Reports on the inference accuracy of quantified models used for monitoring the performance of artificial intelligence or machine learning.
By quantifying the inference accuracy level of AI/ML models through UE reports, the problem of network nodes having difficulty monitoring the accuracy of different UE models is solved, thus improving the efficiency and accuracy of AI/ML performance monitoring.
Patent Information
- Application Number
- CN202480090474.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-02
- Publication Date
- 2026-11-10
AI Technical Summary
In wireless communication, network nodes struggle to efficiently monitor the inference accuracy of quantified artificial intelligence or machine learning (AI/ML) models deployed across different user equipment (UEs), resulting in inefficient AI/ML performance monitoring.
The UE reports the inference accuracy level of the quantized AI/ML model to the network node, and the network node monitors the AI/ML performance based on this indication, thereby gaining an understanding of the accuracy level of the quantized AI/ML model used by different UEs.
It improves the efficiency of network nodes in performing AI/ML performance monitoring, ensures the appropriateness and accuracy of monitoring parameters, and avoids overly aggressive or conservative monitoring.
Smart Images

Figure CN123077035A_ABST
Abstract
Description
Technical Field
[0001] All aspects of this disclosure relate to wireless communication in general, and more specifically to techniques, apparatus, and methods for reporting the inference accuracy of quantified models for monitoring the performance of artificial intelligence or machine learning. Background Technology
[0002] Wireless communication systems are widely deployed to provide a variety of services, including voice, text, messaging, video, data, and / or other services. Services may include unicast, multicast, and / or broadcast services, etc. Typical wireless communication systems employ multiple access radio access technologies (RATs) capable of supporting communication with multiple users by sharing available system resources (e.g., time-domain resources, frequency-domain resources, spatial-domain resources, and / or device transmit power, etc.). Examples of such multiple access RATs include Code Division Multiple Access (CDMA) systems, Time Division Multiple Access (TDMA) systems, Frequency Division Multiple Access (FDMA) systems, Orthogonal Frequency Division Multiple Access (OFDMA) systems, Single-Carrier Frequency Division Multiple Access (SC-FDMA) systems, and Time Division Synchronous Code Division Multiple Access (TD-SCDMA) systems.
[0003] The aforementioned Multiple Access RATs have been adopted in various telecommunications standards to provide a common protocol enabling different wireless communication devices to communicate at the city, national, regional, or global level. An example telecommunications standard is New Radio (NR). NR (also known as 5G) is part of the continuous evolution of mobile broadband announced by the 3rd Generation Partnership Project (3GPP). NR (and other mobile broadband evolutions beyond NR) can be designed to better support the Internet of Things (IoT) and reduced-capacity device deployments, industrial connectivity, millimeter-wave (mmWave) expansion, licensed and unlicensed spectrum access, non-terrestrial network (NTN) deployments, sidelinks and other device-to-device direct communication technologies (e.g., cellular vehicle-to-everything (CV2X) communications), massive MIMO, decomposed network architectures and network topology expansion, multi-subscriber implementations, high-precision positioning and / or radio frequency (RF) sensing, and more. As the demand for mobile broadband access continues to grow, further improvements to NR can be implemented, and other radio access technologies (such as 6G) can be introduced to further advance mobile broadband evolution. Summary of the Invention
[0004] Some aspects described herein relate to a user equipment (UE) for wireless communication. The UE may include one or more memories and one or more processors coupled to the one or more memories. The one or more processors may be configured to send to a network node an indication of an inference accuracy level associated with a quantized AI / ML model for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model for the AI / ML task. The one or more processors may be configured to receive signaling associated with AI / ML performance monitoring from the network node based at least in part on this indication of the inference accuracy level.
[0005] Some aspects described herein relate to a network node for wireless communication. The network node may include one or more memories and one or more processors coupled to the one or more memories. The one or more processors may be configured to receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The one or more processors may be configured to send signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0006] Some aspects described herein relate to a server device for wireless communication. The server device may include one or more memories and one or more processors coupled to the one or more memories. The one or more processors may be configured to send a quantized AI / ML model for an AI / ML task to a UE, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task. The one or more processors may be configured to send an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0007] Some aspects described herein relate to a method for wireless communication performed by a UE. The method may include: sending to a network node an indication of an inference accuracy level associated with a quantized AI / ML model used for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The method may also include: receiving signaling associated with AI / ML performance monitoring from the network node based at least in part on the indication of the inference accuracy level.
[0008] Some aspects described herein relate to a method for wireless communication performed by a network node. The method may include: receiving an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The method may also include: sending signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0009] Some aspects described herein relate to a method of wireless communication performed by a server device. The method may include: transmitting to a UE a quantized AI / ML model for an AI / ML task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task. The method may also include: transmitting an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0010] Some aspects described herein relate to a non-transitory computer-readable medium storing a set of instructions for wireless communication by a UE. When executed by one or more processors of the UE, the set of instructions enables the UE to: send an indication to a network node of an inference accuracy level associated with a quantized AI / ML model for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model for the AI / ML task. When executed by one or more processors of the UE, the set of instructions enables the UE to: receive signaling associated with AI / ML performance monitoring from the network node, at least in part, based on the indication of the inference accuracy level.
[0011] Some aspects described herein relate to a non-transitory computer-readable medium storing a set of instructions for wireless communication by a network node. When executed by one or more processors of the network node, the set of instructions enables the network node to: receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. When executed by one or more processors of the network node, the set of instructions enables the network node to: send signaling associated with AI / ML performance monitoring to the UE, at least in part, based on the indication of the inference accuracy level.
[0012] Some aspects described herein relate to a non-transitory computer-readable medium storing a set of instructions for wireless communication by a server device. When executed by one or more processors of the server device, the set of instructions enables the server device to: send a quantized AI / ML model for an AI / ML task to a UE, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task. When executed by one or more processors of the server device, the set of instructions enables the server device to: send an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0013] Some aspects described herein relate to an apparatus for wireless communication. The apparatus may include: means for transmitting to a network node an indication of an inference accuracy level associated with a quantized AI / ML model for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model for the AI / ML task. The apparatus may also include: means for receiving from the network node signaling associated with AI / ML performance monitoring, at least in part based on the indication of the inference accuracy level.
[0014] Some aspects described herein relate to an apparatus for wireless communication. The apparatus may include: means for receiving an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The apparatus may also include: means for sending signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0015] Some aspects described herein relate to an apparatus for wireless communication. The apparatus may include: means for transmitting to a UE a quantized AI / ML model for an AI / ML task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task. The apparatus may also include: means for transmitting an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0016] Various aspects of this disclosure may be implemented or be implemented as described in whole by or embodied in the methods, apparatus, systems, computer program products, non-transitory computer-readable media, user equipment, base stations, network nodes, network entities, wireless communication devices and / or processing systems as fully described in the specification and drawings and illustrated in the specification and drawings.
[0017] The preceding paragraphs of this section have broadly summarized some aspects of this disclosure. These and additional aspects and their associated advantages will be described below. The disclosed aspects can serve as the basis for modifying or designing other aspects for performing the same or similar purposes of this disclosure. Such equivalent aspects do not depart from the scope of the appended claims. The characteristics of the aspects disclosed herein, their organization and operation, and their associated advantages will be better understood from the following description taken in conjunction with the accompanying drawings.
[0018] While aspects and implementations have been described in this application by way of example, those skilled in the art will understand that additional specific implementations and use cases can arise in many different arrangements and scenarios. The innovations described herein can be implemented across many different platform types, devices, systems, shapes, sizes, and package arrangements. For example, implementations and / or uses can arise via integrated chip implementations and other non-module-based devices (e.g., end-user devices, vehicles, communication devices, computing devices, industrial equipment, retail / purchasing devices, medical devices, AI-enabled devices, etc.). While some examples may or may not specifically point to a use case or application, the applicability of various types of the described innovations is evident. The scope of implementations can range from chip-level or module-level components to non-module, non-chip-level implementations, and further to aggregated, distributed, or original equipment manufacturer (OEM) devices or systems incorporating one or more aspects of the described innovations. In some practical settings, devices incorporating the described aspects and features may also include additional components and features for implementing and practicing the implementations protected and described in the claims. For example, the transmission and reception of wireless signals necessarily involve multiple components for analog and digital purposes (e.g., hardware components including antennas, RF chains, power amplifiers, modulators, buffers, processors, interleavers, adders / summers, etc.). The innovations described herein are intended to be implemented in a variety of devices, chip-level components, systems, distributed arrangements, end-user equipment, etc., with different sizes, shapes, and constructions. Attached Figure Description
[0019] The accompanying drawings illustrate some aspects of this disclosure but do not limit its scope, as other aspects can be achieved by this description. Each drawing in the drawings is provided for illustrative and descriptive purposes and not as a definition of limitation of the claims. Identical or similar reference numerals in different drawings may identify identical or similar elements.
[0020] Figure 1 This is a diagram illustrating an example of a wireless communication network according to the present disclosure.
[0021] Figure 2 This is a diagram illustrating communication between an example network node and an example user equipment (UE) in a wireless network according to the present disclosure.
[0022] Figure 3 This is a diagram illustrating an example decomposed base station architecture according to this disclosure.
[0023] Figure 4 This is a diagram illustrating an example architecture of a functional framework for intelligent radio access networks implemented by data collection, according to the present disclosure.
[0024] Figure 5 This is a diagram illustrating an example of beam management based on artificial intelligence / machine learning (AI / ML) according to this disclosure.
[0025] Figure 6 This is a diagram illustrating an example AI / ML architecture for communication between a first wireless device and a second wireless device according to this disclosure.
[0026] Figures 7 to 8 This is an illustration of an example associated with a report on the inference accuracy of a quantified model used for AI / ML performance monitoring, according to this disclosure.
[0027] Figure 9 This is a diagram illustrating an example process performed, for example, at the UE or a device of the UE, according to this disclosure.
[0028] Figure 10 This is a diagram illustrating an example process performed, for example, at a network node or a device of a network node, according to the present disclosure.
[0029] Figure 11 This is a diagram illustrating an example process performed, for example, at a server device or a device of a server device, according to this disclosure.
[0030] Figures 12 to 14 This is a diagram illustrating an example device for wireless communication according to the present disclosure. Detailed Implementation
[0031] Various aspects of this disclosure are described below with reference to the accompanying drawings. However, aspects of this disclosure may be embodied in many different forms and should not be construed as limited to any specific aspect illustrated or described with reference to the drawings or otherwise presented in this disclosure. Rather, these aspects are provided so that this disclosure will be comprehensive and complete, and will fully convey the scope of protection of this disclosure to those skilled in the art. Those skilled in the art will understand that the scope of this disclosure is intended to cover any aspect of this disclosure disclosed herein, whether implemented independently of or in combination with any other aspect of this disclosure. For example, various combinations or numbers of aspects set forth herein may be used to implement an apparatus or practice. Furthermore, the scope of this disclosure is intended to cover apparatuses having structures and / or functionalities other than those available for practicing the various aspects of this disclosure set forth herein, or methods practiced using these other structures and / or functionalities. Any aspect of this disclosure disclosed herein may be embodied by one or more elements of the claims.
[0032] Various methods, operations, apparatuses, and techniques will now be presented with reference to them. These methods, operations, apparatuses, and techniques will be described in detail below and illustrated in the accompanying drawings by various blocks, modules, components, circuits, steps, processes, or algorithms (collectively, “elements”). These elements may be implemented using hardware, software, or a combination of hardware and software. Whether these elements are implemented as hardware or software depends on the specific application and the design constraints imposed on the system as a whole.
[0033] Some aspects and techniques described herein can be implemented, at least in part, using artificial intelligence (AI) programs, such as those including machine learning (ML) or artificial neural network (ANN) models. Example AI or ML (AI / ML) models may include mathematical representations or definitions of computational capabilities for inferring from input data based on patterns or relationships identified in the input data. As used herein, the term "inference" may include one or more of decision, prediction, determination, or value, which may represent the output of an AI / ML model. Computational capabilities may be defined based on certain parameters of the AI / ML model, such as weights and biases. Weights may indicate a relationship between certain input data and certain outputs of the AI / ML model, and biases are offsets that may indicate the starting point of the AI / ML model's output. An example AI / ML model operating on input data may begin with an initial output based on the biases and then update its output based on a combination of input data and weights.
[0034] AI / ML models can be deployed in one or more devices (e.g., network nodes and / or user equipment (UE)) and can be configured to enhance various aspects of wireless communication systems. For example, AI / ML models can be trained to identify patterns or relationships in data corresponding to networks, devices, or air interfaces. AI / ML models can support operational decisions involving one or more aspects associated with wireless communication devices, networks, or services. For example, AI / ML models can be used to support or improve aspects such as signal decoding / decoding, network routing, energy saving, transceiver circuit control, frequency synchronization, timing synchronization, channel state estimation, channel equalization, channel state feedback, modulation, demodulation, device location, beamforming, load balancing, operation and management functions, security, etc.
[0035] In some examples, an AI / ML model can be trained for an AI / ML task at a network node or server device (e.g., a model server) and delivered to one or more UEs so that the UEs can use the AI / ML model to perform inference (e.g., prediction) for the AI / ML task. To improve efficiency (e.g., to enable the AI / ML model to be deployed at the UE), the floating-point AI / ML model trained for the AI / ML task can be quantized to produce a quantized AI / ML model for the AI / ML task. "Quantization" is a model size reduction technique that reduces the number of bits used to represent model parameters. For example, quantization can convert model weights (and / or other model parameters) from a high-precision (e.g., floating-point) representation to a lower-precision representation (e.g., a fixed-value representation or an integer value representation). Quantization of a model (e.g., a floating-point model) can reduce the model's memory usage and computational complexity.
[0036] In some examples, different UEs may have different UE-specific model quantization preferences or constraints. For instance, different UE AI / ML hardware accelerators may be associated with different quantization preferences or constraints. This could result in different UEs deploying different quantized AI / ML models for the same floating-point AI / ML model used for AI / ML tasks, with these quantized AI / ML models having different inference accuracy performance levels. Network nodes can perform AI / ML performance monitoring for UEs deploying different quantized AI / ML models for AI / ML tasks. For example, network nodes can schedule the transmission of auxiliary reference signals (RS) for prediction accuracy verification and / or configure inference error thresholds / numbers for triggering performance alarm messages, etc. However, network nodes may not be aware of the different model inference accuracy levels of the quantized AI / ML models deployed at different UEs. This could lead to inefficient AI / ML performance monitoring by network nodes, such as using overly aggressive or conservative AI / ML performance monitoring parameters for some UEs.
[0037] The report generally pertains to the reporting of the inference accuracy of quantized AI / ML models. Some aspects more specifically involve reporting the inference accuracy level of quantized AI / ML models deployed at the UE for AI / ML performance monitoring. In some aspects, the UE may send an indication to the network node of the inference accuracy level associated with the quantized AI / ML model used for the AI / ML task. The inference accuracy level may indicate the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The network node may send signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level, and the UE may receive signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level.
[0038] In some aspects, the server device may send a quantized AI / ML model to the UE for use in AI / ML tasks. The quantized AI / ML model may be associated with a floating-point AI / ML model used for the AI / ML tasks. In some aspects, the server device may send an indication to the network node of the inference accuracy level associated with the quantized AI / ML model to be used by the UE in AI / ML tasks. In some examples, the network node may send signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level received from the server device at the network node, and the UE may receive the signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level received from the server device at the network node.
[0039] Specific aspects of the subject matter described in this disclosure can be implemented to achieve one or more of the following potential advantages. In some examples, by reporting from the UE or server device the inference accuracy level of the quantized AI / ML model to be used by the UE, the described techniques can be used to enable network nodes to know the inference accuracy levels of different quantized AI / ML models used by different UEs, which can improve the efficiency of AI / ML performance monitoring performed by the network nodes.
[0040] Multiple access radio access technology (RAT) has been adopted in various telecommunications standards to provide a common protocol that enables wireless communication devices to communicate at the city, enterprise, national, regional, or global level. For example, 5G New Radio (NR) is part of the continuous mobile broadband evolution announced by the 3rd Generation Partnership Project (3GPP). 5G NR supports a variety of technologies and use cases, including enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine-type communication (mMTC), millimeter wave (mmWave) technology, beamforming, network slicing, edge computing, Internet of Things (IoT) connectivity and management, and network function virtualization (NFV).
[0041] As the demand for broadband access increases and as the technologies supported by wireless communication networks evolve, further technological improvements can be adopted or implemented in 5G NR or future RATs (such as 6G) to further advance the evolution of wireless communication for a variety of existing and new use cases and applications. Such technological improvements can be associated with new frequency band extensions, licensed and unlicensed spectrum access, overlapping spectrum use, small cell deployments, non-terrestrial network (NTN) deployments, decomposed network architectures and network topology extensions, device aggregation, advanced duplex communication, sidelinks and other device-to-device direct communication, IoT (including passive or environmental IoT) networks, reduced-capacity (RedCap) UE functionality, industrial connectivity, multi-subscriber implementations, high-precision positioning, radio frequency (RF) sensing and / or AI / ML, and more. These technological improvements can support use cases such as wireless backhaul, wireless data centers, extended reality (XR) and metaverse applications, meta-services for supporting vehicle connectivity, holographic and mixed reality communications, autonomous and collaborative robots, vehicle platooning and collaborative manipulation, sensor networks, posture monitoring, brain-computer interfaces, digital twin applications, asset management, and general coverage applications using off-ground and / or aerial platforms, among others. The methods, operations, apparatuses, and techniques described herein can implement one or more of the foregoing technologies and / or support one or more of the foregoing use cases.
[0042] Figure 1 This is a diagram illustrating an example of a wireless communication network 100 according to the present disclosure. The wireless communication network 100 may be a 5G (or NR) network or a 6G network, or may include elements of a 5G (or NR) network or elements of a 6G network, etc. The wireless communication network 100 may include a plurality of network nodes 110, shown as network node (NN) 110a, network node 110b, network node 110c, and network node 110d. Network nodes 110 may support communication with a plurality of UEs 120 (shown as UE120a, UE 120b, UE 120c, UE 120d, and UE 120e).
[0043] Network nodes 110 and UEs 120 of wireless communication network 100 can communicate using the electromagnetic spectrum, which can be subdivided into various categories, frequency bands, carriers, and / or channels according to frequency or wavelength. For example, devices of wireless communication network 100 can communicate using one or more operating frequency bands. In some aspects, multiple wireless networks 100 can be deployed in a given geographical area. Each wireless communication network 100 can support a specific RAT (which may also be referred to as an air interface) and can operate on one or more carrier frequencies in one or more frequency ranges. Examples of RATs include 4G RATs, 5G / NRRATs, and / or 6G RATs, etc. In some examples, when multiple RATs are deployed in a given geographical area, each RAT in that geographical area can operate on a different frequency to avoid interference with each other.
[0044] Various operating frequency bands have been defined as frequency ranges designated FR1 (410 MHz to 7.125 GHz), FR2 (24.25 GHz to 52.6 GHz), FR3 (7.125 GHz to 24.25 GHz), FR4a or FR4-1 (52.6 GHz to 71 GHz), FR4 (52.6 GHz to 114.25 GHz), and FR5 (114.25 GHz to 300 GHz). Although a portion of FR1 is greater than 6 GHz, in some documents and articles, FR1 is often (interchangeably) referred to as the “sub-6 GHz” band. Similarly, in some documents and articles, FR2 is often (interchangeably) referred to as the “millimeter wave” band, but this is different from the Very High Frequency (EHF) band (30 GHz to 300 GHz) identified as the “millimeter wave” band by the International Telecommunication Union (ITU). The frequencies between FR1 and FR2 are often referred to as the mid-band frequencies, including FR3. Frequency bands falling within FR3 can inherit FR1 or FR2 characteristics, thereby effectively extending the characteristics of FR1 or FR2 into mid-band frequencies. Therefore, "below 6 GHz" (if used herein) can broadly refer to frequencies less than 6 GHz, within FR1, and / or included in mid-band frequencies. Similarly, the term "millimeter wave" (if used herein) can broadly refer to frequencies included in mid-band frequencies, within FR2, FR4, FR4-a, FR4-1, or FR5, and / or within the EHF band. Higher frequency bands can extend 5G NR operation, 6G operation, and / or other RATs above 52.6 GHz. For example, each of FR4a, FR4-1, FR4, and FR5 falls within the EHF band. In some examples, the wireless communication network 100 can implement dynamic spectrum sharing (DSS), where multiple RATs (e.g., 4G / LTE and 5G / NR) are implemented within a single frequency band using dynamic bandwidth allocation (e.g., based on user demand). It is conceivable that the frequencies included in these operating frequency bands (e.g., FR1, FR2, FR3, FR4, FR4-a, FR4-1 and / or FR5) can be modified, and the techniques described herein are applicable to those modified frequency ranges.
[0045] Network node 110 may include one or more devices, components, or systems that enable communication between UE 120 and one or more devices, components, or systems of wireless communication network 100. Network node 110 may be, may include, or may also be referred to as an NR network node, 5G network node, 6G network node, node B, eNB, gNB, access point (AP), transmit / receive point (TRP), mobility element, core, network entity, network element, network equipment, and / or another type of device, component, or system included in a radio access network (RAN).
[0046] Network node 110 may be implemented as a single physical node (e.g., a single physical structure) or as two or more physical nodes (e.g., two or more different physical structures). For example, network node 110 may be a device or system implementing a portion of a radio protocol stack, a device or system implementing a complete radio protocol stack (such as a complete gNB protocol stack), or a collection of devices or systems collectively implementing a complete radio protocol stack. For example, and as shown, network node 110 may be an aggregated network node (with an aggregated architecture), meaning that network node 110 can implement a complete radio protocol stack physically and logically integrated within a single node (e.g., a single physical structure) in the wireless communication network 100. For example, aggregated network node 110 may consist of a single standalone base station or a single TRP that uses the complete radio protocol stack to implement or facilitate communication between UE 120 and the core network of wireless communication network 100.
[0047] Alternatively, and also as shown in the figure, network node 110 can be a decomposed network node (sometimes referred to as a decomposed base station), meaning that network node 110 can realize a radio protocol stack that is physically distributed and / or logically distributed among two or more nodes in the same or different geographical locations. For example, a decomposed network node may have a decomposed architecture. In some deployments, decomposed network node 110 may be used in integrated access and backhaul (IAB) networks, in open radio access networks (O-RAN) (such as network configurations conforming to O-RAN Alliance standards), or in virtualized radio access networks (vRAN) (also referred to as cloud radio access networks (C-RAN)) to facilitate scaling by decomposing base station functionality into multiple units that can be deployed independently.
[0048] Network nodes 110 of the wireless communication network 100 may include one or more central units (CUs), one or more distributed units (DUs), and / or one or more radio units (RUs). CUs may host one or more higher-layer control functions, such as Radio Resource Control (RRC) functions, Packet Data Convergence Protocol (PDCP) functions, and / or Service Data Adaptation Protocol (SDAP) functions, etc. DUs may host one or more of the Radio Link Control (RLC) layer, Media Access Control (MAC) layer, and / or one or more higher physical (PHY) layers, at least in part, according to functional splits (such as functional splits defined by 3GPP). In some examples, DUs may also host one or more lower PHY layer functions, such as Fast Fourier Transform (FFT), Inverse FFT (iFFT), beamforming, Physical Random Access Channel (PRACH) extraction and filtering, and / or scheduling of resources for one or more UEs 120, etc. RUs may host RF processing functions or lower PHY layer functions, such as FFT, iFFT, beamforming, or PRACH extraction and filtering, etc., according to functional splits (such as lower-layer functional splits). In this architecture, each RU can be operated to handle over-the-air (OTA) communications with one or more UE 120s.
[0049] In some aspects, a single network node 110 may include a combination of one or more CUs, one or more DUs, and / or one or more RUs. Additionally or alternatively, network node 110 may include one or more near real-time (near RT) RAN Intelligent Controllers (RICs) and / or one or more non-real-time (non-RT) RICs. In some examples, CUs, DUs, and / or RUs may be implemented as virtual units, such as Virtual Central Units (VCUs), Virtual Distributed Units (VDUs), or Virtual Radio Units (VRUs), etc. Virtual units may be implemented as virtual network functions, such as those associated with cloud deployments.
[0050] Some network nodes 110 (e.g., base stations, RUs, or TRPs) can provide communication coverage for a specific geographic area. In 3GPP, the term "cell" can refer to the coverage area of network node 110 or to network node 110 itself, depending on the context in which the term is used. Network node 110 can support one or more (e.g., three) cells. In some examples, network node 110 can provide communication coverage for macro cells, pico cells, femto cells, or another type of cell. A macro cell can cover a relatively large geographic area (e.g., a radius of several kilometers) and can allow unrestricted access by UE 120 with a service subscription. A pico cell can cover a relatively small geographic area and can allow unrestricted access by UE 120 with a service subscription. A femto cell can cover a relatively small geographic area (e.g., a residential area) and can allow restricted access by UE 120 associated with that femto cell (e.g., UE 120 in a Closed Subscriber Group (CSG)). The network node 110 used for a macro cell can be referred to as a macro network node. Network node 110 used for a picocell may be referred to as a pico network node. Network node 110 used for a femtocell may be referred to as a femto network node or a home network node. In some examples, the cell may not necessarily be stationary. For example, the geographical area of the cell may move depending on the location of the associated mobile network node 110 (e.g., a train, satellite base station, drone, or NTN network node).
[0051] The wireless communication network 100 can be a heterogeneous network, comprising different types of network nodes 110, such as macro network nodes, piconet nodes, femtonet nodes, relay network nodes, aggregation network nodes, and / or decomposition network nodes, etc. Figure 1 In the example shown, network node 110a can be a macro network node for macro cell 130a, network node 110b can be a pico network node for pico cell 130b, and network node 110c can be a femto network node for femto cell 130c. Various types of network nodes 110 can typically transmit at different power levels, serve different coverage areas, and / or have different effects on interference in the wireless communication network 100 compared to other types of network nodes 110. For example, macro network nodes can have high transmit power levels (e.g., 5 watts to 40 watts), while pico network nodes, femto network nodes, and relay network nodes can have lower transmit power levels (e.g., 0.1 watts to 2 watts).
[0052] In some examples, network node 110 may be, may include, or may operate as a RU, TRP, or base station communicating with one or more UEs 120 via a radio access link (which may be referred to as a "Uu" link). The radio access link may include a downlink and an uplink. "Downlink" (or "DL") refers to the communication direction from network node 110 to UE 120, and "uplink" (or "UL") refers to the communication direction from UE 120 to network node 110. Downlink channels may include one or more control channels and one or more data channels. Downlink control channels may be used to transmit downlink control information (DCI) (e.g., scheduling information, reference signals, and / or configuration information) from network node 110 to UE 120. Downlink data channels may be used to transmit downlink data (e.g., user data associated with UE 120) from network node 110 to UE 120. Downlink control channels may include one or more physical downlink control channels (PDCCH), and downlink data channels may include one or more physical downlink shared channels (PDSCH). The uplink channel may similarly include one or more control channels and one or more data channels. The uplink control channel can be used to transmit uplink control information (UCI) from UE 120 to network node 110 (e.g., transmitting corresponding reference signals and / or feedback with one or more downlinks). The uplink data channel can be used to transmit uplink data (e.g., user data associated with UE 120) from UE 120 to network node 110. The uplink control channel may include one or more physical uplink control channels (PUCCH), and the uplink data channel may include one or more physical uplink shared channels (PUSCH). The downlink and uplink may each include a set of resources on which network node 110 and UE 120 can communicate.
[0053] Downlink and uplink resources may include time-domain resources (frames, subframes, time slots, and / or symbols), frequency-domain resources (bands, component carriers, subcarriers, resource blocks, and / or resource elements), and / or spatial-domain resources (specific transmission directions and / or beam parameters). Frequency-domain resources in some bands may be subdivided into bandwidth portions (BWPs). A BWP may be a contiguous block of frequency-domain resources allocated to one or more UEs 120 (e.g., a contiguous block of resource blocks). A UE 120 may be configured with both an uplink BWP and a downlink BWP (where the uplink BWP and downlink BWP may be the same BWP or different BWPs). BWPs may be dynamically configured and / or reconfigured (e.g., by sending DCI configuration to one or more UEs 120 via network node 110), meaning that BWPs may be adjusted in real-time (or near real-time) based on changing network conditions in the wireless communication network 100 and / or based on the specific requirements of one or more UEs 120. This allows for more efficient use of available frequency domain resources in the wireless communication network 100, as fewer frequency domain resources can be allocated to the BWP for UE 120 (which reduces the number of frequency domain resources that UE 120 needs to monitor), thus allowing more frequency domain resources to be distributed across multiple UE 120s. Therefore, the BWP can also assist in the implementation of such UE 120s by facilitating the configuration of smaller bandwidths for communications performed by lower-capacity UE 120s.
[0054] As described above, in some aspects, the wireless communication network 100 may be an IAB network, may include an IAB network, or may be included in an IAB network. In an IAB network, at least one network node 110 is an anchor network node communicating with a core network. The anchor network node 110 may also be referred to as an IAB donor (or "IAB donor"). The anchor network node 110 may be connected to the core network via a wired backhaul link. For example, the Ng interface of the anchor network node 110 may terminate at the core network. Additionally or alternatively, the anchor network node 110 may be connected to one or more devices in the core network that provide core access and mobility management functions (AMF). An IAB network typically also includes multiple non-anchor network nodes 110, which may also be referred to as relay network nodes or simply IAB nodes (or "IAB-nodes"). Each non-anchor network node 110 can directly communicate with the anchor network node 110 via a wireless backhaul link to access the core network, or can indirectly communicate with the anchor network node 110 via one or more other non-anchor network nodes 110 and an associated wireless backhaul link forming a backhaul path to the core network. Some anchor network nodes 110 or other non-anchor network nodes 110 can also directly communicate with one or more UEs 120 via a wireless access link carrying access services. In some examples, network resources used for wireless communication (such as time resources, frequency resources, and / or spatial resources) can be shared between the access link and the backhaul link.
[0055] In some examples, any network node 110 relaying communication may be referred to as a relay network node, a relay station, or simply a repeater. A repeater may receive communications from an upstream station (e.g., another network node 110 or UE 120) and transmit communications to a downstream station (e.g., UE 120 or another network node 110). In this case, the wireless communication network 100 may include or be referred to as a "multi-hop network." Figure 1 In the example shown, network node 110d (e.g., a relay network node) can communicate with network node 110a (e.g., a macro network node) and UE 120d to facilitate communication between network node 110a and UE 120d. Additionally or alternatively, UE 120 can be a relay station capable of relaying transmissions to or from other UE 120s, or can operate as such a relay station. UE 120 relaying communication can be referred to as a UE repeater or relay UE, etc.
[0056] UE 120 may be physically distributed throughout the wireless communication network 100, and each UE 120 may be stationary or mobile. UE 120 may be, may include, or may be included in the following: access terminal, another terminal, mobile station, or subscriber unit. UE 120 may be, or may include, a cellular phone (e.g., a smartphone), a personal digital assistant (PDA), a wireless modem, a wireless communication device, a handheld device, a laptop computer, a cordless phone, a wireless local loop (WLL) station, a tablet computer, a camera, a gaming device, a netbook, a smartbook, an ultrabook, a medical device, a biometric device, a wearable device (e.g., a smartwatch, smart clothing, smart glasses, a smart wristband and / or smart jewelry (such as a smart ring or smart bracelet)), an entertainment device (e.g., a music device, a video device and / or a satellite radio), an XR device, a vehicle component or sensor, a smart meter or sensor, industrial manufacturing equipment, a Global Navigation Satellite System (GNSS) device (such as a Global Positioning System device or another type of positioning device), a UE function of a network node, and / or any other suitable device or function that can communicate via a wireless medium, or may be coupled to them.
[0057] UE 120 and / or network node 110 may include one or more chips, system-on-a-chip (SoC), chipsets, packages, or devices that individually or collectively constitute or include a processing system. The processing system includes processor (or “processing”) circuitry in the form of one or more processors, microprocessors, processing units (such as central processing units (CPUs), graphics processing units (GPUs), neural processing units (NPUs), and / or digital signal processors (DSPs)), processing blocks, application-specific integrated circuits (ASICs), programmable logic devices (PLDs) (such as field-programmable gate arrays (FPGAs)), or other discrete gate or transistor logic components or circuits (all of which are generally referred to herein individually as “processors” or collectively as “processors” or “processor circuitry”). One or more of these processors may be individually or collectively configured to perform the various functions or operations described herein. A processor group that can be configured or configured to perform a set of functions may include a first processor that can be configured or configured to perform a first function in the set, and a second processor that can be configured or configured to perform a second function in the set, or may include the entire processor group that is configured or can be configured to perform the set of functions.
[0058] The processing system may also include memory circuitry in the form of one or more memory devices, memory blocks, memory elements, or other discrete gate or transistor logic components or circuits, each of which may include tangible storage media such as random access memory (RAM) or read-only memory (ROM) or combinations thereof (all of which are generally referred to herein individually as "memory" or collectively as "memory" or "memory circuitry"). One or more of these memories may be coupled to one or more processors in the processor (e.g., operatively coupled, communicatively coupled, electronically coupled, or electrically coupled) and may store processor-executable code (such as software) individually or collectively, which, when executed by one or more processors in the processor, may configure one or more processors in the processor to perform the various functions or operations described herein. Additionally or alternatively, in some examples, one or more processors in the processor may be pre-configured to perform the various functions or operations described herein without being configured by software. The processing system may also include or be coupled to one or more modems (such as Wi-Fi (e.g., IEEE compliant) modems or cellular (e.g., 3GPP 4G LTE, 5G, or 6G compliant) modems). In some embodiments, one or more processors of the processing system include or implement one or more modems among the modems. The processing system may also include, or be coupled to, multiple radio components (collectively, “radio components”), multiple RF chains, or multiple transceivers, each of which may in turn be coupled to one or more antennas among multiple antennas. In some embodiments, one or more processors of the processing system include or implement one or more of the radio components, RF chains, or transceivers. UE 120 may include, or may be included in, a housing that accommodates components associated with UE 120, including the processing system.
[0059] Some UEs 120 may be considered Machine Type Communication (MTC) UEs, Evolved or Enhanced Machine Type Communication (eMTC) UEs, Further Enhanced eMTC (feMTC) UEs, or Enhanced feMTC (efeMTC) UEs, or further evolutions thereof, all of which may be collectively referred to as "MTC UEs". MTC UEs may be, may include, or may be included in or coupled with the following: robots, unmanned aerial vehicles, remote devices, sensors, instruments, monitors, and / or location tags. Some UEs 120 may be considered IoT devices and / or may be implemented as NB-IoT (Narrowband IoT) devices. IoT UEs or NB-IoT devices may be, may include, or may be included in or coupled with the following: industrial machines, appliances, refrigerators, doorbell camera devices, home automation devices, and / or lighting fixtures, etc. Some UEs 120 may be considered customer premises equipment, which may include telecommunications equipment installed at a customer location (such as a home or office) to enable access to a service provider’s network (such as being included in or communicating with the wireless communication network 100).
[0060] Some UEs 120 can be categorized according to different categories associated with varying levels of complexity and / or capabilities. UEs 120 in the first category facilitate large-scale IoT within the wireless communication network 100 and offer lower complexity and / or cost compared to UEs 120 in the second category. UEs 120 in the second category may include mission-critical IoT devices capable of URLLC, enhanced mobile broadband (eMBB), and / or precise positioning within the wireless communication network 100, legacy UEs, baseline UEs, high-level UEs, advanced UEs, full-capability UEs, and / or premium UEs. UEs 120 in the third category may have intermediate-level complexity and / or capabilities (e.g., capabilities between first-category UEs 120 and second-capability UEs 120). UEs 120 in the third category may be referred to as reduced-capability UEs (“RedCap UEs”), intermediate-level UEs, NR lightweight UEs, and / or NR-Lite UEs, etc. RedCap UEs bridge the gap in capabilities and complexity between NB-IoT devices and / or eMTC UEs and mission-critical IoT devices and / or premium UEs. RedCap UEs can include, for example, wearable devices, IoT devices, industrial sensors, and / or cameras associated with limited bandwidth, power capacity, and / or transmission range. RedCap UEs can support healthcare environments, building automation, power distribution, process automation, transportation and logistics, and / or smart city deployments, among others.
[0061] In some examples, two or more UEs 120 (e.g., shown as UE 120a and UE 120e) can communicate directly with each other using sidelink communication (e.g., without communication through a network node 110 acting as an intermediary). As an example, UE 120a can directly send data, control information, or other signaling to UE 120e as sidelink communication. This contrasts with, for example, UE 120a first sending data to network node 110 in UL communication, and then that network node sending data to UE 120e in DL communication. In various examples, UE 120 can use peer-to-peer (P2P) communication protocols, device-to-device (D2D) communication protocols, vehicle-to-everything (V2X) communication protocols (which may include vehicle-to-vehicle (V2V) protocols, vehicle-to-infrastructure (V2I) protocols, and / or vehicle-to-pedestrian (V2P) protocols), and / or mesh network communication protocols to send and receive sidelink communication. In some deployments and configurations, network node 110 may schedule and / or allocate resources for sidelink communication between UEs 120 in the wireless communication network 100. In some other deployments and configurations, UE 120 (instead of network node 110) may perform or cooperate with or negotiate with one or more other UEs to perform scheduling operations, resource selection operations, and / or other operations for sidelink communication.
[0062] In various examples, in addition to half-duplex operation, some network nodes and UEs in the wireless communication network 100, including network node 110 and UE 120, can also be configured for full-duplex operation. Network node 110 or UE 120 operating in half-duplex mode can perform only one of transmission or reception during a specific time resource period (such as a specific time slot, symbol, or other time period). Half-duplex operation may involve time division duplex (TDD), where the DL transmission of network node 110 and the UL transmission of UE 120 do not occur in the same time resource (i.e., the transmissions do not overlap in time). In contrast, network node 110 or UE 120 operating in full-duplex mode can transmit and receive communications concurrently (e.g., within the same time resource). By operating in full-duplex mode, network node 110 and / or UE 120 can generally increase the capacity of the network and radio access links. In some examples, full-duplex operation may involve frequency division duplex (FDD), in which network node 110 performs DL transmission in a first frequency band or on a first component carrier, and UE 120 performs transmission in a second frequency band or on a second component carrier, the second frequency band or the second component carrier being different from the first frequency band or the first component carrier, respectively. In some examples, full-duplex operation may be enabled for UE 120 but not for network node 110. For example, UE 120 may simultaneously transmit UL to the first network node 110 and receive DL transmissions from the second network node 110 in the same time resources. In some other examples, full-duplex operation may be enabled for network node 110 but not for UE 120. For example, network node 110 may simultaneously transmit DL to the first UE 120 and receive UL transmissions from the second UE 120 in the same time resources. In some other examples, full-duplex operation may be enabled for both network node 110 and UE 120.
[0063] In some examples, UE 120 and network node 110 can perform MIMO communication. "MIMO" generally refers to the simultaneous transmission or reception of multiple signals (such as multiple layers or multiple data streams) using the same time and frequency resources. MIMO techniques typically utilize multipath propagation. MIMO can be implemented using various spatial processing or spatial multiplexing operations. In some examples, MIMO can support simultaneous transmission to multiple receivers, which is called multi-user MIMO (MU-MIMO). Some RATs can employ advanced MIMO techniques such as mTRP operations (including redundant transmission or reception on multiple TRPs), reciprocity in the time or frequency domain, single-frequency network (SFN) transmission, or noncoherent joint transmission (NC-JT).
[0064] In some aspects, UE 120 may include a communication manager 140. As described in more detail elsewhere herein, the communication manager 140 may send to a network node an indication of an inference accuracy level associated with a quantized AI / ML model used for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and receive from the network node signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level. Additionally or alternatively, the communication manager 140 may perform one or more other operations described herein.
[0065] In some aspects, network node 110 may include communication manager 150. As described in more detail elsewhere herein, communication manager 150 may receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by the UE for AI / ML tasks, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML tasks; and send signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0066] In some respects, as described in more detail elsewhere herein, the communication manager 150 may send to the UE a quantized AI / ML model for an AI / ML task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task; and send an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model. Additionally or alternatively, the communication manager 150 may perform one or more other operations described herein.
[0067] As indicated above, Figure 1 This is provided as an example. Other examples are available with reference to [the relevant information]. Figure 1 The examples described are different.
[0068] Figure 2 This is a diagram illustrating communication between an example network node 110 and an example UE 120 in a wireless network according to the present disclosure.
[0069] like Figure 2As shown, network node 110 may include a data source 212, a transmit processor 214, a transmit (TX) MIMO processor 216, a set of modems 232 (shown as 232a to 232t, where t≥1), a set of antennas 234 (shown as 234a to 234v, where v≥1), a MIMO detector 236, a receive processor 238, a data sink 239, a controller / processor 240, a memory 242, a communication unit 244, a scheduler 246, and / or a communication manager 150, etc. In some configurations, one or a combination of antennas 234, modems 232, MIMO detectors 236, receive processors 238, transmit processors 214, and / or TX MIMO processors 216 may be included in the transceiver of network node 110. The transceiver may be under the control of and used by one or more processors (such as controller / processor 240), and in some respects, may perform aspects of the methods, procedures and / or operations described herein in conjunction with processor-readable code stored in memory 242. In some respects, network node 110 may include one or more interfaces, communication components and / or other components that facilitate communication with UE 120 or another network node.
[0070] The terms “processor,” “controller,” or “controller / processor” can refer to one or more controllers and / or one or more processors. For example, references to “processor,” “controller / processor,” etc. (in the singular) should be understood as referring to a combination of… Figure 2 The processor described refers to any one or more processors, such as a single processor or a combination of multiple different processors. The reference to "one or more processors" should be understood as a combination of references. Figure 2 Any one or more processors described herein. For example, one or more processors of network node 110 may include transmit processor 214, TX MIMO processor 216, MIMO detector 236, receive processor 238, and / or controller / processor 240. Similarly, one or more processors of UE 120 may include MIMO detector 256, receive processor 258, transmit processor 264, TX MIMO processor 266, and / or controller / processor 280.
[0071] In some aspects, a single processor can perform all operations described as being performed by one or more processors. In some aspects, a first set of one or more processors can perform a first operation described as being performed by that one or more processors, and a second set of one or more processors can perform a second operation described as being performed by that one or more processors. The first set of processors and the second set of processors can be the same set of processors or can be different sets of processors. The reference to "one or more memories" should be understood to refer to any one or more memories of the corresponding device, such as combined... Figure 2 The memory described. For example, an operation described as being performed by one or more memories can be performed by the same subset of the one or more memories or by different subsets of the one or more memories.
[0072] For downlink communication from network node 110 to UE 120, transmitting processor 214 may receive data (“downlink data”) intended for use by UE 120 (or a set of UEs including UE 120) from data source 212 (such as a data pipeline or data queue). In some examples, transmitting processor 214 may select one or more MCSs for UE 120 based on one or more Channel Quality Indicators (CQIs) received from UE 120. Network node 110 may process the data (e.g., including encoding the data) based on the MCS selected for UE 120 for transmission to UE 120 on the downlink, thereby generating data symbols. Transmitting processor 214 may process system information (e.g., semi-static resource partitioning information (SRPI)) and / or control information (e.g., CQI requests, grants, and / or upper-layer signaling) and provide overhead symbols and / or control symbols. The transmitting processor 214 can generate reference symbols for reference signals (e.g., cell-specific reference signals (CRS), demodulation reference signals (DMRS), or channel state information (CSI) reference signals (CSI-RS)) and / or synchronization signals (e.g., primary synchronization signal (PSS) or secondary synchronization signal (SSS)).
[0073] The TX MIMO processor 216 can perform space processing (e.g., pre-decoding) on data symbols, control symbols, overhead symbols, and / or reference symbols where applicable, and can output a set of symbol streams (e.g., TA set of output symbol streams is provided to modem 232. For example, each output symbol stream may be provided to a corresponding modulator component (shown as MOD) of modem 232. Each modem 232 may use the corresponding modulator component to process (e.g., modulate) the corresponding output symbol stream (e.g., for orthogonal frequency division multiplexing (OFDM)) to obtain an output sample stream. Each modem 232 may further use the corresponding modulator component to process (e.g., convert to analog, amplify, filter, and / or upconvert) the output sample stream to obtain a time-domain downlink signal. Modems 232a to 232t may transmit the downlink signal set (e.g., ...) together via the corresponding set of antennas 234. T (One downlink signal).
[0074] Downlink signals may include DCI communication, MAC control element (MAC-CE) communication, RRC communication, downlink reference signals, or another type of downlink communication. Downlink signals may be transmitted on the PDCCH, PDSCH, and / or on another downlink channel. Downlink signals may carry one or more transport blocks (TBs) of data. A TB may be a data unit transmitted via the air interface in the wireless communication network 100. A data stream (e.g., from data source 212) may be encoded into multiple TBs for transmission via the air interface. The number of TBs used to carry data associated with a particular data stream may be associated with a TB size shared by multiple TBs. The TB size may be based on the radio channel conditions of the air interface, the MCS used to encode the data, downlink resources allocated for transmitting data, and / or other parameters, or otherwise associated with them. Generally, a larger TB size allows for a larger amount of data to be transmitted in a single transmission, reducing signaling overhead. However, a larger TB size may be more prone to transmission and / or reception errors than a smaller TB size, but such errors can be mitigated through more robust error correction techniques.
[0075] For uplink communication from UE 120 to network node 110, the uplink signal from UE 120 may be received by antenna 234, processed by modem 232 (e.g., demodulator component of modem 232, shown as DEMOD), detected by MIMO detector 236 (e.g., receive (Rx) MIMO processor) where applicable, and / or further processed by receive processor 238 to obtain decoded data and / or control information. Receive processor 238 may provide the decoded data to data sink 239 (which may be a data pipeline, data queue, and / or another type of data sink) and provide the decoded control information to processors such as controller / processor 240.
[0076] Network node 110 may use scheduler 246 to schedule one or more UEs 120 for downlink or uplink communication. In some aspects, scheduler 246 may use DCI to dynamically schedule DL transmissions to and / or UL transmissions from UE 120. In some examples, scheduler 246 may allocate repetitive time-domain and / or frequency-domain resources that UE 120 may use to transmit and / or receive communication using RRC configuration (e.g., semi-static configuration), for example, to perform semi-persistent scheduling (SPS) or to configure configuration grant (CG) for UE 120.
[0077] One or more of the following may be included in the RF chain of network node 110: transmit processor 214, TX MIMO processor 216, modem 232, antenna 234, MIMO detector 236, receive processor 238, and / or controller / processor 240. The RF chain may include one or more filters, mixers, oscillators, amplifiers, analog-to-digital converters (ADCs), and / or other devices for converting between analog signals (such as those used for transmission or reception via an air interface) and digital signals (such as those used for processing by one or more processors of network node 110). In some aspects, the RF chain may be a transceiver of network node 110, or may be included in such a transceiver.
[0078] In some examples, network node 110 may use communication unit 244 to communicate with the core network and / or other network nodes. Communication unit 244 may support wired and / or wireless communication protocols and / or connections, such as Ethernet, fiber optic, Common Public Radio Interface (CPRI), and / or wired or wireless backhaul, etc. Network node 110 may use communication unit 244 to send and / or receive data associated with UE 120, or to perform network control signaling, etc. Communication unit 244 may include transceivers and / or interfaces, such as network interfaces.
[0079] UE 120 may include a collection of antennas 252 (shown as antennas 252a to 252r, where r ≥ 1), a collection of modems 254 (shown as modems 254a to 254u, where u ≥ 1), a MIMO detector 256, a receive processor 258, a data sink 260, a data source 262, a transmit processor 264, a TX MIMO processor 266, a controller / processor 280, a memory 282, and / or a communication manager 140, etc. One or more components of UE 120 may be included in housing 284. In some aspects, one or a combination of antenna 252, modem 254, MIMO detector 256, receive processor 258, transmit processor 264, or TX MIMO processor 266 may be included in a transceiver included in UE 120. The transceiver may be under the control of and used by one or more processors (such as controller / processor 280), and in some respects, may perform aspects of the methods, procedures, or operations described herein in conjunction with processor-readable code stored in memory 282. In some respects, UE 120 may include another interface, another communication component, and / or another component that facilitates communication with network node 110 and / or another UE 120.
[0080] For downlink communication from network node 110 to UE 120, the set of antennas 252 can receive downlink communication or signals from network node 110, and can transmit the set of received downlink signals (e.g., R Each received signal is provided to a set of modems 254. For example, each received signal may be provided to a corresponding demodulator component (shown as DEMOD) of modem 254. Each modem 254 may use the corresponding demodulator component to condition (e.g., filter, amplify, down-convert, and / or digitize) the received signal to obtain an input sample. Each modem 254 may use the corresponding demodulator component to further demodulate or process the input sample (e.g., for OFDM) to obtain a received symbol. MIMO detector 256 may obtain the received symbols from the set of modems 254, perform MIMO detection on the received symbols where applicable, and provide the detected symbols. Receiver processor 258 may process (e.g., decode) the detected symbols, provide the decoded data for UE 120 to data sink 260 (which may include data pipelines, data queues, and / or applications executed on UE 120), and provide the decoded control information and system information to controller / processor 280.
[0081] For uplink communication from UE 120 to network node 110, the transmitting processor 264 may receive and process data (“uplink data”) from data source 262 (such as data pipelines, data queues, and / or applications running on UE 120) and control information from controller / processor 280. The control information may include one or more parameters, feedback, one or more signal measurements, and / or other types of control information. In some aspects, the receiving processor 258 and / or controller / processor 280 may determine one or more parameters related to the transmission of uplink communication for received signals (such as those received from network node 110 or another UE). One or more parameters may include a Reference Signal Received Power (RSRP) parameter, a Received Signal Strength Indicator (RSSI) parameter, a Reference Signal Received Quality (RSRQ) parameter, a CQI parameter, or a Transmit Power Control (TPC) parameter, etc. The control information may include indications of the RSRP parameter, RSSI parameter, RSRQ parameter, CQI parameter, TPC parameter, and / or another parameter. Control information can facilitate parameter selection and / or scheduling for UE 120 by network node 110.
[0082] Transmit processor 264 can generate reference symbols for one or more reference signals, such as uplink DMRS, uplink sounding reference signal (SRS), and / or another type of reference signal. Symbols from transmit processor 264 can be pre-decoded by TX MIMO processor 266 where applicable, and further processed by a set of modems 254 (e.g., for DFT-s-OFDM or CP-OFDM). TX MIMO processor 266 can perform spatial processing (e.g., pre-decoding) on data symbols, control symbols, overhead symbols, and / or reference symbols where applicable, and can output a set of symbol streams (e.g., ... U A set of output symbol streams is provided to modem 254. For example, each output symbol stream may be provided to a corresponding modulator component (shown as MOD) of modem 254. Each modem 254 may use the corresponding modulator component to process (e.g., modulate) the corresponding output symbol stream (e.g., for OFDM) to obtain an output sample stream. Each modem 254 may further use the corresponding modulator component to process (e.g., convert to analog, amplify, filter, and / or upconvert) the output sample stream to obtain an uplink signal.
[0083] Modems 254a to 254u can transmit uplink signal sets (e.g., via corresponding sets of antennas 252) R One uplink signal or UUplink signals may include UCI communication, MAC-CE communication, RRC communication, or another type of uplink communication. Uplink signals may be transmitted on PUSCH, PUCCH, and / or another type of uplink channel. Uplink signals may carry one or more TBs of data. Sidelink data and control transmission (i.e., transmission directly between two or more UEs 120) may typically use techniques similar to those described for uplink data and control transmission, and may use sidelink-specific channels such as the Physical Sidelink Shared Channel (PSSCH), Physical Sidelink Control Channel (PSCCH), and / or Physical Sidelink Feedback Channel (PSFCH).
[0084] One or more antennas in the set of antennas 252 or the set of antennas 234 may include one or more antenna panels, one or more antenna groups, one or more sets of antenna elements, or one or more antenna arrays, etc., or may be included in one or more antenna panels, one or more antenna groups, one or more sets of antenna elements, or one or more antenna arrays, etc. Antenna panels, antenna groups, sets of antenna elements, or antenna arrays may include one or more antenna elements (within a single housing or multiple housings), a set of coplanar antenna elements, a set of non-coplanar antenna elements, or with one or more transmitting or receiving components (such as...) Figure 2 An antenna module is a device that integrates one or more antenna elements (such as one or more components of an antenna). As used herein, "antenna" can refer to one or more antennas, one or more antenna panels, one or more antenna groups, one or more collections of antenna elements, or one or more antenna arrays. "Antenna panel" can refer to an array or panel of antenna elements arranged in an array or panel that can facilitate beamforming by manipulating the parameters of the antenna group. "Antenna module" can refer to a circuit that includes one or more antennas, and may also include one or more other components (such as filters, amplifiers, or processors) associated with integrating the antenna module into a wireless communication device.
[0085] In some examples, each antenna element of antenna 234 or antenna 252 may include one or more sub-elements for radiating or receiving radio frequency signals. For example, a single antenna element may include a first sub-element cross-polarized with a second sub-element that can be used to independently transmit the cross-polarized signal. Antenna elements may include patch antennas, dipole antennas, and / or other types of antennas arranged in a linear pattern, a two-dimensional pattern, or another pattern. The spacing between antenna elements can allow signals transmitted individually by antenna elements at desired wavelengths to interact or interfere with each other constructively and destructively in various directions (such as to form a desired beam). For example, given a desired wavelength or frequency range, the spacing may provide a quarter wavelength, half a wavelength, or another fraction of the wavelength between adjacent antenna elements to allow for desired constructive and destructive interference patterns of signals transmitted by individual antenna elements within that desired range.
[0086] The amplitude and / or phase of signals transmitted via antenna elements and / or sub-elements can be modulated and (e.g., by manipulating phase shifts, phase offsets, and / or amplitudes) shifted relative to each other to generate one or more beams; this is known as beamforming. The term "beam" can refer to the directional transmission of a wireless signal toward a receiving device or otherwise in a desired direction. "Beam" can also generally refer to the direction associated with such directional signal transmission, the set of directional resources associated with the signal transmission (e.g., angle of arrival, horizontal direction, and / or vertical direction), and / or a set of parameters indicating one or more aspects of the directional signal, the direction associated with the signal, and / or the set of directional resources associated with the signal. In some implementations, antenna elements can be individually selected or deselected for the directional transmission of a signal (or multiple signals) by controlling the amplitude of one or more corresponding amplifiers and / or the phase of the signal to form one or more beams. The shape of the beam (such as amplitude, width, and / or the presence of sidelobes) and / or the direction of the beam (such as the angle of the beam relative to the surface of the antenna array) can be dynamically controlled by modifying the phase shifts, phase offsets, and / or amplitudes of multiple signals relative to each other.
[0087] Different UEs 120 or network nodes 110 may include different numbers of antenna elements. For example, UE 120 may include a single antenna element, two antenna elements, four antenna elements, eight antenna elements, or different numbers of antenna elements. As another example, network node 110 may include eight antenna elements, 24 antenna elements, 64 antenna elements, 128 antenna elements, or different numbers of antenna elements. Generally speaking, a larger number of antenna elements provides increased control over the parameters used for beamforming compared to a smaller number of antenna elements, while a smaller number of antenna elements may be less complex to implement and can use less power. Multiple antenna elements can support multi-layer transmission, in which the same time and frequency resources are used to utilize spatial multiplexing to transmit a first layer of communication (which may include a first data stream) and a second layer of communication (which may include a second data stream).
[0088] Although Figure 2 The boxes in the diagram are illustrated as different components, but the functions described above with respect to these boxes may be implemented in a single hardware, software, or combined component, or in various combinations of components. For example, the functions described with respect to transmit processor 264, receive processor 258, and / or TX MIMO processor 266 may be performed by or under the control of controller / processor 280.
[0089] Figure 3 This is an illustration of an example disaggregated base station architecture 300 according to the present disclosure. One or more components of the example disaggregated base station architecture 300 may be one or more network nodes (such as one or more network nodes 110), may include, or may be included in one or more network nodes. The disaggregated base station architecture 300 may include a CU 310, which may communicate directly with the core network 320 via a backhaul link, or may communicate indirectly with the core network 320 via one or more disaggregated control units (such as non-RT RIC 350 and / or near-RT RIC 370 associated with a Service Management and Orchestration (SMO) framework 360 (e.g., via an E2 link)). The CU 310 may communicate with one or more DU 330 via a corresponding midhaul link (such as via an F1 interface). Each DU 330 may communicate with one or more RU 340 via a corresponding fronthaul link. Each RU 340 may communicate with one or more UE 120 via a corresponding RF access link. In some deployments, UE 120 can be served by multiple RU 340s simultaneously.
[0090] Each component of the disassembled base station architecture 300 (including CU 310, DU 330, RU 340, near-RT RIC 370, non-RT RIC 350, and SMO frame 360) may include one or more interfaces or may be coupled to one or more interfaces for receiving or transmitting signals, such as data or information, via wired or wireless transmission media.
[0091] In some respects, the CU 310 can be logically divided into one or more CU user plane (CU-UP) units and one or more CU control plane (CU-CP) units. When implemented in an O-RAN configuration, the CU-UP units can communicate bidirectionally with the CU-CP units via an interface such as an E1 interface. The CU 310 can be deployed to communicate with one or more DU 330s for network control and signaling, as needed. Each DU 330 may correspond to a logical unit that includes one or more base station functions for controlling the operation of one or more RU 340s. For example, the DU 330 may host various layers, such as the RLC layer, MAC layer, or one or more PHY layers (such as one or more high PHY layers or one or more low PHY layers). Each layer (which may also be referred to as a module) can be implemented using an interface for signaling to other layers (and modules) hosted by the DU 330, or for signaling to control functions hosted by the CU 310. Each RU 340 may implement lower-layer functionality. In some respects, the real-time and non-real-time aspects of communication with the control plane and user plane of the RU 340 can be controlled by the corresponding DU 330.
[0092] The SMO framework 360 supports RAN deployment and provisioning for both non-virtualized and virtualized network elements. For non-virtualized network elements, the SMO framework 360 supports the deployment of dedicated physical resources for RAN coverage requirements, which can be managed via operation and maintenance interfaces such as the O1 interface. For virtualized network elements, the SMO framework 360 can interact with cloud computing platforms such as the Open Cloud (O-Cloud) platform 390 to perform network element lifecycle management (such as instantiating virtualized network elements) via cloud computing platform interfaces such as the O2 interface. Virtualized network elements may include, but are not limited to, CU 310, DU 330, RU 340, non-RT RIC 350, and / or near-RT RIC 370. In some aspects, the SMO framework 360 can communicate with hardware aspects of the 4G RAN, 5G NR RAN, and / or 6G RAN (such as the Open eNB (O-eNB) 380) via the O1 interface. Additionally or alternatively, the SMO framework 360 can communicate directly with each of one or more RUs 340 via the corresponding O1 interface. In some deployments, this configuration enables each DU 330 and CU 310 to be implemented in a cloud-based RAN architecture, such as a vRAN architecture.
[0093] The non-RT RIC 350 may include or implement logic functions that enable non-real-time control and optimization of RAN elements and resources, including AI / ML workflows for model training and updates, and / or policy-based guidance of applications and / or features in the near-RT RIC 370. The non-RT RIC 350 may be coupled to or communicate with the near-RT RIC 370, such as via an A1 interface. The near-RT RIC 370 may include or implement logic functions that enable near real-time control and optimization of RAN elements and resources via an interface, such as an E2 interface, through data collection and action, connecting one or more CU 310s, one or more DU 330s, and / or O-eNBs to the near-RT RIC 370.
[0094] In some respects, to generate AI / ML models to be deployed in the near-RT RIC 370, the non-RT RIC 350 may receive parameters or external enrichment information from an external server. This information can be utilized by the near-RT RIC 370 and may be received from non-network data sources or network functions at the SMO framework 360 or the non-RT RIC 350. In some examples, the non-RT RIC 350 or near-RT RIC 370 may modulate RAN behavior or performance. For example, the non-RT RIC 350 may monitor long-term trends and patterns of performance and may perform corrective actions using AI / ML models created via the SMO framework 360 (such as reconfiguration via the O1 interface) or via RAN management policies (such as A1 interface policies).
[0095] As indicated above, Figure 3 This is provided as an example. Other examples are available with reference to [the relevant information]. Figure 3 The examples described are different.
[0096] Figure 1 , Figure 2 or Figure 3 Network node 110, its controller / processor 240, UE 120, UE 120's controller / processor 280, CU 310, DU 330, RU 340, or any other component may implement one or more techniques associated with reporting the inference accuracy of the quantized model for AI / ML performance monitoring, or perform one or more operations associated with reporting the inference accuracy of the quantized model for AI / ML performance monitoring, as described in more detail elsewhere herein. For example, network node 110's controller / processor 240, UE 120's controller / processor 280, CU 310, DU 330, RU 340, or any other component may implement one or more techniques associated with reporting the inference accuracy of the quantized model for AI / ML performance monitoring, as described in more detail elsewhere herein. Figure 2 Any other component, CU 310, DU 330, or RU 340 may (alone or in combination with one or more other processors) perform or direct, for example... Figure 9 The process 900 Figure 10 Process 1000 Figure 11 The operation of process 1100 or other processes as described herein. In some aspects, the server device described herein is network node 110, included in network node 110, or comprising... Figure 2One or more components of the network node 110 shown. Memory 242 may store data and program code for network node 110, CU 310, DU 330, or RU 340. Memory 282 may store data and program code for UE 120. In some examples, memory 242 or memory 282 may include a non-transitory computer-readable medium storing a set of instructions (e.g., code or program code) for wireless communication. Memory 242 may include one or more memories, such as a single memory or multiple different memories (of the same or different types). Memory 282 may include one or more memories, such as a single memory or multiple different memories (of the same or different types). For example, the set of instructions may be executed by one or more processors of network node 110, UE 120, CU 310, DU 330, or RU 340 (e.g., directly, or after compilation, transformation, or interpretation). Figure 9 The process 900 Figure 10 Process 1000 Figure 11 The process 1100 or other processes as described herein. In some examples, the execution instructions may include run instructions, transform instructions, compile instructions, and / or interpret instructions, etc.
[0097] In some aspects, the UE (e.g., UE 120) includes: components for transmitting to a network node an indication of an inference accuracy level associated with a quantized AI / ML model for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model for the AI / ML task; and / or components for receiving signaling associated with AI / ML performance monitoring from the network node based at least in part on the indication of the inference accuracy level. Components for the UE to perform the operations described herein may include, for example, one or more of a communication manager 140, an antenna 252, a modem 254, a MIMO detector 256, a receive processor 258, a transmit processor 264, a TX MIMO processor 266, a controller / processor 280, or a memory 282.
[0098] In some aspects, a network node (e.g., network node 110) includes: components for receiving an indication of an inference accuracy level associated with a quantized AI / ML model to be used by the UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and / or components for sending signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level. Components for the network node to perform the operations described herein may include, for example, one or more of the following: a communication manager 150, a transmit processor 214, a TX MIMO processor 216, a modem 232, an antenna 234, a MIMO detector 236, a receive processor 238, a controller / processor 240, a memory 242, or a scheduler 246.
[0099] In some aspects, the server device includes: components for transmitting to the UE a quantized AI / ML model for an AI / ML task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task; and / or components for transmitting an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model. In some aspects, components for the server device to perform the operations described herein may include, for example, one or more of a communication manager 150, a transmit processor 214, a TX MIMO processor 216, a modem 232, an antenna 234, a MIMO detector 236, a receive processor 238, a controller / processor 240, a memory 242, or a scheduler 246.
[0100] As indicated above, Figure 3 This is provided as an example. Other examples are available with reference to [the relevant information]. Figure 3 The examples described are different.
[0101] Figure 4This is a diagram illustrating an example architecture 400 of a functional framework for RAN intelligence enabled by data collection according to this disclosure. In some scenarios, the functional framework for RAN intelligence can be further enhanced by use cases and / or examples through data collection. For example, principles or algorithms for RAN intelligence enabled by AI / ML and associated functional frameworks (e.g., inputs / outputs of optimized AI functionality and / or components for enabling AI) have been utilized or studied to identify the benefits of AI-enabled RAN through possible use cases (e.g., beam management, energy saving, load balancing, mobility management and / or coverage optimization, etc.). In one example, as shown by architecture 400, the functional framework for RAN intelligence may include multiple logical entities such as a model training host 402, a model inference host 404, a data source 406, and an actor 408.
[0102] Model inference host 404 can be configured to run an AI / ML model based on inference data provided by data source 406, and model inference host 404 can utilize inference data input to participant 408 to produce outputs (e.g., predictions). Participant 408 can be an element or entity of the core network or RAN. For example, participant 408 can be a UE, network node, base station (e.g., gNB), CU, DU, and / or RU, etc. Furthermore, participant 408 can also depend on the type of task performed by model inference host 404, the type of inference data provided to model inference host 404, and / or the type of output produced by model inference host 404. For example, if the output from model inference host 404 is associated with beam management, participant 408 can be a UE, DU, or RU. In some examples, model inference host 404 can be hosted on participant 408. For example, a UE can be participant 408 and can host model inference host 404. In some aspects, a UE (e.g., participant 408) can be data source 406. For example, the UE can perform measurements (e.g., NR measurements), input the measurements to the AI / ML model at the model inference host 404 (or provide the measurements to the model inference host 404), and take actions based on the output of the AI / ML model.
[0103] After participant 408 receives output from model inference host 404, participant 408 may determine whether to take action based on that output. For example, if participant 408 is a DU or RU, and the output from model inference host 404 is associated with beam management, participant 408 may determine whether to change / modify the Tx / Rx beam. If participant 408 determines to take action based on that output, in some examples, participant 408 may instruct at least one action subject 410 to take the action. For example, if participant 408 determines to change / modify the Tx / Rx beam used for communication between participant 408 and action subject 410 (e.g., UE), participant 408 may send a beam (re)configuration or beam switching instruction to action subject 410. Participant 408 may modify its Tx / Rx beam based on beam (re)configuration, such as switching to a new Tx / Rx beam or applying different parameters to the Tx / Rx beam, etc. As another example, participant 408 may be a UE, and the output from model inference host 404 may be associated with beam management. For example, the output may be one or more predicted measurements for one or more beams. Participant 408 (e.g., UE) may determine to send a measurement report (e.g., a Layer 1 (L1) RSRP report) to a network node.
[0104] Data source 406 can also be configured to collect data that can be used as training data for training an ML model or as inference data for feeding ML model inference operations. For example, data source 406 may collect data from one or more core network and / or RAN entities (which may include participant 408 or action agent 410) and provide the collected data to model training host 402 for ML model training. In some aspects, model training host 402 may co-located with model inference host 404 and / or participant 408. For example, participant 408 or action agent 410 may provide performance feedback associated with beam configuration to data source 406, which can be used by model training host 402 to monitor or evaluate ML model performance, such as whether the output (e.g., prediction) provided to participant 408 is accurate. In some examples, if the output provided by participant 408 is inaccurate (or below an accuracy threshold), model training host 402 may determine to modify or retrain the ML model used by the model inference host, such as via ML model deployment / update.
[0105] As indicated above, Figure 4 This is provided as an example. Other examples are available with reference to [the relevant information]. Figure 4 The examples described are different.
[0106] Figure 5 This is a diagram illustrating example 500 of AI / ML-based beam management according to this disclosure. Figure 5 As shown, AI / ML model 510 may be deployed at or on UE 120. For example, a model inference host (such as a model inference host) may be deployed at or on UE 120. AI / ML model 510 enables UE 120 to determine one or more inferences or predictions based on data input to AI / ML model 510.
[0107] For example, as indicated by reference numeral 515, the input to AI / ML model 510 may include measurements associated with a first set of beams. For example, network node 110 may transmit one or more signals using a corresponding beam from the first set of beams. UE 120 may perform measurements on the first set of beams (e.g., L1 RSRP measurement, L1 signal-to-interference-plus-noise ratio (SINR) measurement, or other measurements) to obtain a first set of measurements. For example, each beam from the first set of beams may be associated with one or more measurements performed by UE 120. UE 120 may input the first set of measurements (e.g., L1 RSRP measurements or L1 SINR measurements) along with information associated with the first set of beams and / or a second set of beams (such as beam orientation (e.g., spatial orientation), beamwidth, beam shape, and / or other characteristics of the corresponding beams from the first set of beams and / or the second set of beams) into AI / ML model 510.
[0108] As indicated by reference numeral 520 in the attached figure, the AI / ML model 510 may output one or more predictions. These predictions may include predicted measurements associated with a second set of beams (e.g., predicted L1 RSRP measurements or predicted SINR measurements). This reduces the number of beam measurements performed by the UE 120, thereby reducing the power of the UE 120 and / or the network resources of all beams included in the first and second sets of beams that would otherwise be used for measurement. This type of prediction may be referred to as codebook-based spatial domain selection or prediction.
[0109] As another example, the output of AI / ML model 510 may include the point orientation, angle of departure (AoD), and / or angle of arrival (AoA) of the beams included in a second set of beams. This type of prediction may be referred to as non-codebook-based spatial domain selection or prediction. As another example, multiple measurement reports or values collected at different time points may be input into AI / ML model 510. This allows AI / ML model 510 to output codebook-based and / or non-codebook-based predictions for beam measurements, AoD, and / or AoA, etc., at future times. As described herein, the output of AI / ML model 510 can facilitate initial access procedures, secondary cell group (SCG) setup procedures, beam refinement procedures, link quality or interference adaptation procedures, beam failure and / or beam blocking prediction, and / or radio link failure prediction, etc.
[0110] In some examples, a first set of beams (e.g., measured) may be referred to as set B beams, and a second set of beams (e.g., associated with the predicted measurement) may be referred to as set A beams. In some examples, the first set of beams (e.g., set B beams) may be a subset of the second set of beams (e.g., set A beams). In some other examples, the first set of beams and the second set of beams may be different beams and / or may be mutually exclusive sets. For example, the first set of beams (e.g., set B beams) may include wide beams (e.g., unthinned beams or beams with a beamwidth that satisfies a first threshold), and the second set of beams (e.g., set A beams) may include narrow beams (e.g., thinned beams or beams with a beamwidth that satisfies a second threshold). In one example, AI / ML model 510 may perform spatial domain beam prediction for beams included in set A beams based on measurements of beams included in set B beams. As another example, AI / ML model 510 can perform temporal beam prediction for beams included in set A based on historical measurements of beams included in set B.
[0111] As indicated above, Figure 5 This is provided as an example. Other examples are available with reference to [the relevant information]. Figure 5 The examples described are different.
[0112] Figure 6 This is a diagram illustrating an example AI / ML architecture 600 for communication between a first wireless device 602 and a second wireless device 604 according to this disclosure. In some aspects, the first wireless device 602 may be a UE (e.g., UE 120), and the second wireless device 604 may be a network node (e.g., network node 110). It should be noted that the example AI / ML architecture of the first wireless device 602 can be applied to the second wireless device 604, and vice versa.
[0113] The first wireless device 602 may be a chip, a SoC, a chipset, a package, or a device (which includes one or more processors, processing blocks, or processing elements (collectively, "processor 610") and one or more memory blocks or elements (collectively, "memory 620")), or may include the chip, the SoC, the chipset, the package, or the device. The processor 610 may be coupled to a transceiver 640, which includes RF circuitry 642 coupled to an antenna 646 via an interface 644 for transmitting or receiving signals.
[0114] One or more AI / ML models 630 (collectively referred to as "AI / ML models 630") may be stored in memory 620 and accessible by processor 610. Individual AI / ML models or groups of such AI / ML models may be associated with corresponding model identifiers (IDs). In some aspects, different AI / ML models 630 that may optionally be associated with different model identifiers may have different characteristics. One or more AI / ML models 630 may be selected based on corresponding features, characteristics, or applications, as well as characteristics or conditions of the first wireless device 602 (such as power state, mobility state, battery reserve, temperature, etc.). For example, AI / ML models 630 may have different inference data and output pairs (such as different types of inference data that produce different types of outputs), different accuracy levels associated with prediction, different latency associated with producing predictions, different ML model sizes, different coefficients, different parameters, etc.
[0115] Processor 610 can deploy AI / ML model 630 to generate corresponding output data based on input data. As an example, AI / ML model 630 can take measurements of one or more reference signals (such as those corresponding to one or more wide beams) as input to predict channel characteristics (e.g., beam measurements) associated with one or more different reference signals (such as those corresponding to one or more narrow beams within each of one or more wide beams, one or more other wide beams, or one or more narrow beams outside of one or more wide beams, etc.). Input data may include, for example, measurements of one or more reference or pilot signals, such as CQI, signal-to-noise ratio (SNR), SINR, signal-to-noise ratio (SNDR), RSSI, RSRP, RSRQ, and / or block error rate (BLER), etc. Output data may include, for example, compressed CSI feedback or one or more predicted measurements (or characteristics) of one or more reference or pilot signals.
[0116] like Figure 6As shown, the first wireless device 602 and / or the second wireless device 604 can communicate with the model server 650. In some aspects, the model server 650 can perform various AI / ML management tasks for the first wireless device 602 and / or the second wireless device 604. For example, the model server 650 can host various types and / or versions of AI / ML models 630 for download by the first wireless device 602 and / or the second wireless device 604. In some examples, the model server 650 can monitor and evaluate the performance of the AI / ML model 630. Additionally or alternatively, in some examples (e.g., where the first wireless device 602 is a UE and the second wireless device is a network node), the second wireless device 604 can perform AI / ML performance monitoring to monitor and evaluate the performance of the AI / ML model 630 used by the first wireless device 602. In some examples, the model server 650 can send signals or provide indications / instructions to activate or deactivate the use of a specific AI / ML model 630 at the first wireless device 602 or the second wireless device 604. In some examples, model server 650 can switch to different AI / ML models 630 used at first wireless device 602 or second wireless device 604, and model server 650 can provide such instructions to the corresponding first wireless device 602 or second wireless device 604. Model server 650 can act as a model training host (such as in conjunction with...). Figure 4 The model training host 602 is discussed for operation, and the training data is used to update the AI / ML model 630. In some cases, the model server 650 can serve as a data source (such as in conjunction with...). Figure 4 The discussed data source 406 is used to operate, in order to collect and host training data, inference data, performance feedback, etc. associated with the AI / ML model 630.
[0117] In some examples, an AI / ML model (e.g., AI / ML model 630) can be trained at a network node or model server (e.g., model server 650) for an AI / ML task (e.g., beam management / prediction), and this AI / ML model can be delivered to one or more UEs so that one or more UEs can use the AI / ML model to perform inference (e.g., prediction) for the AI / ML task. To improve efficiency (e.g., deploying the AI / ML model at the UE), the floating-point AI / ML model used for the AI / ML task can be quantized to produce a quantized AI / ML model for the AI / ML task. "Quantization" is a model size reduction technique that reduces the number of bits used to represent model parameters. For example, quantization can convert model weights (and / or other model parameters) from a high-precision (e.g., floating-point) representation to a lower-precision representation (e.g., fixed-value or integer value representation). Quantization of the model (e.g., a floating-point model) can reduce the model's memory usage and computational complexity.
[0118] In some examples, different UEs may have different UE-specific model quantization preferences or constraints. For example, different UE AI / ML hardware accelerators may have different quantization preferences or constraints. This could result in different UEs deploying different quantized AI / ML models for the same floating-point AI / ML model used for AI / ML tasks, with these quantized AI / ML models having different inference accuracy performance levels. Examples of quantization preferences or constraints that can be determined for the quantization applied to a floating-point AI / ML model to generate a corresponding quantized AI / ML model may include metrics used to determine the floating-point to fixed-point representation (e.g., clustering, K-means, entropy, uniform, non-uniform, etc.), quantization bit width and scheme (e.g., fixed-point, integer, etc.), component-specific preferences (e.g., whether the activation function should be quantized, preference for fixed-point 16 instead of integer 16 for pooling layers, etc.). In some examples, chipsets from different RF circuit vendors may be associated with different preferences or constraints regarding quantization schemes. In some examples, even for chipsets from the same RF circuit vendor, different original equipment manufacturers (OEMs) may have different custom quantization preferences or constraints, for example, due to specific power or thermal budget considerations. Furthermore, there may be various modules or types of hardware accelerators associated with different quantization preferences or limitations, even from the same RF circuit supplier.
[0119] In some aspects, model server 650 can provide the UE with a quantized UE-side AI / ML model. In some examples, model server 650 can be a third-party server (e.g., a server associated with an entity other than the network node and / or the operator of the wireless communication network to which the UE is connected). For example, a model server (e.g., model server 650) associated with different vendors (e.g., the UE and / or the RF circuit vendor) can provide a quantized UE-side AI / ML model associated with a particular AI / ML task, functionality, feature, or model ID (e.g., an AI / ML task / functionality / feature / model ID associated with a wireless communication standard such as a 3GPP standard). As an example, model server (e.g., model server 650) can store trained floating-point AI / ML models for AI / ML tasks of wide-to-narrow spatial beam prediction (e.g., via L1-RSRP prediction), having model IDs associated with 8 sets of B beams and 32 sets of A beams. The UE can communicate with the model server to report quantization preferences and / or constraints associated with the UE. For example, the model server could be a third-party server (e.g., associated with a vendor of the UE's AI / ML hardware accelerator), and the UE could communicate with the model server, at least in part, based on a proprietary protocol not specified in wireless communication standards (e.g., 3GPP standards). The model server could then prepare a quantized AI / ML model (e.g., by quantizing a floating-point AI / ML model) for the UE to download, according to the UE's quantization preferences and / or constraints. In some aspects, the associated model inference accuracy level of the quantized AI / ML model (e.g., the average L1-RSRP prediction accuracy based on comparisons with a floating-point AI / ML model or with actual measured L1-RSRP, which could also be calculated by the model server) could also be downloaded along with the quantized model.
[0120] Different UEs can deploy different quantized AI / ML models associated with the same AI / ML task, and these different quantized AI / ML models may have different initial model inference accuracy levels (e.g., due to different quantization schemes used for different AI / ML models). Network nodes can perform AI / ML performance monitoring for UEs deploying different quantized AI / ML models for the AI / ML task. For example, a network node can schedule the transmission of auxiliary RS for prediction accuracy verification and / or configure inference error thresholds / numbers for triggering performance alarm messages, etc. However, the network node may not be aware of the different initial model inference accuracy levels of the quantized AI / ML models deployed on different UEs. This can lead to inefficient AI / ML performance monitoring, such as using overly aggressive or conservative AI / ML performance monitoring parameters for some UEs.
[0121] In some aspects described herein, the UE can report to the network node the inference accuracy level associated with the quantized AI / ML model to be used by the UE for AI / ML tasks. In some aspects described herein, a server device (e.g., model server 650) providing the quantized AI / ML model to the UE can report the inference accuracy level associated with the quantized AI / ML model to the network node. In this way, the network node can be aware of the inference accuracy levels of different quantized AI / ML models used by different UEs, which can improve the efficiency of AI / ML performance monitoring performed by the network node.
[0122] As indicated above, Figure 6 This is provided as an example. Other examples may be provided relative to... Figure 6 The examples described are different.
[0123] Figure 7 This is an illustration of example 700 associated with a report on the inference accuracy of a quantified model used for AI / ML performance monitoring, according to this disclosure. Figure 7 As shown, Example 700 includes communication between network node 110, UE 120, and model server 650. In some aspects, model server 650 may be a third-party server associated with an entity other than the operator of network node 110 (e.g., the operator of the wireless communication network associated with network node 110). In some other aspects, model server may be associated with the same operator as network node 110.
[0124] like Figure 7 As shown by reference numeral 705 in the accompanying drawing, UE 120 can obtain a quantized AI / ML model for an AI / ML task. The quantized AI / ML model can be associated with a floating-point AI / ML model for the AI / ML task. For example, a quantized AI / ML model can be generated by applying quantization to a floating-point AI / ML model. That is, the quantized AI / ML model can be a quantized version of a floating-point AI / ML model for the AI / ML task. The quantized AI / ML model can have the same inputs and outputs as the floating-point AI / ML model (e.g., the inputs and outputs associated with the AI / ML task).
[0125] In some respects, AI / ML tasks can be beam measurement or beam management tasks. For example, an AI / ML task can be associated with predicting beam measurements for a set of beams. In such an example, a quantized AI / ML model (and a floating-point AI / ML model) can take beam measurements (e.g., L1-RSRP measurements or other beam measurements) for a first set of beams (e.g., set B beams) as input and can output predicted beam measurements (e.g., predicted L1-RSRP measurements or other beam measurements) for a second set of beams (e.g., set A beams). In another example, an AI / ML task can be associated with predicting a certain number of top beams in a set of beams. In such an example, a quantized AI / ML model (and a floating-point AI / ML model) can take beam measurements (e.g., L1-RSRP measurements or other beam measurements) for a first set of beams (e.g., set B beams) and can output a certain number of beams in a second set of beams (e.g., set A beams) predicted to have the highest / best beam measurements. N (a) beam (e.g., the one predicted to have the highest L1-RSRP measurement) N The identifier of a beam (e.g., beam identifier, RS identifier, etc.). In some other respects, the quantized AI / ML model can be associated with other AI / ML tasks.
[0126] In some aspects, such as Figure 7As shown, UE 120 can obtain (e.g., receive or download) a quantized AI / ML model from model server 650. In such examples, model server 650 can send a quantized AI / ML model for an AI / ML task, and UE 120 can receive a quantized AI / ML model for an AI / ML task. Model server 650 can manage both floating-point AI / ML models and quantized AI / ML models for AI / ML tasks. For example, model server 650 can store floating-point AI / ML models for AI / ML tasks. In some examples, the floating-point AI / ML model and / or AI / ML task (e.g., AI / ML functionality) can be associated with a logical model ID. That is, the logical model ID can identify an AI / ML logical model, which can be defined as a model having (e.g., corresponding to an AI / ML task) certain inputs and outputs. In some aspects, model server 650 can combine communication with UE 120 to generate a quantized AI / ML model. For example, UE 120 may communicate with model server 650 to request a quantized AI / ML model for an AI / ML task, and indicate one or more quantization preferences and / or constraints associated with UE 120. Model server 650 can then generate / prepare a quantized AI / ML model for the AI / ML task by performing quantization of the floating-point AI / ML model according to the one or more quantization preferences and / or constraints indicated by UE 120. In some other aspects, model server 650 may store a floating-point AI / ML model for the AI / ML task and multiple quantized AI / ML models for the AI / ML task, which are generated (e.g., by model server 650) by performing quantization of the floating-point AI / ML model using different quantization schemes and / or parameters. In this case, UE 120 can select a quantized AI / ML model from the multiple quantized AI / ML models stored at model server 650 for the AI / ML task. In some respects, UE 120 can communicate with model server 650 via a protocol transparent to network node 110 and / or a wireless communication standard (e.g., 3GPP standard) used for communication between UE 120 and network node 110 to obtain a quantized AI / ML model.
[0127] although Figure 7The illustration shows UE 120 obtaining a quantized AI / ML model from model server 650. However, in some other aspects, UE 120 may not obtain the quantized AI / ML model from model server 650. In some aspects, UE 120 may obtain the quantized AI / ML model via network node 110. For example, UE 120 may receive the quantized AI / ML model from network node 110 or via network node 110 from another network entity (e.g., a core network device). In some aspects, UE 120 may obtain a quantized AI / ML model for an AI / ML task by performing quantization of a floating-point AI / ML model for the AI / ML task. For example, UE 120 may receive a floating-point AI / ML model from model server 650, network node 110, or another network entity, or UE 120 may train a floating-point AI / ML model for the AI / ML task and perform quantization of the floating-point AI / ML model to generate a quantized AI / ML model.
[0128] like Figure 7 As further shown by reference numeral 710 in the accompanying drawings, UE 120 can send an indication of the inference accuracy level of the quantized AI / ML model, and network node 110 can receive the indication of the inference accuracy level of the quantized AI / ML model. In some aspects, the inference accuracy level can indicate the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for an AI / ML task (e.g., a corresponding floating-point AI / ML model quantized to generate the quantized AI / ML model). In some aspects, UE 120 can send the indication of the inference accuracy level of the quantized AI / ML model before UE 120 utilizes the quantized AI / ML model and / or the AI / ML task is activated (e.g., activated by network node 110). For example, UE 120 can send the indication of the inference accuracy level of the quantized AI / ML model (e.g., together with an indication of the model ID associated with the AI / ML task) as UE capability information during initial access. In some respects, the inference accuracy level can be the initial inference accuracy level of the quantified AI / ML model (e.g., before AI / ML performance monitoring performed by network node 110).
[0129] In some aspects, the definition of an inference accuracy level may be pre-configured / predefined for UE 120 (e.g., specified in a wireless communication standard), or configured / indicated by configuration information transmitted via network node 110 and received by UE 120 (e.g., via RRC signaling, MAC-CE, or DCI). In such examples, different definitions of inference accuracy may be pre-configured / predefined or configured / indicated for different AI / ML tasks (e.g., different AI / ML functional or logical models) (e.g., different measurements of the relative accuracy of a quantized AI / ML model relative to a floating-point AI / ML model). In some aspects, multiple candidate inference accuracy levels may be associated with an AI / ML task, and the indication of an inference accuracy level may indicate a candidate inference accuracy level selected from multiple candidate inference accuracy levels. In some examples, the multiple candidate inference accuracy levels associated with an AI / ML task may be pre-configured / predefined (e.g., specified in a wireless communication standard). In some other examples, network node 110 may send configuration information indicating multiple candidate inference accuracy levels for an AI / ML task, and UE 120 may receive configuration information indicating multiple candidate inference accuracy levels for an AI / ML task. In some aspects, the inference accuracy levels may be different for different model IDs associated with the same AI / ML task (e.g., the same AI / ML functionality). In some examples, multiple candidate inference accuracy levels may be associated with different model IDs, and the selected candidate inference accuracy level (indicated in the inference accuracy level indication) may be associated with the model ID of the quantized AI / ML model. In some examples, multiple candidate inference accuracy levels may include multiple model IDs corresponding to different inference accuracy levels (for the same AI / ML task), and UE 120 may indicate the model ID from the multiple model IDs corresponding to the inference accuracy level of the quantized AI / ML model.
[0130] In some aspects, the inference accuracy level can be based at least in part on the statistical difference between certain model output features provided by a floating-point AI / ML model for the AI / ML task and a quantized AI / ML model to be used by UE 120 for the AI / ML task, and / or the statistical difference between certain model output features (e.g., the true values of certain model output features) provided by a sprite-assisted detector / model for the AI / ML task and a quantized AI / ML model to be used by UE 120 (associated with a dataset (e.g., a reference dataset) including model input and output features provided by the sprite-assisted model) for the AI / ML task. In some aspects, the output values of one or more target output features determined by applying the floating-point AI / ML model for the AI / ML task to the reference dataset can be error values of the true values (e.g., values determined via the sprite-assisted detector) included in the reference dataset of one or more target output features. X Internal. For example, error values. X This can be the error tolerance of a floating-point AI / ML model used for AI / ML tasks, specified in the wireless communication standard. In some examples, X This could be an error tolerance applied to each of one or more target features, an error tolerance for the average error on one or more target output features, an error tolerance for the total error on one or more target output features, or an error tolerance for the maximum error among one or more target output features, etc. In some aspects, the inference accuracy level of a quantized AI / ML model can be the difference between the output values of one or more target output features determined by applying the quantized AI / ML model to a reference dataset and the output values of one or more target features determined by applying a floating-point AI / ML model to a reference dataset. Y For example, difference Y It can be the average difference between the output value of the quantized AI / ML model and the output value of the floating-point AI / ML model on one or more target output features, the total difference between the output value of the quantized AI / ML model and the output value of the floating-point AI / ML model on one or more target output features, the maximum difference between the output value of the quantized AI / ML model and the corresponding output value of the floating-point AI / ML model on one or more target output features, or a vector of differences between the output value of the quantized AI / ML model and the corresponding output value of the floating-point AI / ML model for each of the one or more target output features, etc.
[0131] In some respects, where an AI / ML task is associated with predicting beam measurements (e.g., L1-RSRP measurements) for a set of beams (e.g., set A beams), the inference accuracy level can be based at least in part on the difference between a first beam measurement (e.g., a first L1-RSRP measurement) and a second beam measurement, the first beam measurement being predicted by a quantized AI / ML model for a subset of the beam set (e.g., a subset of set A beams) based on model inputs included in a reference dataset (e.g., beam measurements for set B beams), and the second beam measurement being predicted by a floating-point AI / ML model for a subset of the beam set (e.g., a subset of set A beams) based on model inputs included in a reference dataset. For example, a quantized AI / ML model (and a floating-point AI / ML model) can take L1-RSRP measurements for set B beams (e.g., a wide beam set) as input and output L1-RSRP measurements for set A beams (e.g., a narrow beam set). The predicted L1-RSRP value determined by the floating-point AI / ML model for the reference dataset and the value for the previous dataset. N (For example, in one example,) N =4) Set A beams (e.g., the set A beams with the highest predicted L1-RSRP value) N The L1-RSRP prediction accuracy between the true L1-RSRP values (e.g., values determined via a sprite-assisted detector) of each beam can be ± X Within dB (of which) X (This is an error tolerance that can be specified in the wireless communication standard). It is used to compare the predicted L1-RSRP value output by the floating-point AI / ML model with the true L1-RSRP value. N The set A beams can be real front beams from the reference dataset. N A beam identifier (e.g., via a sprite-assisted detector). N The value can be predefined or preconfigured (e.g., specified in wireless communication standards). In this example, the inference accuracy level of the quantized AI / ML model can be the L1-RSRP value predicted by the quantized AI / ML model for a reference dataset and the value predicted by the floating-point AI / ML model for the previous dataset. N A set of A beams (e.g., the real front beams identified in the reference dataset) N The difference between the predicted L1-RSRP values determined by the reference dataset of a set of A beams. Y For example, quantified AI / ML models targeting the previous... N A set of A beams (e.g., real front beams) N The L1-RSRP value predicted by the set of A beams can be obtained from the floating-point AI / ML model for the previous... N± L1-RSRP values predicted by the A-beam set Y Within dB. For example, Y It can be the front N The average L1-RSRP difference on beam A of each set, the front N The total L1-RSRP difference on beam A of each set, the front N The maximum L1-RSRP difference on beam A of the set or the previous N A vector of L1-RSRP differences for each A-beam. In some respects, an indication of the inference accuracy level of a quantized AI / ML model can be... Y Instructions, or may include instructions on Y Indications. In some examples, indications of the inference accuracy level of a quantized AI / ML model may include... Y Instructions for the model ID associated with the AI / ML task.
[0132] In some respects, when an AI / ML task is associated with a certain number of top beams in a set of predicted beams (e.g., identifying which...), N The set A beam is the front N The inference accuracy level can be based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in a reference dataset. For example, the inference accuracy level can be based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in a reference dataset relative to the percentage of top beams predicted by a floating-point AI / ML model based on model inputs included in the dataset. In this example, the reference accuracy level... Y It can be compared with the predictions made by applying a floating-point AI / ML model to a reference dataset. N Compared to a set of A beams, this method accurately predicts the front beams by applying a quantized AI / ML model to a reference dataset. N The percentage of A-beams in the set. In some respects, this can be indicated in the indication of the inference accuracy level of the quantized AI / ML model. Y That value.
[0133] In some respects, model server 650 can indicate the inference accuracy level to UE 120 (e.g., when UE 120 receives a quantized AI / ML model from model server 650). For example, model server 650 can send a value to the quantized AI / ML model. Y The value is indicated, and the UE 120 can receive the value of the quantized AI / ML model. Y The value is an indication. The model server 650 can provide information about the quantized AI / ML model.Y The value indication is sent to UE 120 along with the quantized AI / ML model. In such an example, model server 650 can determine the inference accuracy level of the quantized AI / ML model by applying the quantized AI / ML model and the floating-point AI / ML model to a reference dataset (e.g., Y The value of . In some other aspects, UE 120 may determine the inference accuracy level of the quantized AI / ML model (e.g., by applying the quantized AI / ML model to a reference dataset and comparing the model output of the quantized AI / ML model for the reference dataset with the model output of a floating-point AI / ML model for the reference dataset (e.g., these model outputs can be received from model server 650) for example. Y (value).
[0134] The reference dataset used to determine inference accuracy may include model inputs and real (e.g., sprite-assisted) model output values. In some aspects, the reference dataset used to determine the level of inference accuracy may be predefined (e.g., specified in a wireless communication standard). For example, a wireless communication standard may specify different reference datasets for different AI / ML tasks (e.g., different AI / ML functionalities). In some aspects, the reference dataset may be indicated via signaling sent by network node 110. In some examples, network node 110 may signal the dataset to one or more participating entities (e.g., a model server) where a quantized AI / ML model can be downloaded by the UE. For example, network node 110 may send the reference dataset, and model server 650 may receive the reference dataset. Additionally or alternatively, network node 110 may configure the reference dataset for UE 120. For example, network node 110 may send the reference dataset, and UE 120 may receive the reference dataset. In some other aspects, one or more third-party entities (e.g., a model server) may signal a reference dataset used by the third-party entity, along with a model ID (e.g., a logical model ID) corresponding to the AI / ML task associated with the reference dataset, to network node 110. For example, model server 650 may send the reference dataset, and network node 110 may receive the reference dataset. In this case, model server 650 may send an indication of the model ID associated with the floating-point AI / ML model and / or AI / ML task along with the reference dataset.
[0135] like Figure 7As further shown by reference numeral 715 in the accompanying drawings, network node 110 may send signaling associated with AI / ML performance monitoring based at least in part on an indication of the inference accuracy level of the quantized AI / ML model, and UE 120 may receive signaling associated with AI / ML performance monitoring based at least in part on an indication of the inference accuracy level of the quantized AI / ML model. In some aspects, network node 110 may determine one or more AI / ML performance monitoring parameters for UE 120 based at least in part on the indicated inference accuracy level of the quantized AI / ML model to be used by UE 120, and network node 110 may send signaling to UE 120 indicating configuration information of one or more AI / ML performance monitoring parameters and / or scheduling information based at least in part on one or more AI / ML performance parameters.
[0136] In some aspects, network node 110 may schedule the transmission of RS (e.g., auxiliary RS) for prediction accuracy verification of UE 120 based at least in part on the inference accuracy level indicated by the quantized AI / ML model. For example, network node 110 may schedule the transmission of RS for prediction accuracy verification to UE 120 more frequently or less frequently (e.g., compared to the default scheduling frequency and / or compared to RS scheduled for other UEs) based at least in part on the inference accuracy level indicated by the quantized AI / ML model. In such examples, signaling associated with AI / ML performance monitoring may include signaling instructions for scheduling RS to be sent to UE 120 and / or signaling for the transmission of RS for prediction accuracy verification from network node 110 to UE 120.
[0137] In some aspects, network node 110 can determine / configure error thresholds for UE 120 to determine whether a performance alert message should be sent to UE 120. For example, network node 110 can determine / configure higher or lower error thresholds for different UEs associated with AI / ML tasks, based at least in part on the corresponding inference accuracy level of the quantized AI / ML model to be used by different UEs. In such examples, signaling associated with AI / ML performance monitoring may include signaling indicating or configuring error thresholds for UE 120 and / or performance alert messages sent to UE 120 based at least in part on error thresholds determined / configured for UE 120.
[0138] As indicated above, Figure 7 This is provided as an example. Other examples may be provided relative to... Figure 7 The examples described are different.
[0139] Figure 8This is an illustration of example 800 associated with a report on the inference accuracy of a quantified model used for AI / ML performance monitoring, according to this disclosure. Figure 8 As shown, Example 800 includes communication between network node 110, UE 120, and model server 650. In some aspects, model server 650 may be a third-party server associated with an entity other than the operator of network node 110 (e.g., the operator of the wireless communication network associated with network node 110). In some other aspects, model server may be associated with the same operator as network node 110.
[0140] like Figure 8 As shown by reference numeral 805 in the accompanying drawings, model server 650 can send a quantized AI / ML model for an AI / ML task, and UE 120 can receive the quantized AI / ML model for the AI / ML task. For example, UE 120 can download the quantized AI / ML model from model server 650. The quantized AI / ML model can be associated with a floating-point AI / ML model for the AI / ML task. For example, the quantized AI / ML model can be generated by applying quantization to the floating-point AI / ML model. That is, the quantized AI / ML model can be a quantized version of the floating-point AI / ML model for the AI / ML task. The quantized AI / ML model can have the same inputs and outputs as the floating-point AI / ML model (e.g., the inputs and outputs associated with the AI / ML task).
[0141] In some respects, AI / ML tasks can be beam measurement or beam management tasks. For example, an AI / ML task can be associated with predicting beam measurements for a set of beams. In such an example, a quantized AI / ML model (and a floating-point AI / ML model) can take beam measurements (e.g., L1-RSRP measurements or other beam measurements) for a first set of beams (e.g., set B beams) as input and can output predicted beam measurements (e.g., predicted L1-RSRP measurements or other beam measurements) for a second set of beams (e.g., set A beams). In another example, an AI / ML task can be associated with predicting a certain number of top beams in a set of beams. In such an example, a quantized AI / ML model (and a floating-point AI / ML model) can take beam measurements (e.g., L1-RSRP measurements or other beam measurements) for a first set of beams (e.g., set B beams) and can output a certain number of beams in a second set of beams (e.g., set A beams) predicted to have the highest / best beam measurements. N (a) beam (e.g., the one predicted to have the highest L1-RSRP measurement) NThe identifier of a beam (e.g., beam identifier, RS identifier, etc.). In some other respects, the quantized AI / ML model can be associated with other AI / ML tasks.
[0142] Model server 650 can manage floating-point AI / ML models and quantized AI / ML models for AI / ML tasks. For example, model server 650 can store floating-point AI / ML models for AI / ML tasks. In some examples, floating-point AI / ML models and / or AI / ML tasks (e.g., AI / ML functionality) can be associated with a logical model ID. That is, a logical model ID can identify an AI / ML logical model, which can be defined as a model having certain inputs and outputs (e.g., corresponding to an AI / ML task). In some aspects, model server 650 can combine communication with UE 120 to generate quantized AI / ML models. For example, UE 120 can communicate with model server 650 to request quantized AI / ML models for AI / ML tasks and indicate one or more quantization preferences and / or constraints associated with UE 120. Then, model server 650 can generate / prepare quantized AI / ML models for AI / ML tasks by performing quantization of floating-point AI / ML models according to one or more quantization preferences and / or restrictions indicated by UE 120. In some other aspects, model server 650 can store floating-point AI / ML models for AI / ML tasks and multiple quantized AI / ML models for AI / ML tasks, which are generated (e.g., by model server 650) by performing quantization of floating-point AI / ML models using different quantization schemes and / or parameters. In this case, UE 120 can select a quantized AI / ML model from the multiple quantized AI / ML models for AI / ML tasks stored at model server 650. In such examples, each of the multiple quantized AI / ML models for AI / ML tasks can be associated with a corresponding model ID (e.g., a corresponding physical model ID). For example, model server 650 may store corresponding quantized AI / ML models associated with multiple physical model IDs, which are associated with the same AI / ML task (e.g., AI / ML functionality) and / or the same logical model ID. In some aspects, UE 120 may communicate with model server 650 via a protocol transparent to network node 110 and / or a wireless communication standard (e.g., 3GPP standard) used for communication between UE 120 and network node 110 to obtain quantized AI / ML models.
[0143] like Figure 8As further shown by reference numeral 810, model server 650 can send an indication of the inference accuracy level of the quantized AI / ML model to be used by UE 120, and network node 110 can receive the indication of the inference accuracy level of the quantized AI / ML model to be used by the UE. In some aspects, the inference accuracy level can indicate the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for an AI / ML task (e.g., a corresponding floating-point AI / ML model quantized to generate the quantized AI / ML model). In some aspects, the inference accuracy level can be the initial inference accuracy level of the quantized AI / ML model (e.g., before AI / ML performance monitoring performed by network node 110). In some aspects, model server 650 can send an indication to network node 110 of the inference accuracy level of the quantized AI / ML model sent to UE 120 (e.g., downloaded by the UE). In such examples, model server 650 may also send indications of a model ID (e.g., physical model ID) associated with the quantized AI / ML model and / or an indication of a logical model ID associated with the floating-point AI / ML model and / or the AI / ML task. For example, model server 650 may send indications of a model ID (e.g., physical model ID) to register the quantized AI / ML model with network node 110 as an AI / ML model for an AI / ML task.
[0144] In some aspects, model server 650 may send to network node 110 an indication of the corresponding inference accuracy level for multiple quantized AI / ML models (e.g., including quantized AI / ML models to be used by UE 120) used for AI / ML tasks. In such examples, model server 650 may also send to network node 110 an indication of the corresponding model ID (e.g., physical model ID) for each of the multiple quantized AI / ML models (e.g., registering multiple quantized AI / ML models with network node 110 for AI / ML tasks). For example, the multiple quantized AI / ML models may be multiple quantized AI / ML models stored on model server 650 and / or already generated by the model server for AI / ML tasks (e.g., associated with the same logical model ID). In some examples, model server 650 may also indicate to network node 110 the logical model ID corresponding to the AI / ML task or floating-point AI / ML model associated with the multiple quantized AI / ML models. In some examples, for each of the multiple floating-point AI / ML models (corresponding to the respective AI / ML tasks) managed by model server 650 (e.g., stored on the model server), model server 650 may send to network node 110 an indication of the corresponding logical model ID and a corresponding indication of the physical model ID and inference accuracy level of one or more quantized AI / ML models associated with that logical model ID.
[0145] In some aspects, model server 650 may send an indication to network node 110 of the inference accuracy level (and physical model ID) of the quantized AI / ML model to be used by UE 120 after sending the quantized AI / ML model to UE 120. For example, in response to UE 120 downloading the quantized AI / ML model from model server 650 (or in response to generating the quantized AI / ML model for UE 120 if model server 650 generates the quantized AI / ML model), model server 650 may send an indication of the inference accuracy level (and physical model ID) of the quantized AI / ML model to network node 110. In some other aspects, model server 650 may send an indication of the inference accuracy level (and physical model ID) of the quantized AI / ML model to be used by UE 120 to network node 110 before sending the quantized AI / ML model to UE 120. For example, if model server 650 stores multiple quantized AI / ML models for AI / ML tasks that can be downloaded by the UE, model server 650 can combine the generated quantized AI / ML models and, independently of when UE 120 downloads the quantized AI / ML models, send an indication of the inference accuracy level and physical model ID of the quantized AI / ML models to network node 110. In some aspects, model server 650 can send the indication of the inference accuracy level and physical model ID of one or more quantized AI / ML models to multiple network nodes (e.g., including network node 110). In some aspects, network node 110 can receive the indication of the inference accuracy level and physical model ID of the quantized AI / ML models from multiple model servers (e.g., multiple third-party model servers).
[0146] In some aspects, the definition of the inference accuracy level may be pre-configured / predefined (e.g., specified in a wireless communication standard) or configured / indicated via signaling from network node 110. In such examples, different definitions of inference accuracy may be pre-configured / predefined or configured / indicated for different AI / ML tasks (e.g., different AI / ML functional or logical models) (e.g., different measurements of the relative accuracy of a quantized AI / ML model relative to a floating-point AI / ML model). In some examples where the definition of the inference accuracy level (e.g., for an AI / ML task) is configured / indicated by network node 110, model server 650 may receive signaling from network node 110. In other examples where the definition of the inference accuracy level (e.g., for an AI / ML task) is configured / indicated by network node 110, UE 120 may receive signaling from the network node and indicate the definition of the inference accuracy level to model server 650 (e.g., when communicating with model server 650 to obtain a quantized AI / ML model). In some aspects, multiple candidate inference accuracy levels can be associated with an AI / ML task, and the indication of the inference accuracy level can indicate the candidate inference accuracy level selected from the multiple candidate inference accuracy levels. In some examples, the multiple candidate inference accuracy levels associated with the AI / ML task can be pre-configured / predefined (e.g., specified in a wireless communication standard). In some other examples, network node 110 can send configuration information indicating multiple candidate inference accuracy levels for the AI / ML task, and model server 650 and / or UE 120 can receive the configuration information indicating multiple candidate inference accuracy levels for the AI / ML task. In some aspects, the inference accuracy levels can be different for different model IDs associated with the same AI / ML task (e.g., the same AI / ML functionality). In some examples, multiple candidate inference accuracy levels can be associated with different model IDs, and the selected candidate inference accuracy level (indicated in the indication of the inference accuracy level) can be associated with the model ID of the quantized AI / ML model.
[0147] In some respects, the inference accuracy level can be based at least in part on the statistical difference between certain model output features provided by a floating-point AI / ML model for the AI / ML task and a quantized AI / ML model to be used by UE 120 for the AI / ML task, and / or the statistical difference between certain model output features (e.g., the true values of certain model output features) provided by a sprite-assisted detector / model for the AI / ML task and a quantized AI / ML model to be used by UE 120 (associated with a dataset (e.g., a reference dataset) that includes model input and output features provided by the sprite-assisted model) for the AI / ML task. In some respects, as described above... Figure 7 The discussed error values refer to the output values of one or more target output features determined by applying a floating-point AI / ML model for an AI / ML task to a reference dataset, which may be the true values (e.g., values determined via a sprite-assisted detector) included in the reference dataset for the one or more target output features. X Internal. For example, error values. X This could be the error tolerance of a floating-point AI / ML model used for AI / ML tasks, specified in a wireless communication standard. In some aspects, as combined with the above... Figure 7 The inference accuracy level discussed can be the difference between the output values of one or more target output features determined by applying the quantized AI / ML model to a reference dataset and the output values of one or more target features determined by applying a floating-point AI / ML model to the reference dataset. Y .
[0148] In some respects, where an AI / ML task is associated with predicting beam measurements (e.g., L1-RSRP measurements) for a set of beams (e.g., set A beams), the inference accuracy level can be based at least in part on the difference between a first beam measurement (e.g., a first L1-RSRP measurement) and a second beam measurement, the first beam measurement being predicted by a quantized AI / ML model for a subset of the beam set (e.g., a subset of set A beams) based on model inputs included in a reference dataset (e.g., beam measurements for set B beams), and the second beam measurement being predicted by a floating-point AI / ML model for a subset of the beam set (e.g., a subset of set A beams) based on model inputs included in a reference dataset. For example, a quantized AI / ML model (and a floating-point AI / ML model) can take L1-RSRP measurements for set B beams (e.g., a wide beam set) as input and output L1-RSRP measurements for set A beams (e.g., a narrow beam set). The predicted L1-RSRP value determined by the floating-point AI / ML model for the reference dataset and the value for the previous dataset. N (For example, in one example,)N =4) Set A beams (e.g., the set A beams with the highest predicted L1-RSRP value) N The L1-RSRP prediction accuracy between the true L1-RSRP values (e.g., values determined via a sprite-assisted detector) of each beam can be ± X Within dB (of which) X (This is an error tolerance that can be specified in the wireless communication standard). It is used to compare the predicted L1-RSRP value output by the floating-point AI / ML model with the true L1-RSRP value. N The set A beams can be real front beams from the reference dataset. N A beam identifier (e.g., via a sprite-assisted detector). N The value can be predefined or preconfigured (e.g., specified in wireless communication standards). In this example, the inference accuracy level of the quantized AI / ML model can be the L1-RSRP value predicted by the quantized AI / ML model for a reference dataset and the value predicted by the floating-point AI / ML model for the previous dataset. N A set of A beams (e.g., the real front beams identified in the reference dataset) N The difference between the predicted L1-RSRP values determined by the reference dataset of a set of A beams. Y For example, quantified AI / ML models targeting the previous... N A set of A beams (e.g., real front beams) N The L1-RSRP value predicted by the set of A beams can be obtained from the floating-point AI / ML model for the previous... N ± L1-RSRP values predicted by the A-beam set Y Within dB. For example, Y It can be the front N The average L1-RSRP difference on beam A of each set, the front N The total L1-RSRP difference on beam A of each set, the front N The maximum L1-RSRP difference on beam A of the set or the previous N A vector of L1-RSRP differences for each A-beam. In some respects, an indication of the inference accuracy level of a quantized AI / ML model can be... Y Instructions, or may include instructions on Y In some examples, the indication of the inference accuracy level of the quantized AI / ML model sent from model server 650 to network node 110 may include... YInstructions for the inference accuracy level and instructions for the logical model ID associated with the AI / ML task. In such examples, instructions for the inference accuracy level sent from model server 650 to network node 110 may also include instructions for the physical model ID associated with the quantized AI / ML model (e.g., registering the physical model ID with network node 110 for the AI / ML task).
[0149] In some respects, when an AI / ML task is associated with a certain number of top beams in a set of predicted beams (e.g., identifying which...), N The set A beam is the front N The inference accuracy level can be based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in a reference dataset. For example, the inference accuracy level can be based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in a reference dataset relative to the percentage of top beams predicted by a floating-point AI / ML model based on model inputs included in the dataset. In this example, the reference accuracy level... Y It can be compared with the predictions made by applying a floating-point AI / ML model to a reference dataset. N Compared to a set of A beams, this method accurately predicts the front beams by applying a quantized AI / ML model to a reference dataset. N The percentage of each A-beam in the set. In some respects, this can be indicated in the indication of the inference accuracy level of the quantized AI / ML model sent from model server 650 to network node 110. Y That value.
[0150] In some respects, the model server 650 can determine the inference accuracy level of the quantized AI / ML model by applying the quantized AI / ML model and the floating-point AI / ML model to a reference dataset (e.g., YThe reference dataset used to determine inference accuracy may include model inputs and real (e.g., sprite-assisted) model output values. In some aspects, the reference dataset used to determine the level of inference accuracy may be predefined (e.g., specified in a wireless communication standard). For example, a wireless communication standard may specify different reference datasets for different AI / ML tasks (e.g., different AI / ML functionalities). In some aspects, the reference dataset may be indicated via signaling sent by network node 110. For example, network node 110 may send the reference dataset, and model server 650 may receive the reference dataset. In some examples, network node 110 may signal the dataset to one or more third-party entities (e.g., model server 650 and / or one or more other model servers) where the quantized AI / ML model can be downloaded by the UE. In some other aspects, model server 650 may send the reference dataset, and network node 110 may receive the reference dataset. In this case, model server 650 may send an indication of the logical model ID associated with the floating-point AI / ML model and / or AI / ML task along with the reference dataset. For example, one or more third-party entities (e.g., including model server 650 and / or one or more other model servers) may signal the reference dataset used by the third-party entity, along with the model ID (e.g., logical model ID) corresponding to the AI / ML task associated with the reference dataset, to network node 110.
[0151] like Figure 8As further shown by reference numeral 815, UE 120 may send an indication of a model ID (e.g., physical model ID) associated with a quantized AI / ML model to be used by UE 120 (e.g., a quantized AI / ML model received from model server 650 at UE 120), and network node 110 may receive an indication of a model ID (e.g., physical model ID) associated with a quantized AI / ML model to be used by the UE (e.g., a quantized AI / ML model received from model server at the UE). Network node 110 may identify the inference accuracy level of the quantized AI / ML model to be used by UE 120 based at least in part on the indication of the physical model ID received from UE 120 and the indication of the inference accuracy level corresponding to the physical model ID received from model server 650. In some examples, UE 120 may also send to network node 110 an indication of a logical model ID corresponding to an AI / ML task and / or floating-point AI / ML model associated with the quantized AI / ML model. In some aspects, UE 120 may send an indication of the physical model ID associated with the quantized AI / ML model to network node 110 before UE 120 is activated using the quantized AI / ML model and / or the AI / ML task (e.g., activated by network node 110). For example, UE 120 may send the indication of the physical model ID associated with the quantized AI / ML model (e.g., together with the indication of the logical model ID associated with the AI / ML task) as UE capability information during initial access.
[0152] like Figure 8 As further shown by reference numeral 820 in the accompanying drawings, network node 110 may send signaling associated with AI / ML performance monitoring based at least in part on an indication of the inference accuracy level of the quantized AI / ML model, and UE 120 may receive signaling associated with AI / ML performance monitoring based at least in part on an indication of the inference accuracy level of the quantized AI / ML model. In some aspects, network node 110 may determine one or more AI / ML performance monitoring parameters for UE 120 based at least in part on the indicated inference accuracy level of the quantized AI / ML model to be used by UE 120, and network node 110 may send signaling to UE 120 indicating configuration information of one or more AI / ML performance monitoring parameters and / or scheduling information based at least in part on one or more AI / ML performance parameters.
[0153] In some aspects, network node 110 may schedule the transmission of RS (e.g., auxiliary RS) for prediction accuracy verification of UE 120 based at least in part on the inference accuracy level indicated by the quantized AI / ML model. For example, network node 110 may schedule the transmission of RS for prediction accuracy verification to UE 120 more frequently or less frequently (e.g., compared to the default scheduling frequency and / or compared to RS scheduled for other UEs) based at least in part on the inference accuracy level indicated by the quantized AI / ML model. In such examples, signaling associated with AI / ML performance monitoring may include signaling instructions for scheduling RS to be sent to UE 120 and / or signaling for the transmission of RS for prediction accuracy verification from network node 110 to UE 120.
[0154] In some aspects, network node 110 can determine / configure error thresholds for UE 120 to determine whether a performance alert message should be sent to UE 120. For example, network node 110 can determine / configure higher or lower error thresholds for different UEs associated with AI / ML tasks, based at least in part on the corresponding inference accuracy level of the quantized AI / ML model to be used by different UEs. In such examples, signaling associated with AI / ML performance monitoring may include signaling indicating or configuring error thresholds for UE 120 and / or performance alert messages sent to UE 120 based at least in part on error thresholds determined / configured for UE 120.
[0155] As indicated above, Figure 8 This is provided as an example. Other examples may be provided relative to... Figure 8 The examples described are different.
[0156] Figure 9 This is a diagram illustrating an example process 900 performed, for example, at a UE or a device of a UE, according to this disclosure. Example process 900 is an example in which a device or UE (e.g., UE 120) performs operations associated with a report on the inference accuracy of a quantified model used for AI / ML performance monitoring.
[0157] like Figure 9 As shown, in some aspects, process 900 may include: sending to a network node an indication of an inference accuracy level associated with a quantized AI / ML model used for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task (box 910). For example, a UE (e.g., using...) Figure 12The sending component 1204 and / or communication manager 1206 described herein can send to network nodes an indication of an inference accuracy level associated with a quantized AI / ML model used for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task, as described above.
[0158] like Figure 9 As further shown, in some aspects, process 900 may include: receiving signaling associated with AI / ML performance monitoring from a network node (box 920) based at least in part on an indication of the inference accuracy level. For example, the UE (e.g., using...) Figure 12 The receiving component 1202 and / or communication manager 1206 described herein can receive signaling associated with AI / ML performance monitoring from network nodes, at least in part, based on an indication of the inference accuracy level, as described above.
[0159] Process 900 may include additional aspects, such as any single aspect or any combination of aspects described below and / or in conjunction with one or more other processes described elsewhere in this document.
[0160] In the first aspect, the indication of the inference accuracy level indicates the candidate inference accuracy level selected from multiple candidate inference accuracy levels associated with the AI / ML task.
[0161] In the second aspect, either alone or in combination with the first aspect, multiple candidate inference accuracy levels are associated with different model identifiers, and the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0162] In a third aspect, either alone or in combination with one or more of the first and second aspects, process 900 includes: receiving configuration information from a network node indicating multiple candidate inference accuracy levels associated with an AI / ML task.
[0163] In the fourth aspect, either alone or in combination with one or more of the first to third aspects, the AI / ML task is associated with predicting beam measurements for a beam set, and the inference of an accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted for a subset of the beam set by a quantized AI / ML model based on model inputs included in the dataset, and the second beam measurement being predicted for a subset of the beam set by a floating-point AI / ML model based on model inputs included in the dataset.
[0164] In the fifth aspect, either alone or in combination with one or more of the first to fourth aspects, a subset of the beam set includes multiple real top beams associated with the dataset.
[0165] In the sixth aspect, either alone or in combination with one or more of the first to fifth aspects, the first beam measurement includes a first L1-RSRP measurement, and the second beam measurement includes a second L1-RSRP measurement.
[0166] In the seventh aspect, either alone or in combination with one or more of the first to sixth aspects, the AI / ML task is associated with a certain number of top beams in the predicted beam set, and the inference accuracy level is based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in the dataset, relative to the percentage of top beams predicted by a floating-point AI / ML model based on model inputs included in the dataset.
[0167] although Figure 9 An example box of process 900 is shown, but in some respects, process 900 may include... Figure 9 The boxes depicted in the diagram may be fewer, different, or arranged differently than additional boxes. Alternatively, two or more boxes in the process 900 may be executed in parallel.
[0168] Figure 10 This is a diagram illustrating an example process 1000 performed, for example, at a network node or a device of a network node, according to this disclosure. Example process 1000 is an example in which a device or network node (e.g., network node 110) performs operations associated with a report on the inference accuracy of a quantified model used for AI / ML performance monitoring.
[0169] like Figure 10 As shown, in some aspects, process 1000 may include: receiving an indication of an inference accuracy level associated with a quantized AI / ML model to be used by the UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task (box 1010). For example, a network node (e.g., using...) Figure 13 The receiving component 1302 and / or communication manager 1306 described herein can receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by the UE for an AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task, as described above.
[0170] like Figure 10As further shown, in some aspects, process 1000 may include: sending signaling associated with AI / ML performance monitoring to the UE (box 1020) based at least in part on an indication of the inference accuracy level. For example, a network node (e.g., using...) Figure 13 The transmitting component 1304 and / or the communication manager 1306 described herein may transmit signaling associated with AI / ML performance monitoring to the UE, at least in part, based on an indication of the inference accuracy level, as described above.
[0171] Process 1000 may include additional aspects, such as any single aspect or any combination of aspects described below and / or in conjunction with one or more other processes described elsewhere in this document.
[0172] In the first aspect, receiving an indication of the inference accuracy level includes: receiving an indication of the inference accuracy level from the UE.
[0173] In the second aspect, receiving an indication of the inference accuracy level, either alone or in combination with the first aspect, includes receiving an indication of the inference accuracy level from a server device associated with the floating-point AI / ML model and the quantized AI / ML model.
[0174] In a third aspect, receiving an indication of an inference accuracy level from a server device associated with a floating-point AI / ML model and a quantized AI / ML model, either alone or in combination with one or more of the first and second aspects, includes receiving from the server device an indication of a corresponding inference accuracy level for a plurality of quantized AI / ML models used for an AI / ML task, wherein the plurality of quantized AI / ML models include quantized AI / ML models.
[0175] In the fourth aspect, either alone or in combination with one or more of the first to third aspects, process 1000 includes: receiving from the UE an indication of a model identifier associated with the quantized AI / ML model.
[0176] In the fifth aspect, either alone or in combination with one or more of the first to fourth aspects, the indication of the inference accuracy level indicates a candidate inference accuracy level selected from multiple candidate inference accuracy levels associated with the AI / ML task.
[0177] In the sixth aspect, either alone or in combination with one or more of the first to fifth aspects, multiple candidate inference accuracy levels are associated with different model identifiers, and the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0178] In the seventh aspect, either alone or in combination with one or more of the first to sixth aspects, process 1000 includes: sending configuration information to the UE indicating multiple candidate inference accuracy levels associated with the AI / ML task.
[0179] In the eighth aspect, either alone or in combination with one or more of the first to seventh aspects, the AI / ML task is associated with predicting beam measurements for a beam set, and the inference of an accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted for a subset of the beam set by a quantized AI / ML model based on model inputs included in the dataset, and the second beam measurement being predicted for a subset of the beam set by a floating-point AI / ML model based on model inputs included in the dataset.
[0180] In the ninth aspect, either alone or in combination with one or more of the first to eighth aspects, a subset of the beam set includes multiple real top beams associated with the dataset.
[0181] In the tenth aspect, either alone or in combination with one or more of the first to ninth aspects, the first beam measurement includes a first L1-RSRP measurement, and the second beam measurement includes a second L1-RSRP measurement.
[0182] In the eleventh aspect, either alone or in combination with one or more of the first to tenth aspects, the AI / ML task is associated with a certain number of top beams in the predicted beam set, and the inference accuracy level is based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in the dataset, relative to the percentage of top beams predicted by a floating-point AI / ML model based on model inputs included in the dataset.
[0183] although Figure 10 An example box of process 1000 is shown, but in some respects, process 1000 may include... Figure 10 The boxes depicted in the diagram are compared to additional boxes, fewer boxes, different boxes, or boxes arranged in a different manner. Additionally or alternatively, two or more boxes in the process 1000 may be executed in parallel.
[0184] Figure 11 This is a diagram illustrating an example process 1100 performed, for example, at a server device or a device of a server device, according to this disclosure. Example process 1100 is an example in which a device or server device (e.g., model server 650) performs operations associated with a report on the inference accuracy of a quantified model for AI / ML performance monitoring.
[0185] like Figure 11As shown, in some aspects, process 1100 may include: sending a quantized AI / ML model for an AI / ML task to the UE, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task (box 1110). For example, a server device (e.g., using...) Figure 14 The transmitting component 1404 and / or communication manager 1406 described herein can transmit a quantized AI / ML model for an AI / ML task to the UE, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task, as described above.
[0186] like Figure 11 Further shown, in some aspects, process 1100 may include: sending an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model (box 1120). For example, a server device (e.g., using...) Figure 14 The sending component 1404 and / or communication manager 1406 described herein can send an indication of an inference accuracy level associated with a quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model, as described above.
[0187] Process 1100 may include additional aspects, such as any single aspect or any combination of aspects described below and / or in conjunction with one or more other processes described elsewhere in this document.
[0188] In the first aspect, sending an indication of the inference accuracy level includes: sending an indication of the inference accuracy level to network nodes.
[0189] In the second aspect, sending an indication of an inference accuracy level to a network node, either alone or in combination with the first aspect, includes sending an indication to the network node of a corresponding inference accuracy level for a plurality of quantized AI / ML models used for an AI / ML task, wherein the plurality of quantized AI / ML models include quantized AI / ML models.
[0190] In the third aspect, sending an indication of the inference accuracy level, either alone or in combination with one or more of the first and second aspects, includes sending an indication of the inference accuracy level to the UE.
[0191] In the fourth aspect, either alone or in combination with one or more of the first to third aspects, the indication of the inference accuracy level indicates a candidate inference accuracy level selected from multiple candidate inference accuracy levels associated with the AI / ML task.
[0192] In the fifth aspect, either alone or in combination with one or more of the first to fourth aspects, multiple candidate inference accuracy levels are associated with different model identifiers, and the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0193] In the sixth aspect, either alone or in combination with one or more of the first to fifth aspects, the AI / ML task is associated with predicting beam measurements for a beam set, and the inference of an accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted for a subset of the beam set by a quantized AI / ML model based on model inputs included in the dataset, and the second beam measurement being predicted for a subset of the beam set by a floating-point AI / ML model based on model inputs included in the dataset.
[0194] In the seventh aspect, either alone or in combination with one or more of the first to sixth aspects, a subset of the beam set includes multiple real top beams associated with the dataset.
[0195] In the eighth aspect, either alone or in combination with one or more of the first to seventh aspects, the first beam measurement includes a first L1-RSRP measurement, and the second beam measurement includes a second L1-RSRP measurement.
[0196] In the ninth aspect, either alone or in combination with one or more of the first to eighth aspects, the AI / ML task is associated with a certain number of top beams in the predicted beam set, and the inference accuracy level is based at least in part on the percentage of top beams accurately predicted by a quantized AI / ML model based on model inputs included in the dataset, relative to the percentage of top beams predicted by a floating-point AI / ML model based on model inputs included in the dataset.
[0197] although Figure 11 An example box of process 1100 is shown, but in some respects, process 1100 may include... Figure 11 The boxes depicted herein are compared to additional boxes, fewer boxes, different boxes, or boxes arranged in a different manner. Additionally or alternatively, two or more boxes in the process 1100 may be executed in parallel.
[0198] Figure 12This is a schematic diagram of an example device 1200 for wireless communication according to the present disclosure. Device 1200 may be a UE, or a UE may include device 1200. In some aspects, device 1200 includes a receiving component 1202, a transmitting component 1204, and / or a communication manager 1206 that can communicate with each other (e.g., via one or more buses and / or one or more other components). In some aspects, communication manager 1206 is combined with... Figure 1 The described communication manager 140. As shown, the device 1200 can communicate with another device 1208 (such as a UE, a network node (such as a CU, DU, RU, or base station)) or a server device using the receiving component 1202 and the transmitting component 1204.
[0199] In some respects, device 1200 can be configured to perform the functions described herein. Figures 4 to 8 One or more operations as described herein. Additionally or alternatively, the apparatus 1200 may be configured to perform one or more processes described herein (such as...). Figure 9 The process 900) or a combination thereof. In some respects, Figure 12 The illustrated device 1200 and / or one or more components may include a combination Figure 2 One or more components of the described UE. Additionally or alternatively, Figure 12 One or more components shown can be combined Figure 2 Implementation within one or more of the described components. Additionally or alternatively, one or more components in the set of components may be implemented at least partially as software stored in one or more memories. For example, a component (or a portion thereof) may be implemented as instructions or code stored in a non-transitory computer-readable medium and executable by one or more controllers or one or more processors to perform the function or operation of the component.
[0200] Receiver 1202 may receive communications from device 1208, such as reference signals, control information, data communications, or combinations thereof. Receiver 1202 may provide the received communications to one or more other components of device 1200. In some aspects, receiver 1202 may perform signal processing on the received communications (such as filtering, amplification, demodulation, analog-to-digital conversion, demultiplexing, deinterleaving, demapping, equalization, interference cancellation, or decoding, etc.) and may provide the processed signals to one or more other components of device 1200. In some aspects, receiver 1202 may include combinations of... Figure 2 The described UE includes one or more antennas, one or more modems, one or more demodulators, one or more MIMO detectors, one or more receiver processors, one or more controllers / processors, one or more memories, or combinations thereof.
[0201] Transmitting component 1204 may transmit communications, such as reference signals, control information, data communications, or combinations thereof, to device 1208. In some aspects, one or more other components of device 1200 may generate communications and provide the generated communications to transmitting component 1204 for transmission to device 1208. In some aspects, transmitting component 1204 may perform signal processing (such as filtering, amplification, modulation, digital-to-analog conversion, multiplexing, interleaving, mapping, or encoding, etc.) on the generated communications and may transmit the processed signals to device 1208. In some aspects, transmitting component 1204 may include combinations of... Figure 2 The described UE includes one or more antennas, one or more modems, one or more modulators, one or more transmit MIMO processors, one or more transmit processors, one or more controllers / processors, one or more memories, or combinations thereof. In some aspects, the transmit component 1204 may co-located with the receive component 1202 in one or more transceivers.
[0202] The communication manager 1206 may support the operation of the receiving component 1202 and / or the transmitting component 1204. For example, the communication manager 1206 may receive information associated with configuring the reception of communications by the receiving component 1202 and / or the transmission of communications by the transmitting component 1204. Additionally or alternatively, the communication manager 1206 may generate control information and / or provide control information to the receiving component 1202 and / or the transmitting component 1204 to control the reception and / or transmission of communications.
[0203] The transmitting component 1204 can send to the network node an indication of an inference accuracy level associated with a quantized AI / ML model used for the AI / ML task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task. The receiving component 1202 can receive signaling associated with AI / ML performance monitoring from the network node based at least in part on the indication of the inference accuracy level.
[0204] The receiving component 1202 can receive configuration information from the network node that indicates multiple candidate inference accuracy levels associated with the AI / ML task.
[0205] Figure 12 The number and arrangement of components shown are provided as an example. In practice, different arrangements may exist. Figure 12 The components shown are compared to additional components, fewer components, different components, or components arranged in a different manner. Furthermore, Figure 12 The two or more components shown can be implemented within a single component, or Figure 12The single component shown can be implemented as multiple distributed components. Additionally or alternatively, Figure 12 The collection of (one or more) components shown is executable and described as being composed of Figure 12 Another set of components shown performs one or more functions.
[0206] Figure 13 This is a schematic diagram of an example device 1300 for wireless communication according to the present disclosure. Device 1300 may be a network node, or a network node may include device 1300. In some aspects, device 1300 includes a receiving component 1302, a transmitting component 1304, and / or a communication manager 1306 that can communicate with each other (e.g., via one or more buses and / or one or more other components). In some aspects, communication manager 1306 is combined with... Figure 1 The described communication manager 150. As shown, device 1300 can communicate with another device 1308 (such as a UE or a network node (such as a CU, DU, RU or base station)) using receiving component 1302 and transmitting component 1304.
[0207] In some respects, device 1300 can be configured to perform the functions described herein. Figures 4 to 8 One or more operations as described herein. Additionally or alternatively, the apparatus 1300 may be configured to perform one or more processes described herein (such as...). Figure 10 The process 1000) or a combination thereof. In some respects, Figure 13 The illustrated device 1300 and / or one or more components may include a combination Figure 2 One or more components of the described network node. Additionally or alternatively, Figure 13 One or more components shown can be combined Figure 2 Implementation within one or more of the described components. Additionally or alternatively, one or more components in the set of components may be implemented at least partially as software stored in one or more memories. For example, a component (or a portion thereof) may be implemented as instructions or code stored in a non-transitory computer-readable medium and executable by one or more controllers or one or more processors to perform the function or operation of the component.
[0208] Receiver 1302 may receive communications from device 1308, such as reference signals, control information, data communications, or combinations thereof. Receiver 1302 may provide the received communications to one or more other components of device 1300. In some aspects, receiver 1302 may perform signal processing on the received communications (such as filtering, amplification, demodulation, analog-to-digital conversion, demultiplexing, deinterleaving, demapping, equalization, interference cancellation, or decoding, etc.) and may provide the processed signals to one or more other components of device 1300. In some aspects, receiver 1302 may include combinations of... Figure 2 The described network node includes one or more antennas, one or more modems, one or more demodulators, one or more MIMO detectors, one or more receiver processors, one or more controllers / processors, one or more memories, or combinations thereof. In some aspects, receiver component 1302 and / or transmitter component 1304 may include or be included in a network interface. The network interface may be configured to acquire and / or output signals for device 1300 via one or more communication links, such as backhaul links, midhaul links, and / or fronthaul links.
[0209] Transmitting component 1304 may transmit communications, such as reference signals, control information, data communications, or combinations thereof, to device 1308. In some aspects, one or more other components of device 1300 may generate communications and provide the generated communications to transmitting component 1304 for transmission to device 1308. In some aspects, transmitting component 1304 may perform signal processing (such as filtering, amplification, modulation, digital-to-analog conversion, multiplexing, interleaving, mapping, or encoding, etc.) on the generated communications and may transmit the processed signals to device 1308. In some aspects, transmitting component 1304 may include combinations of... Figure 2 The described network node includes one or more antennas, one or more modems, one or more modulators, one or more transmit MIMO processors, one or more transmit processors, one or more controllers / processors, one or more memories, or combinations thereof. In some aspects, the transmit component 1304 may co-located with the receive component 1302 in one or more transceivers.
[0210] The communication manager 1306 may support the operation of the receiving component 1302 and / or the transmitting component 1304. For example, the communication manager 1306 may receive information associated with configuring the reception of communications by the receiving component 1302 and / or the transmission of communications by the transmitting component 1304. Additionally or alternatively, the communication manager 1306 may generate control information and / or provide control information to the receiving component 1302 and / or the transmitting component 1304 to control the reception and / or transmission of communications.
[0211] The receiving component 1302 may receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by the UE for AI / ML tasks, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML tasks. The transmitting component 1304 may transmit signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0212] The receiving component 1302 can receive from the UE an indication of the model identifier associated with the quantized AI / ML model.
[0213] The transmitting component 1304 can send configuration information to the UE indicating multiple candidate inference accuracy levels associated with the AI / ML task.
[0214] Figure 13 The number and arrangement of components shown are provided as an example. In practice, different arrangements may exist. Figure 13 The components shown are compared to additional components, fewer components, different components, or components arranged in a different manner. Furthermore, Figure 13 The two or more components shown can be implemented within a single component, or Figure 13 The single component shown can be implemented as multiple distributed components. Additionally or alternatively, Figure 13 The collection of (one or more) components shown is executable and described as being composed of Figure 13 Another set of components shown performs one or more functions.
[0215] Figure 14 This is a schematic diagram of an example device 1400 for wireless communication according to the present disclosure. Device 1400 may be a server device, or a server device may include device 1400. In some aspects, device 1400 includes a receiving component 1402, a transmitting component 1404, and / or a communication manager 1406 that can communicate with each other (e.g., via one or more buses and / or one or more other components). In some aspects, communication manager 1406 is combined with... Figure 1 The described communication manager 150. As shown, device 1400 can communicate with another device 1408 (such as a UE or a network node (such as a CU, DU, RU or base station)) using receiving component 1402 and transmitting component 1404.
[0216] In some respects, device 1400 can be configured to perform the functions described herein. Figures 4 to 8 One or more operations as described herein. Additionally or alternatively, the apparatus 1400 may be configured to perform one or more processes described herein (such as...). Figure 11 The process 1100) or a combination thereof. In some respects, Figure 14 The illustrated device 1400 and / or one or more components may include a combination Figure 2 One or more components of the described server device. Additionally or alternatively, Figure 14 One or more components shown can be combined Figure 2 Implementation within one or more of the described components. Additionally or alternatively, one or more components in the set of components may be implemented at least partially as software stored in one or more memories. For example, a component (or a portion thereof) may be implemented as instructions or code stored in a non-transitory computer-readable medium and executable by one or more controllers or one or more processors to perform the function or operation of the component.
[0217] Receiver 1402 may receive communications from device 1408, such as reference signals, control information, data communications, or combinations thereof. Receiver 1402 may provide the received communications to one or more other components of device 1400. In some aspects, receiver 1402 may perform signal processing on the received communications (such as filtering, amplification, demodulation, analog-to-digital conversion, demultiplexing, deinterleaving, demapping, equalization, interference cancellation, or decoding, etc.) and may provide the processed signals to one or more other components of device 1400. In some aspects, receiver 1402 may include combinations of... Figure 2 The described server device includes one or more antennas, one or more modems, one or more demodulators, one or more MIMO detectors, one or more receiver processors, one or more controllers / processors, one or more memories, or combinations thereof.
[0218] Transmitting component 1404 may transmit communications, such as reference signals, control information, data communications, or combinations thereof, to device 1408. In some aspects, one or more other components of device 1400 may generate communications and provide the generated communications to transmitting component 1404 for transmission to device 1408. In some aspects, transmitting component 1404 may perform signal processing (such as filtering, amplification, modulation, digital-to-analog conversion, multiplexing, interleaving, mapping, or encoding, etc.) on the generated communications and may transmit the processed signals to device 1408. In some aspects, transmitting component 1404 may include combinations of... Figure 2 The described server device includes one or more antennas, one or more modems, one or more modulators, one or more transmit MIMO processors, one or more transmit processors, one or more controllers / processors, one or more memories, or combinations thereof. In some aspects, the transmit component 1404 may co-located with the receive component 1402 in one or more transceivers.
[0219] The communication manager 1406 may support the operation of the receiving component 1402 and / or the transmitting component 1404. For example, the communication manager 1406 may receive information associated with configuring the reception of communications by the receiving component 1402 and / or the transmission of communications by the transmitting component 1404. Additionally or alternatively, the communication manager 1406 may generate control information and / or provide control information to the receiving component 1402 and / or the transmitting component 1404 to control the reception and / or transmission of communications.
[0220] The transmitting component 1404 can transmit to the UE a quantized AI / ML model for an AI / ML task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task. The transmitting component 1404 can also transmit an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0221] Figure 14 The number and arrangement of components shown are provided as an example. In practice, different arrangements may exist. Figure 14 The components shown are compared to additional components, fewer components, different components, or components arranged in a different manner. Furthermore, Figure 14 The two or more components shown can be implemented within a single component, or Figure 14 The single component shown can be implemented as multiple distributed components. Additionally or alternatively, Figure 14 The collection of (one or more) components shown is executable and described as being composed of Figure 14 Another set of components shown performs one or more functions.
[0222] The following provides an overview of some aspects of this disclosure: Aspect 1: A method of wireless communication performed by a user equipment (UE), the method comprising: sending to a network node an indication of an inference accuracy level associated with a quantized AI / ML model for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model for the AI / ML task; and receiving from the network node signaling associated with AI / ML performance monitoring based at least in part on the indication of the inference accuracy level.
[0223] Aspect 2: According to the method of aspect 1, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
[0224] Aspect 3: According to the method of aspect 2, wherein the plurality of candidate inference accuracy levels are associated with different model identifiers, and wherein the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0225] Aspect 4: The method according to any one of Aspects 2 to 3, the method further comprising: receiving from the network node configuration information indicating the plurality of candidate inference accuracy levels associated with the AI / ML task.
[0226] Aspect 5: The method according to any one of Aspects 1 to 4, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
[0227] Aspect 6: According to the method of aspect 5, the subset of the beam set includes a plurality of real top beams associated with the dataset.
[0228] Aspect 7: The method according to any one of Aspects 5 to 6, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
[0229] Aspect 8: The method according to any one of Aspects 1 to 4, wherein the AI / ML task is associated with a certain number of top beams in a predicted beam set, and wherein the inference accuracy level is based at least in part on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
[0230] Aspect 9: A method of wireless communication performed by a network node, the method comprising: receiving an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a user equipment (UE) for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and sending signaling associated with AI / ML performance monitoring to the UE based at least in part on the indication of the inference accuracy level.
[0231] Aspect 10: According to the method of aspect 9, receiving the indication of the inference accuracy level includes: receiving the indication of the inference accuracy level from the UE.
[0232] Aspect 11: According to the method of aspect 9, receiving the indication for the inference accuracy level includes: receiving the indication for the inference accuracy level from a server device associated with the floating-point AI / ML model and the quantized AI / ML model.
[0233] Aspect 12: According to the method of aspect 11, receiving the indication of the inference accuracy level from the server device associated with the floating-point AI / ML model and the quantized AI / ML model includes: receiving from the server device an indication of a corresponding inference accuracy level for a plurality of quantized AI / ML models for the AI / ML task, wherein the plurality of quantized AI / ML models includes the quantized AI / ML model.
[0234] Aspect 13: The method according to aspect 12, the method further comprising: receiving from the UE an indication of a model identifier associated with the quantized AI / ML model.
[0235] Aspect 14: The method according to any one of Aspects 9 to 13, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
[0236] Aspect 15: According to the method of aspect 14, wherein the plurality of candidate inference accuracy levels are associated with different model identifiers, and wherein the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0237] Aspect 16: The method according to any one of Aspects 14 to 15, the method further comprising: sending configuration information to the UE indicating the plurality of candidate inference accuracy levels associated with the AI / ML task.
[0238] Aspect 17: The method according to any one of Aspects 9 to 16, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
[0239] Aspect 18: According to the method of aspect 17, the subset of the beam set includes a plurality of real top beams associated with the dataset.
[0240] Aspect 19: The method according to any one of Aspects 17 to 18, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
[0241] Aspect 20: The method according to any one of Aspects 9 to 16, wherein the AI / ML task is associated with a certain number of top beams in a set of predicted beams, and wherein the inference accuracy level is based at least in part on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
[0242] Aspect 21: A method of wireless communication performed by a server device, the method comprising: transmitting to a user equipment (UE) a quantized AI / ML model for an artificial intelligence or machine learning (AI / ML) task, wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task; and transmitting an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
[0243] Aspect 22: According to the method of aspect 21, sending the indication of the inference accuracy level includes: sending the indication of the inference accuracy level to a network node.
[0244] Aspect 23: According to the method of aspect 22, sending the indication of the inference accuracy level to the network node includes: sending the indication of a corresponding inference accuracy level of a plurality of quantized AI / ML models for the AI / ML task to the network node, wherein the plurality of quantized AI / ML models includes the quantized AI / ML model.
[0245] Aspect 24: According to the method of aspect 21, sending the indication of the inference accuracy level includes: sending the indication of the inference accuracy level to the UE.
[0246] Aspect 25: The method according to any one of Aspects 21 to 24, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
[0247] Aspect 26: According to the method of aspect 25, wherein the plurality of candidate inference accuracy levels are associated with different model identifiers, and wherein the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
[0248] Aspect 27: The method according to any one of Aspects 21 to 26, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
[0249] Aspect 28: According to the method of aspect 27, the subset of the beam set includes a plurality of real top beams associated with the dataset.
[0250] Aspect 29: The method according to any one of Aspects 27 to 28, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
[0251] Aspect 30: The method according to any one of Aspects 21 to 26, wherein the AI / ML task is associated with a certain number of top beams in a set of predicted beams, and wherein the inference accuracy level is based at least in part on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
[0252] Aspect 31: An apparatus for wireless communication at a device, the apparatus comprising: one or more processors; one or more memories coupled to the one or more processors; and instructions stored in the one or more memories and executable by the one or more processors to cause the apparatus to perform the method according to one or more of aspects 1 to 30.
[0253] Aspect 32: An apparatus for wireless communication at a device, the apparatus comprising: one or more memories; and one or more processors coupled to the one or more memories, the one or more processors being configured to cause the device to perform the method according to one or more of aspects 1 to 30.
[0254] Aspect 33: An apparatus for wireless communication, the apparatus comprising at least one component for performing the method according to one or more of aspects 1 to 30.
[0255] Aspect 34: A non-transitory computer-readable medium storing code for wireless communication, said code including instructions executable by one or more processors to perform the methods described in one or more of aspects 1 to 30.
[0256] Aspect 35: A non-transitory computer-readable medium storing a set of instructions for wireless communication, the set of instructions comprising one or more instructions that, when executed by one or more processors of a device, cause the device to perform the method according to one or more of aspects 1 to 30.
[0257] Aspect 36: A device for wireless communication, the device including a processing system comprising one or more processors and one or more memories coupled to the one or more processors, the processing system being configured to cause the device to perform the method according to one or more of aspects 1 to 30.
[0258] Aspect 37: An apparatus for wireless communication at a device, the apparatus comprising: one or more memories; and one or more processors coupled to the one or more memories, the one or more processors being individually or collectively configured to cause the device to perform the method according to one or more of aspects 1 to 30.
[0259] While the foregoing disclosure provides examples and descriptions, it is not intended to be exhaustive or to limit the aspects to the precise form disclosed. Modifications and variations may be made based on the foregoing disclosure, or from various forms of practice.
[0260] As used herein, the term "component" is intended to be broadly interpreted as hardware or a combination of hardware and at least one of software or firmware. "Software" should be broadly interpreted as instructions, instruction sets, code, code segments, program code, programs, subroutines, software modules, applications, software applications, software packages, routines, subroutines, objects, executable programs, threads of execution, procedures, or functions, whether referred to as software, firmware, middleware, microcode, hardware description languages, or other terms. As used herein, a "processor" is implemented in hardware or a combination of hardware and software. It will be apparent that the systems or methods described herein may be implemented in various forms of hardware or combinations of hardware and software. The actual dedicated control hardware or software code used to implement these systems or methods is not limited in any way. Therefore, the operation and behavior of these systems or methods are described herein without reference to specific software code, as those skilled in the art will understand that the software and hardware can be designed to implement these systems or methods, at least in part, based on the description herein. Unless otherwise stated, a component configured to perform a function means that the component has the capability to perform that function, but it is not necessary for the component to actually perform that function.
[0261] As used in this article, depending on the context, "meeting the threshold" can mean a value greater than the threshold, greater than or equal to the threshold, less than the threshold, less than or equal to the threshold, equal to the threshold, not equal to the threshold, etc.
[0262] As used in this article, the phrase “at least one of” in a list of items refers to any combination of these items, including a single member. As an example, “at least one of the following: a, b, or c” is intended to cover a, b, c, a+b, a+c, b+c, and a+b+c, as well as any combination with multiple of the same element (e.g., a+a, a+a+a, a+a+b, a+a+c, a+b+b, a+c+c, b+b, b+b+b, b+b+c, c+c, and c+c+c, or any other ordering of a, b, and c).
[0263] No element, action, or instruction used herein should be construed as essential or necessary unless explicitly stated otherwise. Furthermore, as used herein, the articles “a” and “an” are intended to include one or more items and are used interchangeably with “one or more.” Similarly, as used herein, the article “described” is intended to include one or more items mentioned in connection with the article “described” and is used interchangeably with “one or more.” Furthermore, as used herein, the terms “set” and “group” are intended to include one or more items and are used interchangeably with “one or more.” If only one item is desired, the phrase “only one” or similar terminology will be used. Moreover, as used herein, the terms “having” and similar terms are intended as open-ended terms that do not limit the elements they modify (e.g., “having” A may also have B). Additionally, the phrase “based on” is intended to mean “based on or otherwise related to” unless otherwise explicitly stated. Furthermore, as used herein, the term “or” is intended to be inclusive when used consecutively and is interchangeable with “and / or” unless otherwise explicitly stated (e.g., if used in conjunction with “either of the two” or “only one of them”). It should be understood that “one or more” is equivalent to “at least one”.
[0264] Although specific combinations of features are set forth in the claims or disclosed in the description, these combinations are not intended to limit the disclosure of various aspects. Many of these features may be combined in ways not specifically stated in the claims or disclosed in the description. The disclosure of various aspects includes each dependent claim in combination with each other claim in the claim set.
Claims
1. A user equipment (UE) for wireless communication, the user equipment (UE) comprising: One or more memory units; and One or more processors, said one or more processors coupled to said one or more memories, said one or more processors being configured to cause the UE to: Sending an indication to network nodes of an inference accuracy level associated with a quantized AI / ML model used for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and Signaling associated with AI / ML performance monitoring is received from the network node based at least in part on the indication of the inference accuracy level.
2. The UE of claim 1, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
3. The UE of claim 2, wherein the plurality of candidate inference accuracy levels are associated with different model identifiers, and wherein the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
4. The UE of claim 2, wherein the one or more processors are further configured to cause the UE to: Receive configuration information from the network node indicating the plurality of candidate inference accuracy levels associated with the AI / ML task.
5. The UE of claim 1, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
6. The UE of claim 5, wherein the subset of the beam set includes a plurality of real top beams associated with the dataset.
7. The UE of claim 5, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
8. The UE of claim 1, wherein the AI / ML task is associated with a number of top beams in a predicted beam set, and wherein the inference accuracy level is at least in part based on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
9. A network node for wireless communication, the network node comprising: One or more memory units; and One or more processors, said one or more processors coupled to said one or more memories, said one or more processors being configured to cause the network node to: Receive an indication of an inference accuracy level associated with a quantized AI / ML model to be used by a user equipment (UE) for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and Signaling associated with AI / ML performance monitoring is sent to the UE based at least in part on the indication of the inference accuracy level.
10. The network node of claim 9, wherein, in order for the network node to receive the indication of the inference accuracy level, the one or more processors are configured to cause the network node to: The UE receives the indication of the inferred accuracy level.
11. The network node of claim 9, wherein, in order for the network node to receive the indication of the inference accuracy level, the one or more processors are configured to cause the network node to: The indication of the inference accuracy level is received from the server device associated with the floating-point AI / ML model and the quantized AI / ML model.
12. The network node of claim 11, wherein, in order for the network node to receive the indication of the inference accuracy level from the server device associated with the floating-point AI / ML model and the quantized AI / ML model, the one or more processors are configured to cause the network node to: The server device receives indications of the corresponding inference accuracy levels for a plurality of quantized AI / ML models used for the AI / ML task, wherein the plurality of quantized AI / ML models include the quantized AI / ML model.
13. The network node of claim 12, wherein the one or more processors are further configured to cause the network node to: The UE receives an indication of the model identifier associated with the quantized AI / ML model.
14. The network node of claim 9, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
15. The network node of claim 14, wherein the plurality of candidate inference accuracy levels are associated with different model identifiers, and wherein the selected candidate inference accuracy level is associated with the model identifier of the quantized AI / ML model.
16. The network node of claim 14, wherein the one or more processors are further configured to cause the network node to: The UE is sent configuration information indicating the plurality of candidate inference accuracy levels associated with the AI / ML task.
17. The network node of claim 9, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
18. The network node of claim 17, wherein the subset of the beam set includes a plurality of real top beams associated with the dataset.
19. The network node of claim 17, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
20. The network node of claim 9, wherein the AI / ML task is associated with a number of top beams in a set of predicted beams, and wherein the inference accuracy level is at least in part based on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
21. A server device for wireless communication, the server device comprising: One or more memory units; and One or more processors, said one or more processors coupled to said one or more memories, said one or more processors being configured to cause the server device to: Sending a quantized AI / ML model for an artificial intelligence or machine learning (AI / ML) task to a user equipment (UE), wherein the quantized AI / ML model is associated with a floating-point AI / ML model for the AI / ML task; and Send an indication of an inference accuracy level associated with the quantized AI / ML model, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to the floating-point AI / ML model.
22. The server device of claim 21, wherein, in order for the server device to send the indication of the inference accuracy level, the one or more processors are configured to cause the server device to: Send the indication of the inference accuracy level to the network node.
23. The server device of claim 22, wherein, in order for the server device to send the indication of the inference accuracy level to the network node, the one or more processors are configured to cause the server device to: Send to the network node an indication of the corresponding inference accuracy level of a plurality of quantized AI / ML models used for the AI / ML task, wherein the plurality of quantized AI / ML models include the quantized AI / ML model.
24. The server device of claim 21, wherein, in order for the server device to send the indication of the inference accuracy level, the one or more processors are configured to cause the server device to: Send the indication of the inference accuracy level to the UE.
25. The server device of claim 21, wherein the indication of the inference accuracy level indicates a candidate inference accuracy level selected from a plurality of candidate inference accuracy levels associated with the AI / ML task.
26. The server device of claim 21, wherein the AI / ML task is associated with predicting beam measurements for a beam set, and wherein the inference accuracy level is based at least in part on the difference between a first beam measurement and a second beam measurement, the first beam measurement being predicted by the quantized AI / ML model for a subset of the beam set based on model inputs included in the dataset, and the second beam measurement being predicted by the floating-point AI / ML model for the subset of the beam set based on the model inputs included in the dataset.
27. The server device of claim 26, wherein the subset of the beam set includes a plurality of real top beams associated with the dataset.
28. The server device of claim 27, wherein the first beam measurement includes a first layer 1 reference signal received power (L1-RSRP) measurement, and wherein the second beam measurement includes a second L1-RSRP measurement.
29. The server device of claim 21, wherein the AI / ML task is associated with a number of top beams in a set of predicted beams, and wherein the inference accuracy level is at least in part based on the percentage of the top beams accurately predicted by the quantized AI / ML model based on model inputs included in the dataset, relative to the top beams predicted by the floating-point AI / ML model based on model inputs included in the dataset.
30. A method for wireless communication performed by a user equipment (UE), the method comprising: Sending an indication to network nodes of an inference accuracy level associated with a quantized AI / ML model used for an artificial intelligence or machine learning (AI / ML) task, wherein the inference accuracy level indicates the relative accuracy of the quantized AI / ML model relative to a floating-point AI / ML model used for the AI / ML task; and Signaling associated with AI / ML performance monitoring is received from the network node based at least in part on the indication of the inference accuracy level.
