams of the UE and one or more transmit beams of the BS. The measurements may include reference signal received power (RSRP), signal-to-noise and interference ratio (SINR), signal-to-noise ratio (SNR), etc.

[0133] Furthermore, in 618 and according to some embodiments, wireless device 604 may accumulate and / or compress the measurements or observations into results (e.g., training samples). In other words, these training samples may be aggregated on wireless device 604. For example, wireless device 604 may first compress and / or aggregate training samples / results / metadata information before transmitting them to a machine learning entity or server, such as model trainer 602.

[0134] Furthermore, compression of the aggregated training samples (e.g., measurement results) and / or metadata may allow wireless device 606 to minimize the number of bits required to transmit them. In other words, through compression, wireless device 606 may be able to encode measurement information using fewer bits than the original representation. Thus, according to some embodiments, the size of the reconstruction or modification of the aggregated measurement information or metadata may be effectively reduced.

[0135] In 620, the wireless device 604 may send the compressed results and / or metadata (e.g., training samples) to a server, such as the model trainer 602 (e.g., the model trainer 510). Additionally, since the sending of training samples may be bandwidth-intensive, it may be beneficial to send the training samples during more suitable conditions. For example, the wireless device 604 may benefit from sending training samples (e.g., measurement results) when connected to external power, when connected to Wi-Fi, under good or high-fidelity signal conditions, during a pause in application activity, or in various other scenarios.

[0136] In some embodiments, the training samples may be accompanied by metadata, such as the date and time of capture (e.g., the date / time of measurement), the location of capture (e.g., the location of measurement), network identification, cell identification, or beam ID, as well as the configuration, model, and software version of the wireless device 604 (and various other metadata). Additionally or alternatively, the metadata may include an assessment by the wireless device itself of its operating environment based on local measurements or sensor data. According to some embodiments, the training samples or results may include information that can be used to generate or update an improved machine learning model, such as a model ID for an encoder / decoder pair (such as the wireless device 604 (encoder) and the network node 606 (decoder)).

[0137] According to some embodiments, various fields may be included as part of the model ID (and as an example, associated with a neural network (NN)). For example, the model ID may include fields such as network (NW) vendor identification, UE vendor identification, public land mobile network (PLMN) ID, use case ID, and / or the number of NNs for the use case. For example, the use case ID may be an identifier associated with observed or predefined channel characteristics. Additionally or alternatively, according to some embodiments, different use cases may be associated with different numbers of NNs based on their complexity or processing requirements. Further, for the model description of the associated model ID, metadata information may also be included. For example, according to some embodiments, the metadata information may indicate the training status (e.g., the trained and tested network) and a potential training dataset indication. Additionally or alternatively, the metadata information may indicate functionality / object, input / output of the machine learning (ML) model, latency benchmark of the ML model, memory requirements of the ML model, accuracy of the ML model, and / or the compression state of the model. Further, the metadata information may indicate the inference or operating conditions, including urban environment, indoor environment, or dense macrocell environment. Additionally or alternatively, the metadata information may indicate preprocessing and / or postprocessing of the measurements for machine learning input / output.

[0138] In some embodiments, network node 606 may associate a label or hash with a measurement resource. For example, network node 606 may configure or associate a label (e.g., an identifier) or a hash value in a CSI-RS resource configuration. According to some embodiments, the label or hash may allow a distributed machine learning entity / server (e.g., as an example, model trainer 602) to understand the measurement conditions through its coordination with a network operator or infrastructure provider (e.g., network vendor), whereby detailed network operation information may not be publicly available. Additionally, in some instances, the network operator or infrastructure provider may share information about network operation conditions through a database that may be shared (e.g., via a commercial agreement) with model trainer 602, and the label and / or hash may provide a key to look up specific network operation information. Additionally or alternatively, model trainer 602 may be configured to aggregate the collected training samples. According to some embodiments, network node 606 may request and receive training samples from wireless device 604.

[0139] In some embodiments, model trainer 602 may be configured to use the training samples received from one or more devices to generate or update (e.g., train) a model associated with a model ID using machine learning techniques. In other words, model trainer 602 may use the training samples provided by wireless device 604 to train an encoder model. In some embodiments, the model IDs may be organized in a way that captures the relationships between the models. For example, after using an initial set of training samples to train an encoder and a decoder, additional samples may be used with the original set to evolve (e.g., update or change) the encoder model while maintaining the same decoder model trained on the initial sample set. One example reason would be to maintain compatibility with the decoder trained at the network based on the initial set, as the network may not have access to the additional training samples. Subsequently, the accumulated training samples may be used to create a new encoder-decoder pair (intended to share the additional samples with the network). Additionally, knowledge of the relationships between the models may allow the devices and the network to determine the best model to use among the compatible models they have access to. In some embodiments, the model relationships may be transmitted to a model server from which the devices may access or request these model relationships.

[0140] Therefore, it may be beneficial for model trainer 602 to associate a model ID with a set of observed statistics. For example, the observed statistics may correspond to one or more device manufacturers, device hardware and software versions, operating environment (channel characteristics), network configuration, and various other characteristics. In some embodiments, the characteristics may be explicitly captured through fields in the model ID or only implicitly captured by including the observations corresponding to the characteristics.

[0141] In addition, the model trainer 602 may associate the received training samples with metadata (which may be included as part of the training samples) by using a clustering algorithm such that the training samples and metadata are arranged into a number of (possibly overlapping) clusters or groups. For example, there may be a cluster that encompasses all the training samples and additional clusters associated with low or high latency spread, low or high signal-to-noise ratio (SNR), and various other possibilities. Additionally or alternatively, an encoder-decoder pair may be trained based on the training samples belonging to each cluster and associated with a model ID (e.g., a model corresponding to the model ID may be generated or updated). According to some embodiments, the model ID may be a number (e.g., a numerical value). In other embodiments, the model ID may include one or more data fields. Additionally or alternatively, the model ID may be characterized as a bit string. As a potential option, multiple encoder-decoder pairs (with associated model IDs) may be trained for one cluster based on different trade-offs between the number of encoded bits and the fidelity of the representation. In other words, some encoder-decoder pair models may be preferentially trained based on the size and quality of the training samples.

[0142] In some embodiments, different clusters may correspond to different operating environments. For example, one cluster may correspond to an indoor environment, whose channel state is significantly different from another cluster corresponding to an outdoor environment. As another example, the clusters may also be associated with different environments, such as a marine environment compared to a mountainous environment. Thus, these clusters will indicate different RF characteristics.

[0143] In addition, an affinity metric model may be associated with each cluster. According to some embodiments, the affinity metric model may vary with the metadata as well as the training samples. For example, a first model associated with model ID "1" may be trained based on samples collected at low speed and high SNR. Additionally or alternatively, a second model associated with model ID "2" may be trained based on samples collected at high speed and corresponding to low SNR. Further, each ID corresponding to index "k" may correspond to a pair of observed characteristics or statistics (e.g., speed k , SNR k ). Thus, in order to select the most appropriate or efficient model when the wireless device is operating at low / high SNR and / or traveling at low / high speed, the device may be able to calculate a metric (e.g., an affinity metric). For example, the device may use the following formula to calculate the metric: α * (current speed - speed k ) 2 + β * (current SNR - SNR k ) 2, where α and β can be coefficients. The index k that minimizes this quantity can be the ID closest to the current operating conditions (e.g., "1" or "2"), and thus can have the greatest affinity or suitability for the determined ID. Additionally, according to this example, and in some embodiments, speed and SNR can be considered metadata rather than the channel measurements themselves.

[0144] In other embodiments, the affinity metric can be calculated based on actual observations. For example, the ID corresponding to index k can be associated with a particular value of the ratio of the largest channel eigenvalue and the Nth largest channel eigenvalue corresponding to the performed observations or measurements. As an example, N can be chosen as the second or third largest channel eigenvalue. Thus, the device can be capable of using its observations or measurements to calculate the eigenvalues of the current channel it is in and calculate the ratio of its largest eigenvalue to the Nth largest eigenvalue. Thus, the device can then pick or select the index k for which the determined or calculated ratio is closest. In other words, the model ID (which can be represented by a number in some instances) can be associated with affinity metrics that can indicate the fidelity or other RF characterization of the model ID via corresponding values or functions. Additionally, according to some embodiments, the wireless device 604 can use the affinity metric model to rank or request the model IDs it has access to based on their suitability (e.g., compatibility and / or enhanced quality) for a particular device and / or network and operating environment.

[0145] In 622, the model trainer 602 can transmit or send the model associated with the model ID and the affinity metric model to another server, such as the model server 600 (e.g., model server 504). In 624, the model trainer 602 can send the training samples associated with the model ID (for the decoder) to a different server, such as the training data server 608 (e.g., training data server 508). Additionally, the training data server 608 can also use a repository (e.g., training data repository 506) to store the said training samples with the model ID (for the decoder). According to some embodiments, the model trainer 602 can choose not to provide training samples that are considered outliers to the model server 600 or the training data server 608. In some embodiments, the functionality of the model trainer 602 can be distributed across multiple processing servers. Additionally or alternatively, the functionality of the model trainer 602 can also be co-located with any of the other entities, including the UE itself. In other words, according to some embodiments, the model trainer 602 can be co-located with the model server 600, the wireless device 604, or the training data server 608. Additionally or alternatively, according to some embodiments, the training data server 608 can be co-located with the model server 600.

[0146] Thus, in 626, the wireless device 604 can potentially rank or request one or more model IDs or associated models to which it has access based on the affinity metrics or measurements (e.g., training samples) from the model server 600. For example, the training samples collected on the device can correspond to affinity metric models associated with different model IDs. Additionally, according to some embodiments, the model IDs can be ranked based on the resulting affinity metrics. Thus, the wireless device 604 can request model IDs in 626 based on the ranking from the affinity metrics. The model server 600 can store the model IDs and affinity metric models in a model repository such as model repository 502. In 628, the model server 600 can send the corresponding model, model ID, and / or affinity metric to the wireless device 604 in response to the request from the wireless device 604.

[0147] In 630, the wireless device 604 can send an ordered list of model IDs to the network node 606. For example, according to some embodiments, based on the affinity metrics and ordered model IDs (and / or models) received from the model server 600, the wireless device may be able to convey a prioritized list of model IDs to the network, which can indicate its preferred model IDs. In other words, the ordered list can include one or more model IDs arranged in a prioritized order based on the affinity metrics associated with the model IDs. According to some embodiments, the wireless device 604 can determine the order of the model IDs to include in the ordered list.

[0148] In 632, after receiving the ordered list of model IDs from the wireless device 604, the network node 606 can request training samples corresponding to the model IDs in the ordered list from the training data server 608. Additionally, in 624, the training data server 608 may have received training samples from the model trainer 602. Thus, in 634, the training data server 608 can send the requested training samples to the network node 606, which can be used to train a decoder model to be compatible with the model IDs from the ordered list of model IDs. In some embodiments, the network node 606 can request decoder training samples (e.g., training samples for the decoder (network node 606)) from the wireless device 604. Thus, the wireless device 604 can use its current encoder model to generate training samples for the decoder and further send these decoder training samples to the network node 606. Then, the network node can use the decoder training samples to train its decoder model. According to some embodiments, the network node 606 can send an additional request to the training data server 608 after 634 for the purpose of using the received training samples for the model in subsequent communication sessions.

[0149] In 636, the network node 606 may select a model ID for communicating with the wireless device 604 based on the decoder model it has trained and the potential preference for the model ID in the ordered list of model IDs. According to some embodiments, the model ID may correspond to the model IDs associated with both the decoder (e.g., the network node) and the encoder (e.g., the wireless device). In other words, the selected model ID may be applicable or compatible with both the decoder and the encoder (e.g., an encoder-decoder pair). Additionally, according to some embodiments, the network node 606 may provide a response to the wireless device 604 indicating which model ID it has selected. Thus, in 638, the wireless device 604 and the network node 606 may use the selected model ID to perform subsequent communication. According to some embodiments, the wireless device 604 providing an ordered list of model IDs and the network node 606 subsequently selecting a model ID may be characterized as model ID negotiation.

[0150] In some embodiments, the communication between the wireless device 604 and the network node 606 may benefit from the selected model ID associated with the observed statistics. For example, the process discussed above may enable a continuous learning process throughout the device (and network) lifespan. Thus, this may allow device (and network) manufacturers to aggregate in-field observations from all their devices and provide customized models for different devices, configurations, and operating environments. Additionally, according to some embodiments, the encoder / decoder pair (e.g., UE / BS) may further benefit its communication by minimizing the air resources consumed for transmitting training samples, minimizing the on-device storage and processing requirements for training, and / or identifying and eliminating outlier samples.

[0151] As another example and according to some embodiments, a network-priority approach may be utilized to select and / or evolve compatible model IDs for UE / BS communication. More specifically, while Figure 6 a model trained based on the samples transmitted from the wireless device 604 to the model trainer 602 is described, an alternative "network-priority" approach may involve at least some similar steps performed by the network node 606 instead of the wireless device 604. For example, instead of the wireless device 604 sending the training samples (e.g., compressed results / metadata) to the model trainer 602 as shown in 618, the wireless device 604 may send them to the network node 606. Thus, the network node 606 may send them to the model trainer 602. Then, the model trainer 602 may train the encoder-decoder pair and transmit the model ID to the model server and the training samples to the training data server for the wireless device 604 and the network node 606 to appropriately request the training samples from it. According to some embodiments, the wireless device 604 may be configured to request the training samples corresponding to the model ID from the training data server 608.

[0152] According to some embodiments, different model trainers (other than model trainer 602) associated with a UE or UE manufacturer may request training samples from the training data server 608. Thus, different model trainers may be able to use the requested training samples to train an encoder model. Additionally or alternatively, encoder models trained by different model trainers may be uploaded to different model servers associated with the UE or UE manufacturer. Further, according to some embodiments, a UE may request a model and / or an associated model ID from the model server. In other words, according to some embodiments, a UE may utilize more than one model trainer and / or model server to train an encoder model and request a model and / or an associated model ID therefrom.

[0153] It is possible that a machine learning model is trained to use channel state information or beam measurement information (as well as various other measurement information) of one or more cells to infer the best model ID or encoder / decoder pair ID of a cell (or possibly multiple cells). For example, a machine learning model may be used to obtain channel state information or beam measurement information of one or more cells and use that information to identify one or more preferred model IDs of one or more UEs or network nodes operating in the one or more cells.

[0154] In some embodiments, at least according to some embodiments, the cell from which channel state information is obtained and the preferred model ID is inferred using the channel state information may or may not be co-located with another cell (possibly in another frequency range). In other words, at least in some embodiments, the input to the machine learning model may, for example, be used as an identifier of the location and orientation of a wireless device to a sufficient extent based on the training information provided to the machine learning model so as to allow for an effective inference of an effective model ID to be used for cellular communication between a cellular base station and the wireless device via the cell or another cell corresponding to a different frequency range. Thus, embodiments are also contemplated in which the machine learning model may be trained (and used as an input) on one or more other types of information that may be related to the model ID to be used for cellular communication between a cellular base station and a wireless device via a cell in a different frequency range, for example, as additional or alternative information to the channel state information for a cell in a first frequency range.

[0155] Thus, at least according to some embodiments, Figure 6 the method of may be used to provide a framework by which the model ID selection of a wireless device may be performed with the assistance of machine learning-based techniques and, thus, at least in some instances, potentially reduce wireless device power consumption and / or improve network resource usage efficiency.

[0156] Figures 7 to 9 and additional information

[0157] Figures 7 to 9 Illustrates additional aspects that can be combined, if desired, with the Figure 6 method. More specifically, Figures 7 to 9 illustrates example aspects of various possible methods for performing enhanced distributed machine learning model maintenance in a wireless communication system according to some embodiments. However, it should be noted that the exemplary details illustrated and described with respect to these figures are not intended to limit the present disclosure as a whole: Many variations and alternatives of the details provided herein are possible and should be considered within the scope of the present disclosure. Figures 7 to 9 within.

[0158] Figure 7 Illustrates an example relationship between a training set and an encoder / decoder pair according to some embodiments. Model IDs can be organized in a way that captures the relationships between the models. For example, after training an encoder and a decoder using an initial training sample set, additional samples can be used with the original set to evolve the encoder model while keeping the same decoder model trained on the initial training sample set. One intention in doing so may be to maintain compatibility with the decoder trained on the initial set at the network (since the network may not have access to the additional samples). Subsequently, the cumulative or additional samples can be used to create a new encoder-decoder pair (with the intention of sharing the additional samples with the network). Knowledge of the relationships between the models can allow the device and the network to determine the best model to use among the compatible models to which they have access.

[0159] According to some embodiments, if additional training samples become available after an encoder-decoder pair has been trained for a certain model ID, the decoder can remain fixed and the encoder can be incrementally retrained to benefit from the additional training samples. This can enable continuous improvement of the encoder model on the device without communicating with a network node. Additionally or alternatively, the user equipment 512 can be capable of incrementally retraining the encoder / decoder pair together for a certain model ID and further generating an additional incremental decoder training data set. The user equipment can then transmit the incremental decoder training data set with the model ID (e.g., an updated model ID version) to the network node to also fine-tune (e.g., update) its decoder.

[0160] As an example, as Figure 7As shown, the encoder 1704 can evolve or update its model ID using the training set 1702 (using training resources requested from the decoder 1706), while maintaining compatibility with the decoder 1706. According to some embodiments, the encoder 1704 can continue to evolve / update its model ID through additional training 710 (e.g., additional training resources) and the associated training set 2708, such that the updated encoder / decoder pair and model ID correspond to an encoder 2712 that has maintained its compatibility with the decoder 1706. Additionally or alternatively, the encoder / decoder pair can continue with even further additional training corresponding to the training sets 1...N714, resulting in an updated encoder N716 and decoder N718 pair.

[0161] Figure 8 An example communication flow diagram between a wireless device 802 and a model server 804 is illustrated. More specifically, Figure 8 An example of how the device 802 can request a model from the model server 804 based on a model ID or metadata is illustrated.

[0162] For example, the model server 804 can maintain the models it receives in a repository. Thus, a device such as the device 802 can be able to request an encoder model and an associated affinity metric model from the model server 804. In some embodiments, the device 802 can request a specific model ID from the model server 804. According to some embodiments, the model ID can be provided to the device during an operating system (OS) update, or discovered during a model ID negotiation with the network. Additionally or alternatively, the device 802 can provide metadata to the model server 804 in order to receive a suitable model or set of models upon request. In some embodiments, in addition to potentially requesting decoder compatibility, the device 802 can also query the model server 804 for an upgraded model based on the existing model IDs it has access to. Thus, the model server 804 can use its knowledge of model relationships to provide a suitable model for the device 802. For example, as shown at 808, according to some embodiments, the model server 804 can provide the device 802 with one or more encoder models and the associated model IDs and affinity metrics.

[0163] Figure 9 An example communication flow diagram between a wireless device 902 and a network node 904 is illustrated. More specifically, Figure 9 An example of how the device 902 communicating with the network entity 904 can participate in a model ID negotiation is illustrated.

[0164] According to some embodiments, device 902 may sort or request the model IDs to which it has access based on an affinity metric. For example, training samples collected on device 902 may correspond to an affinity metric model associated with different model IDs. In some embodiments, the model IDs may be sorted based on the resulting affinity metric initially provided to the model trainer. Thus, as part of the model ID negotiation, device 902 may transmit an ordered list of model IDs to network 904 at 906. Additionally or alternatively, device 902 may choose not to include some of the model IDs to which it has access. For example, device 902 may sometimes choose to conserve battery by restricting its own use to models that require fewer computational resources, thus consuming less power. According to some embodiments, device 902 may suppress or not include one or more model IDs in the ordered list based on one or more model IDs being outliers or associated with edge scenarios.

[0165] In 908, according to some embodiments, network 904 may select or pick from the ordered list of model IDs those for which it has a suitable decoder and communicate that selection or pick to the device. As a supplement or alternative to exchanging model IDs, device 902 may also transmit the location (e.g., a Uniform Resource Locator (URL)) of training data samples associated with the selected model IDs. In some embodiments, network 904 may similarly request from device 902 the training data samples (or training data server locations) for the model IDs to which the device has access. Thus, this may allow network 904 to learn decoders for new or updated encoders over time.

[0166] According to some embodiments, Figure 9The model ID negotiation shown may occur at different times. For example, model ID negotiation may occur when a connection is established, when the operating environment changes, after a tracking area update, and / or when a data or application session changes, as well as in various other scenarios. Additionally or alternatively, rather than the device 902 transmitting the location of the training data sample associated with the model ID to the network 904, the network 904 may request such location from a server (e.g., a training data server) that stores a list of model IDs and corresponding training data sample locations. In some embodiments, the device 902 may transmit a timestamp to the network 904 that may indicate that it has all model IDs trained up to that timestamp. Therefore, the network 904 may be able to use the timestamp to retrieve the corresponding model ID list by querying a server such as a training data server. Knowing that the device 902 may prefer newer models over older models, the timestamp may enable the network 904 to explicitly or implicitly recreate an ordered list of model IDs. Additionally or alternatively, the history of the network interacting with several devices using different model IDs can allow the network 904 to form its own preferences between different models by measuring key performance indicators such as throughput, latency, capacity, and various other indicators. Furthermore, according to some embodiments, the network can use these preferences to make its selection or choice during model negotiation.

[0167] The machine learning technique may be operable to identify the best encoder / decoder pair or model ID by generating training samples using measurement information of one or more cells of the environment, and further creating or updating the model ID using a more efficient communication configuration. In some embodiments, the measurement information may be for a cell that is co-located or not co-located with the cell providing the measurement. For example, it may be the case that the UE may operate using carrier aggregation and / or dual connectivity mode (e.g., to potentially have links with cells in multiple frequency ranges). For dual connectivity mode, it may be the case that the two nodes may perform some coordination of information related to the measurements and / or model IDs.

[0168] Still another set of embodiments may include a cellular base station comprising: one or more processors; and a memory having instructions stored thereon that, when executed by the one or more processors, perform the steps of a method according to one of the foregoing examples.

[0169] Yet another set of embodiments may include a computer program product comprising computer instructions that, when executed by one or more processors, perform the steps of the method described in any of the preceding examples.

[0170] Yet another example embodiment may include a method comprising: performing, by a wireless device, any or all of the foregoing examples.

[0171] Another exemplary embodiment may include an apparatus that includes: an antenna; radio components coupled to the antenna; and a processing element operatively coupled to the radio components, wherein the apparatus is configured to implement any or all parts of the foregoing examples.

[0172] Another set of exemplary embodiments may include a non-transitory computer-accessible memory medium that includes program instructions that, when executed at a device, cause the device to implement any or all parts of any one of the foregoing examples.

[0173] Yet another set of exemplary embodiments may include a computer program that includes instructions for performing any or all parts of any one of the foregoing examples.

[0174] Still another set of exemplary embodiments may include an apparatus that includes means for performing any or all elements of any one of the foregoing examples.

[0175] A further exemplary collection of embodiments may include an apparatus that includes a processing element configured to cause a wireless device to perform any or all elements of any one of the foregoing examples.

[0176] It is well known that the use of personally identifiable information should follow privacy policies and practices that are generally recognized as meeting or exceeding industry or government requirements for maintaining user privacy. Specifically, personally identifiable information data should be managed and processed to minimize the risk of inadvertent or unauthorized access or use, and the nature of authorized use should be clearly explained to users.

[0177] By interpreting each message / signal X received by a user equipment (UE) in the downlink as a message / signal X sent by a base station, and interpreting each message / signal Y sent by the UE in the uplink as a message / signal Y received by the base station, any one of the methods for operating a UE described herein may form the basis for a corresponding method for operating a base station.

[0178] The embodiments of the present disclosure may be implemented in any form of a variety of forms. For example, in some embodiments, the subject matter may be implemented as a computer-implemented method, a computer-readable memory medium, or a computer system. In other embodiments, the subject matter may be implemented using one or more custom-designed hardware devices such as an ASIC. In other embodiments, the subject matter may be implemented using one or more programmable hardware elements such as an FPGA.

[0179] In some embodiments, a non-transitory computer-readable memory medium (e.g., a non-transitory memory element) may be configured to store program instructions and / or data, where if the computer system executes these program instructions, the computer system is caused to execute a method, such as any of the method embodiments described herein, or any combination of the method embodiments described herein, or any subset of any of the method embodiments described herein, or any combination of such subsets.

[0180] In some embodiments, a device (e.g., a UE) may be configured to include a processor (or a set of processors) and a memory medium (or a memory element), where the memory medium stores program instructions, where the processor is configured to read and execute these program instructions from the memory medium, where these program instructions are executable to implement any of the various method embodiments described herein (or any combination of the method embodiments described herein, or any subset of any of the method embodiments described herein, or any combination of such subsets). The device may be implemented in any of a variety of forms.

[0181] Although the above embodiments have been described in considerable detail, many variations and modifications will become apparent to those skilled in the art once the above disclosure is fully understood. It is intended that the following claims be interpreted to cover all such variations and modifications.

Claims

1. An apparatus, comprising: at least one processor, the at least one processor being configured to cause a user equipment (UE) to: receive one or more reference signals from a network node; use the one or more reference signals to perform one or more measurements; compress the one or more measurements into one or more measurement results; send the one or more measurement results to a server; request from the server at least one of one or more identifiers (IDs) or one or more models associated with the one or more IDs; receive from the server the at least one of the one or more IDs or the one or more models, wherein the one or more IDs or the one or more models are provided based on the one or more measurement results; send an ordered list of the one or more IDs to the network node; receive from the network node a response indicating a selection of an ID among the one or more IDs; and communicate with the network node using the ID.

2. The apparatus according to claim 1, wherein the one or more reference signals are channel state information-reference signals (CSI-RS).

3. The apparatus according to claim 1, wherein the UE is configured to send the one or more measurement results to the server when operating according to at least one of the following conditions: when connected to external power; when connected to Wi-Fi; when operating under high-fidelity signal conditions; or during a pause in application activity.

4. The apparatus according to claim 1, wherein the one or more measurement results include metadata corresponding to the one or more IDs, and wherein the metadata indicates at least one of the following: the training status of the one or more IDs; the functionality, object, input, or output of the one or more IDs; the latency benchmark, memory requirement, or accuracy of the one or more IDs; the compression status of the one or more IDs; the inference or operating conditions of the one or more IDs; or the preprocessing and postprocessing information of the one or more measurements.

5. The apparatus according to claim 1, wherein the one or more measurements include at least one of the following: one or more channel state information (CSI) measurements; or one or more beam scanning measurements.

6. The apparatus according to claim 1, wherein the server is a machine learning model trainer co-located with a machine learning model server.

7. The apparatus according to claim 1, wherein the ordered list of the one or more IDs is arranged in a priority order based on one or more affinity metrics associated with the one or more IDs.

8. The apparatus according to claim 1, further comprising: a radio component, the radio component being operatively coupled to the at least one processor.

9. An apparatus, comprising: at least one processor, the at least one processor being configured to cause a network node to: receive from a user equipment (UE) a request for one or more training resources; send the one or more training resources to the UE; Receive an ordered list of one or more model identifiers (IDs) from the UE; Request one or more training samples associated with the one or more model IDs from a server; Receive the one or more training samples associated with the one or more model IDs from the server; Select a model ID from the one or more model IDs based at least in part on the one or more training samples; Send a response indicating the model ID to the UE; and Communicate with the UE using the model ID.

10. The apparatus according to claim 9, wherein the one or more training samples include metadata information corresponding to at least one of the following: Capture date; Capture time; Capture location; Network identifier; Cell identifier; Beam configuration and identifier; Device model and software version; or An assessment of the operating environment of the UE based on local measurements and sensors of the UE.

11. The apparatus according to claim 9, wherein the one or more model IDs include information corresponding to at least one of the following: Network provider identifier; UE provider identifier; Public Land Mobile Network (PLMN) ID; Use case ID; or The number of neural networks for one or more use cases.

12. The apparatus according to claim 9, wherein the UE and the network node operate as an encoder-decoder pair, respectively.

13. The apparatus according to claim 9, wherein the at least one processor is further configured to cause the network node to: Associate at least one of a tag or a hash value with the one or more training resources.

14. The apparatus according to claim 13, wherein the at least one of the tag or the hash value indicates the measurement conditions of the one or more training samples.

15. The apparatus according to claim 9, wherein the at least one processor is further configured to cause the network node to: Request the one or more training samples associated with the one or more model IDs from the UE.

16. A network node, comprising: At least one processor, the at least one processor being configured to cause the network node to: Receive a request for one or more reference signals from a user equipment (UE); Send the one or more reference signals to the UE; Receive one or more compressed measurement results from the UE; Send the one or more compressed measurement results to a first server; Receive an ordered list of one or more model IDs from the UE; Request one or more training samples corresponding to the one or more model IDs from a second server; Receive the one or more training samples from the second server; Select a model ID from the one or more model IDs based at least in part on the one or more training samples; Send a response indicating the model ID to the UE; and Communicate with the UE using the model ID.

17. The network node according to claim 16, wherein the one or more reference signals are channel state information-reference signals (CSI-RS).

18. The network node according to claim 16, wherein the one or more compressed measurement results include metadata indicating at least one of the following: The training status of the one or more model IDs; The functionality, object, input, or output of the one or more model IDs; The latency benchmark, memory requirement, or accuracy of the one or more model IDs; The compression status of the one or more model IDs; The inference or operating conditions of the one or more model IDs; Or The preprocessing and postprocessing information of the one or more compressed measurement results.

19. The network node according to claim 16, wherein the one or more compressed measurement results are based on at least one of the following: One or more channel state information (CSI) measurements; or One or more beam sweep measurements.

20. The network node according to claim 16, wherein the ordered list of the one or more model IDs is arranged in a priority order based on one or more affinity metrics.

Images