Processing timeline considerations for channel state information

Abstract

A first wireless device, such as a user equipment, generates a message indicating a processing time for at least one of: training a neural network for channel state information (CSI) derivation or reporting CSI based on the trained neural network. The first wireless device transmits the message indicating the processing time to a second wireless device. The second wireless device may be a network entity, such as a base station, a transmission reception point, or another UE.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit and priority of Greek patent application serial number SN20200100496, entitled “Processing Timeline Considerations for Channel State Information,” filed on August 18, 2021, which application is expressly incorporated herein by reference in its entirety. Technical Field

[0003] The present disclosure relates generally to communication systems, and more particularly to encoding data sets using the operation of a neural network.

[0004] introduction

[0005] Wireless communication systems are widely deployed to provide various telecommunication services such as telephony, video, data, messaging, and broadcast. Typical wireless communication systems may employ multiple access technologies capable of supporting communication with multiple users by sharing available system resources. Examples of such multiple access technologies 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.

[0006] These multiple access technologies have been adopted in various telecommunication standards to provide common protocols that enable different wireless devices to communicate at city, country, region, and even global levels. An example telecommunication standard is 5G New Radio (NR). 5G NR is a part of the continuous mobile broadband evolution promulgated by the Third Generation Partnership Project (3GPP) to meet new requirements associated with latency, reliability, security, scalability (e.g., with the Internet of Things (IoT)) and other requirements. 5GNR includes services associated with enhanced mobile broadband (eMBB), massive machine type communication (mMTC), and ultra-reliable low latency communication (URLLC). Some aspects of 5G NR can be based on 4G Long Term Evolution (LTE) standards. There is a need for further improvements to 5G NR technology. These improvements may also be applicable to other multiple access technologies and telecommunication standards that employ these technologies.

[0007] Brief Overview

[0008] A brief summary of one or more aspects is given below to provide a basic understanding of such aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or critical elements of all aspects nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be presented later.

[0009] In one aspect of the present disclosure, a wireless communication method for wireless communication is provided. The method includes: generating a message indicating a processing time at a user equipment (UE) for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; and transmitting the message indicating the processing time to a first network entity.

[0010] In another aspect of the present disclosure, an apparatus for wireless communication is provided. The apparatus includes means for generating a message indicating a processing time at a UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; and means for transmitting the message indicating the processing time to a first network entity.

[0011] In another aspect of the present disclosure, an apparatus for wireless communication is provided. The apparatus includes a memory and at least one processor coupled to the memory, the memory and the at least one processor being configured to: generate a message indicating a processing time at a UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; and transmit the message indicating the processing time to a first network entity.

[0012] In another aspect of the present disclosure, a computer-readable medium storing computer-executable code for wireless communication at a UE is provided, which, when executed by a processor, causes the processor to: generate a message indicating a processing time at the UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; and transmit the message indicating the processing time to a first network entity.

[0013] In another aspect of the present disclosure, a method for wireless communication is provided. The method includes: receiving a processing time from a user equipment (UE) for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; transmitting a configuration to the UE based on the processing time; and receiving CSI from the UE based on the configuration.

[0014] In another aspect of the present disclosure, an apparatus for wireless communication is provided. The apparatus includes a device for receiving a processing time from a UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; a device for transmitting a configuration to the UE based on the processing time; and a device for receiving CSI from the UE based on the configuration.

[0015] In another aspect of the present disclosure, an apparatus for wireless communication is provided. The apparatus includes a memory and at least one processor coupled to the memory, the memory and the at least one processor being configured to: receive a processing time from a UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; transmitting a configuration to the UE based on the processing time; and receiving CSI from the UE based on the configuration.

[0016] In another aspect of the present disclosure, a computer-readable medium storing computer-executable code for wireless communication at a UE is provided, which, when executed by a processor, causes the processor to: receive from the UE a processing time for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; transmit a configuration to the UE based on the processing time; and receive CSI from the UE based on the configuration.

[0017] To achieve the foregoing and related ends, the one or more aspects include the features fully described below and particularly pointed out in the claims. The following description and drawings set forth in detail certain illustrative features of the one or more aspects. However, these features are merely indicative of several of the various ways in which the principles of the various aspects may be employed, and this description is intended to cover all such aspects and their equivalents. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a diagram illustrating an example of a wireless communication system and an access network.

[0020] Figure 2A is a diagram illustrating an example of a first frame according to various aspects of the present disclosure.

[0021] Figure 2B is a diagram illustrating an example of DL channels within a subframe according to various aspects of the present disclosure.

[0022] Figure 2C is a diagram illustrating an example of a second frame according to various aspects of the present disclosure.

[0023] Figure 2D is a diagram illustrating an example of UL channels within a subframe according to various aspects of the present disclosure.

[0024] Figure 3 is a diagram illustrating an example of a base station and a user equipment (UE) in an access network.

[0025] Figure 4A is a diagram illustrating an example of an encoding device and a decoding device using previously stored channel state information according to various aspects of the present disclosure.

[0026] Figure 4B is a diagram illustrating an example associated with an encoding device and a decoding device according to various aspects of the present disclosure.

[0027] Figure 5-Figure 8 is a diagram illustrating examples associated with using a neural network to encode and decode a data set for uplink communications in accordance with various aspects of the present disclosure.

[0028] Fig. 9 and Fig.10 is a diagram illustrating example processes associated with using a neural network to encode a data set for uplink communication in accordance with various aspects of the present disclosure.

[0029] Fig.11 Example timing for channel state information (CSI) measurements is illustrated.

[0030] Fig.12 Example timing of configurations for CSI measurements is illustrated.

[0031] Fig.13 An example communication flow between a UE and a network device is illustrated.

[0032] Fig.14A and 14B is a flow chart of a wireless communication method that includes transmitting information about processing time for training a neural network for CSI derivation or reporting CSI based on the trained neural network.

[0033] Fig.15 is a diagram illustrating an example of a hardware implementation for an example device configured to transmit information regarding processing time for training a neural network for CSI derivation or reporting CSI based on the trained neural network.

[0034] Fig.16 is a flow chart of a wireless communication method that includes receiving information about a processing time for training a neural network for CSI derivation or reporting CSI based on the trained neural network.

[0035] Fig.17is a diagram illustrating an example of a hardware implementation for an example device configured to receive information regarding processing time for training a neural network for CSI derivation or reporting CSI based on the trained neural network.

[0036] Detailed Description

[0037] A coding device operating in a network may measure reference signals, etc., to report to a network entity. For example, the coding device may measure reference signals during a beam management process to implement channel state feedback (CSF), may measure the received power of reference signals from a serving cell and / or a neighbor cell, may measure signal strength of an inter-radio access technology (e.g., WiFi) network, may measure sensor signals for detecting the location of one or more objects within an environment, etc. However, reporting this information may consume communication and / or network resources.

[0038] In some aspects described herein, a coding device (e.g., a UE, a base station, a transmit reception point (TRP), a network device, a low earth orbit (LEO) satellite, a medium earth orbit (MEO) satellite, a geostationary orbit (GEO) satellite, a high elliptical orbit (HEO) satellite, etc.) may train one or more neural networks to learn the dependence of each measured quality on an individual parameter, isolate the measured qualities through various layers (also referred to as "operations") of the one or more neural networks, and compress the measurements in a manner that limits compression losses. In some aspects, the coding device may use the properties of the number of bits being compressed to construct a process for extracting and compressing each feature (also referred to as a dimension) that affects the number of bits. In some aspects, the number of bits may be associated with sampling of one or more reference signals and / or may indicate channel state information. For example, the coding device may encode the measurements using one or more extraction operations and compression operations associated with the neural network to produce compressed measurements, wherein the one or more extraction operations and compression operations are based at least in part on a feature set of the measurements.

[0039] The encoding device may transmit the compressed measurements to a network entity such as a server, a TRP, another UE, a base station, etc. Although the examples described herein cite a base station as a decoding device, the decoding device may be any network entity. A network entity may be referred to herein as a "decoding device".

[0040] The decoding device may decode the compressed measurement using one or more decompression operations and reconstruction operations associated with the neural network. The one or more decompression and reconstruction operations may be based at least in part on a feature set of the compressed data set to produce a reconstructed measurement. The decoding device may use the reconstructed measurement as channel state information feedback.

[0041] In some aspects, the UE may determine a processing time for training a neural network and / or reporting CSI using a trained neural network, and may provide the processing time to a network entity. The UE may then receive a configuration for training the neural network and / or for reporting CSI based on the processing time. For example, the UE may determine the processing time based on parameters of the neural network, such as the number of layers of the neural network, the number of weights of the neural network, the type of layers of the neural network, whether it is a previously trained neural network, whether multiple neural networks are to be trained simultaneously, the procedure for the neural network to be trained, the requested accuracy, the number of layers to be trained within the layers of the neural network, the sequence of layers. The UE may determine the processing time after training based on any of the following: encoder output vectors, encoder input vectors, vector combinations, the number of layers of the neural network, the number of elements in input / output / intermediate vectors, the type of layers in the neural network, or the sequence of layers.

[0042] The UE may report training and / or processing capabilities per layer, per layer type, per layer combination, per input / output / intermediate vector length, per number of layers, or per sequence of layers. The UE may report different levels of processing times, e.g., slower times and faster times, for the same neural network, same layer, same combination of layers, or same sequence of layers. The UE may then receive a configuration to use one of these processing times.

[0043] The detailed description set forth below in conjunction with the accompanying drawings is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein may be practiced. This detailed description includes specific details to provide a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form in order to avoid diluting such concepts.

[0044] Several aspects of telecommunication systems will now be presented with reference to various apparatus and methods. These apparatus and methods will be described in the following detailed description and illustrated in the accompanying drawings by various blocks, components, circuits, processes, algorithms, etc. (collectively referred to as "elements"). These elements may be implemented using electronic hardware, computer software, or any combination thereof. Whether such elements are implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system.

[0045] As an example, an element, or any part of an element, or any combination of elements may be implemented as a "processing system" including one or more processors. Examples of processors include: microprocessors, microcontrollers, graphics processing units (GPUs), central processing units (CPUs), application processors, digital signal processors (DSPs), reduced instruction set computing (RISC) processors, systems on chip (SoCs), baseband processors, field programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gating logic, discrete hardware circuits, and other suitable hardware configured to perform various functionalities described throughout this disclosure. One or more processors in a processing system may execute software. Software should be broadly interpreted as meaning instructions, instruction sets, codes, code segments, program codes, programs, subroutines, software components, applications, software applications, software packages, routines, subroutines, objects, executables, threads of execution, procedures, functions, etc., whether it is described in software, firmware, middleware, microcode, hardware description languages, or other terms.

[0046] Accordingly, in one or more example embodiments, the functions described may be implemented in hardware, software, or any combination thereof. If implemented in software, each function may be stored or encoded on a computer-readable medium as one or more instructions or codes. Computer-readable media include computer storage media. Storage media may be any available medium that can be accessed by a computer. As an example and not limitation, such computer-readable media may include random access memory (RAM), read-only memory (ROM), electrically erasable programmable ROM (EEPROM), optical disk storage, magnetic disk storage, other magnetic storage devices, combinations of these types of computer-readable media, or any other medium that can be used to store computer-accessible instructions or data structure forms of computer executable code.

[0047] Although various aspects are described in this application by explanation of some examples, it will be understood by those skilled in the art that additional implementations and use cases can be generated in many different arrangements and scenarios. The various aspects described herein can be implemented across many different platform types, devices, systems, shapes, sizes, and packaging arrangements. For example, each implementation and / or use can be generated via an integrated chip implementation and other devices based on non-module components (e.g., end-user devices, vehicles, communication equipment, computing equipment, industrial equipment, retail / shopping equipment, medical equipment, devices that enable artificial intelligence (AI), etc.). Although some examples may or may not be specifically for each use case or application, the wide applicability of the described aspects may occur. The scope of each implementation can range from chip-level or module components to non-module, non-chip-level implementations, and further to the aggregation, distributed or original equipment manufacturer (OEM) equipment or system that incorporates one or more aspects of the described technology. In some actual environments, the equipment incorporating the various aspects and features described may also include additional components and features for implementing and practicing the claimed and described aspects. For example, the transmission and reception of wireless signals necessarily include several components for analog and digital purposes (e.g., hardware components, including antennas, RF chains, power amplifiers, modulators, buffers, processor(s), interleavers, adders / summers, etc.). The various aspects described herein are intended to be practiced in a wide variety of devices of various sizes, shapes, and configurations, chip-level components, systems, distributed arrangements, aggregated or disaggregated components (e.g., associated with user equipment (UE) and / or base stations), end-user devices, etc.

[0048] Figure 1 1 is a diagram illustrating an example of a wireless communication system and access network 100. The wireless communication system (also referred to as a wireless wide area network (WWAN)) includes a base station 102, a UE 104, an evolved packet core (EPC) 160, and another core network 190 (e.g., a 5G core (5GC)). The base station 102 may include a macro cell (a high-power cellular base station) and / or a small cell (a low-power cellular base station). A macro cell includes a base station. A small cell includes a femto cell, a pico cell, and a micro cell.

[0049] The base station 102 configured for 4G LTE (collectively referred to as the Evolved Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access Network (E-UTRAN)) can interface with the EPC 160 via a first backhaul link 132 (e.g., an S1 interface). The base station 102 configured for 5G NR (collectively referred to as the Next Generation RAN (NG-RAN)) can interface with the core network 190 via a second backhaul link 184. Among other functions, the base station 102 can also perform one or more of the following functions: delivery of user data, radio channel encryption and decryption, integrity protection, header compression, mobility control functions (e.g., handover, dual connectivity), inter-cell interference coordination, connection establishment and release, load balancing, distribution of non-access stratum (NAS) messages, NAS node selection, synchronization, radio access network (RAN) sharing, multimedia broadcast multicast service (MBMS), subscriber and equipment tracking, RAN information management (RIM), paging, positioning, and delivery of alert messages. The base stations 102 may communicate with each other directly or indirectly (eg, through the EPC 160 or the core network 190) over a third backhaul link 134 (eg, an X2 interface). The first backhaul link 132, the second backhaul link 184, and the third backhaul link 134 may be wired or wireless.

[0050] Base stations 102 may communicate wirelessly with UEs 104. Each base station 102 may provide communication coverage for a respective geographic coverage area 110. There may be overlapping geographic coverage areas 110. For example, a small cell 102' may have a coverage area 110' that overlaps with the coverage area 110 of one or more macro base stations 102. A network that includes both small cells and macro cells may be referred to as a heterogeneous network. A heterogeneous network may also include a home evolved Node B (eNB) (HeNB), which may provide services to a restricted group referred to as a closed subscriber group (CSG). A communication link 120 between a base station 102 and a UE 104 may include an uplink (UL) (also referred to as a reverse link) transmission from the UE 104 to the base station 102 and / or a downlink (DL) (also referred to as a forward link) transmission from the base station 102 to the UE 104. The communication link 120 may use multiple-input multiple-output (MIMO) antenna technology, including spatial multiplexing, beamforming, and / or transmit diversity. These communication links may be over one or more carriers. For each carrier allocated in a carrier aggregation of up to Yx MHz (x component carriers) for transmission in each direction, the base station 102 / UE 104 may use a spectrum of up to Y MHz (e.g., 5, 10, 15, 20, 100, 400 MHz, etc.) bandwidth. These carriers may or may not be adjacent to each other. The allocation of carriers may be asymmetric with respect to DL and UL (e.g., more or fewer carriers may be allocated to DL compared to UL). The component carrier may include a primary component carrier and one or more secondary component carriers. The primary component carrier may be referred to as a primary cell (PCell), and the secondary component carrier may be referred to as a secondary cell (SCell).

[0051] In some aspects, UE 104 may be configured to access a radio cell supported by a non-terrestrial network (NTN) device 107, and network 100 may be referred to as an NTN. NTN device 107 may be referred to as a space-borne vehicle or an airborne vehicle. In some examples, NTN device 107 may be configured to operate as a relay for communication between UE 104 and base station 102 or 180. In such examples, NTN device 107 may be referred to as a transparent payload, and base station 102 or 180 may be referred to as a ground base station. In some examples, NTN device 107 may include an onboard substrate and / or a decomposed base station. In such examples, NTN device 107 may be referred to as a regenerative payload and / or an NTN base station. Feeder link 109 may be provided between NTN device 107 and a gateway device, and service link 111 may be provided between UE 104 and NTN device 107.

[0052] Some UEs 104 may communicate with each other using a device-to-device (D2D) communication link 158. The D2D communication link 158 may use DL / UL WWAN spectrum. The D2D communication link 158 may use one or more sidelink channels, such as a physical sidelink broadcast channel (PSBCH), a physical sidelink discovery channel (PSDCH), a physical sidelink shared channel (PSSCH), and a physical sidelink control channel (PSCCH). The D2D communication may be through a variety of wireless D2D communication systems, such as, for example, WiMedia, Bluetooth, ZigBee, Wi-Fi based on the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standard, LTE, or NR.

[0053] The wireless communication system may further include a Wi-Fi access point (AP) 150 in communication with a Wi-Fi station (STA) 152 via a communication link 154, e.g., in a 5 GHz unlicensed spectrum, etc. When communicating in an unlicensed spectrum, the STA 152 / AP 150 may perform a clear channel assessment (CCA) prior to communication to determine whether the channel is available.

[0054] The small cell 102' may operate in a licensed and / or unlicensed spectrum. When operating in an unlicensed spectrum, the small cell 102' may employ NR and use the same unlicensed spectrum (e.g., 5 GHz, etc.) as used by the Wi-Fi AP 150. The small cell 102' employing NR in the unlicensed spectrum may boost the coverage of the access network and / or increase the capacity of the access network.

[0055] The electromagnetic spectrum is typically subdivided into various classes, bands, channels, etc. based on frequency / wavelength. In 5G NR, two initial operating bands have been identified as frequency range designations FR1 (410 MHz–7.125 GHz) and FR2 (24.25 GHz–52.6 GHz). Frequencies between FR1 and FR2 are typically referred to as mid-band frequencies. Although a portion of FR1 is greater than 6 GHz, FR1 is typically (interchangeably) referred to as the “sub-6 GHz band” in various documents and articles. Similar naming issues sometimes arise regarding FR2, which is typically (interchangeably) referred to as the “millimeter wave” band in various documents and articles, although it is different from the extremely high frequency (EHF) band (30 GHz–300 GHz) identified as the “millimeter wave” band by the International Telecommunication Union (ITU).

[0056] Frequencies between FR1 and FR2 are generally referred to as mid-band frequencies. Recent 5G NR research has identified the operating bands of these mid-band frequencies as frequency range designation FR3 (7.125GHz–24.25GHz). The frequency bands falling within FR3 can inherit FR1 characteristics and / or FR2 characteristics, and thus the features of FR1 and / or FR2 can be effectively extended to mid-band frequencies. In addition, higher frequency bands are currently being explored to extend 5G NR operations to above 52.6GHz. For example, three higher operating bands have been identified as frequency range designations FR4a or FR4-1 (52.6GHz–71GHz), FR4 (52.6GHz–114.25GHz), and FR5 (114.25GHz–300GHz). Each of these higher frequency bands falls within the EHF band.

[0057] In view of the above aspects, unless otherwise specifically stated, it should be understood that if used herein, the term sub-"6 GHz" and the like can broadly refer to frequencies that can be less than 6 GHz, can be within FR1, or can include mid-band frequencies. In addition, unless otherwise specifically stated, it should be understood that if used herein, the term "millimeter wave" and the like can broadly refer to frequencies that can include mid-band frequencies, can be within FR2, FR4, FR4-a or FR4-1 and / or FR5, or can be within the EHF band.

[0058] Whether a small cell 102' or a large cell (e.g., a macro base station), the base station 102 may include and / or be referred to as an eNB, a gB node (gNB), or another type of base station. Some base stations (such as gNB 180) may operate in the traditional sub-6 GHz spectrum, in millimeter wave frequencies, and / or near millimeter wave frequencies to communicate with UE 104. When the gNB 180 operates in millimeter wave frequencies or near millimeter wave frequencies, the gNB 180 may be referred to as a millimeter wave base station. The millimeter wave base station 180 may utilize beamforming 182 with the UE 104 to compensate for path loss and short range. The base station 180 and the UE 104 may each include multiple antennas, such as antenna elements, antenna panels, and / or antenna arrays to facilitate beamforming.

[0059] Base station 180 may transmit beamformed signals to UE 104 in one or more transmit directions 182'. UE 104 may receive beamformed signals from base station 180 in one or more receive directions 182". UE 104 may also transmit beamformed signals to base station 180 in one or more transmit directions. Base station 180 may receive beamformed signals from UE 104 in one or more receive directions. Base station 180 / UE 104 may perform beam training to determine the best receive direction and transmit direction for each of base station 180 / UE 104. The transmit direction and receive direction of base station 180 may be the same or may be different. The transmit direction and receive direction of UE 104 may be the same or may be different.

[0060] The EPC 160 may include a mobility management entity (MME) 162, other MMEs 164, a serving gateway 166, a multimedia broadcast multicast service (MBMS) gateway 168, a broadcast multicast service center (BM-SC) 170, and a packet data network (PDN) gateway 172. The MME 162 may be in communication with a home subscriber server (HSS) 174. The MME 162 is a control node that handles signaling between the UE 104 and the EPC 160. In general, the MME 162 provides bearer and connection management. All user Internet Protocol (IP) packets are delivered through the serving gateway 166, which itself is connected to the PDN gateway 172. The PDN gateway 172 provides UE IP address allocation and other functions. The PDN gateway 172 and the BM-SC 170 are connected to IP services 176. The IP services 176 may include the Internet, an intranet, an IP multimedia subsystem (IMS), a PS streaming service, and / or other IP services. The BM-SC 170 may provide functionality for MBMS user service provisioning and delivery. The BM-SC 170 may serve as an entry point for content provider MBMS transmissions, may be used to authorize and initiate MBMS bearer services within a public land mobile network (PLMN), and may be used to schedule MBMS transmissions. The MBMS Gateway 168 may be used to distribute MBMS traffic to base stations 102 belonging to a multicast broadcast single frequency network (MBSFN) area broadcasting a specific service, and may be responsible for session management (start / stop) and for collecting eMBMS related charging information.

[0061] The core network 190 may include an access and mobility management function (AMF) 192, other AMFs 193, a session management function (SMF) 194, and a user plane function (UPF) 195. The AMF 192 may be in communication with a unified data management (UDM) 196. The AMF 192 is a control node that handles signaling between the UE 104 and the core network 190. In general, the AMF 192 provides QoS flow and session management. All user Internet Protocol (IP) packets are delivered through the UPF 195. The UPF 195 provides UE IP address allocation and other functions. The UPF 195 is connected to an IP service 197. The IP service 197 may include the Internet, an intranet, an IP multimedia subsystem (IMS), a packet switching (PS) streaming (PSS) service, and / or other IP services.

[0062] A base station may include and / or be referred to as a gNB, a Node B, an eNB, an access point, a base transceiver station, a radio base station, a radio transceiver, a transceiver function, a basic service set (BSS), an extended service set (ESS), a transmission reception point (TRP), or some other suitable term. The base station 102 provides an access point to the EPC 160 or the core network 190 for the UE 104. Examples of UE 104 include a cellular phone, a smart phone, a Session Initiation Protocol (SIP) phone, a laptop, a personal digital assistant (PDA), a satellite radio, a global positioning system, a multimedia device, a video device, a digital audio player (e.g., an MP3 player), a camera, a game console, a tablet device, a smart device, a wearable device, a vehicle, an electric meter, a gas pump, a large or small kitchen appliance, a health care device, an implant, a sensor / actuator, a display, or any other similar functional device. Some UEs 104 may be referred to as IoT devices (e.g., parking meters, gas pumps, ovens, vehicles, heart monitors, etc.). UE 104 may also be referred to as a station, a mobile station, a subscriber station, a mobile unit, a subscriber unit, a wireless unit, a remote unit, a mobile device, a wireless device, a wireless communication device, a remote device, a mobile subscriber station, an access terminal, a mobile terminal, a wireless terminal, a remote terminal, a handset, a user agent, a mobile client, a client, or some other suitable terminology.

[0063] Reference again Figure 1 In certain aspects, a UE 104 may include a neural network capability / configuration component 198 configured to determine a processing time for at least one of: training a neural network for CSI derivation or reporting CSI based on a trained neural network; and providing the processing time to a network entity (such as a base station 102 or 180, a transmit reception point (TRP) 103, or another UE 104).

[0064] A network entity (such as base station 102 or 180, TRP 103, or UE 104) may include a neural network configuration component 199 configured to: receive a processing time from a UE for at least one of: training a neural network for CSF or reporting CSF based on the trained neural network; transmit a configuration to the UE based on the processing time; and receive CSI from the UE based on the configuration.

[0065] Although the following description may focus on 5G NR, the concepts described herein may be applicable to other similar areas such as LTE, LTE-A, CDMA, GSM, and other wireless technologies.

[0066] Figure 2A 200 is a diagram illustrating an example of a first subframe within a 5G NR frame structure. Figure 2B is a diagram 230 illustrating an example of DL channels within a 5G NR subframe. Figure 2C 250 is a diagram illustrating an example of a second subframe within a 5G NR frame structure. Figure 2D 280 is a diagram illustrating an example of UL channels within a 5G NR subframe. The 5G NR frame structure can be frequency division duplex (FDD), where for a particular set of subcarriers (carrier system bandwidth), subframes within the subcarrier set are dedicated to DL or UL; or can be time division duplex (TDD), where for a particular set of subcarriers (carrier system bandwidth), subframes within the subcarrier set are dedicated to both DL and UL. Figure 2A , 2C In the example provided, the 5G NR frame structure is assumed to be TDD, where subframe 4 is configured with slot format 28 (mostly DL) and subframe 3 is configured with slot format 1 (all UL), where D is DL, U is UL, and F is for flexible use between DL / UL. Although subframes 3 and 4 are shown as having slot formats 1 and 28, respectively, any particular subframe may be configured with any of the various available slot formats 0-61. Slot formats 0 and 1 are full DL and full UL, respectively. Other slot formats 2-61 include a mixture of DL, UL, and flexible symbols. The UE is configured with a slot format (dynamically configured by DL control information (DCI) or semi-statically / statically configured by radio resource control (RRC) signaling) through the received slot format indicator (SFI). Note that the following description also applies to the 5GNR frame structure for TDD.

[0067] Figure 2A-2DThe frame structure is explained, and various aspects of the present disclosure may be applicable to other wireless communication technologies that may have different frame structures and / or different channels. A frame (10ms) may be divided into 10 equally sized subframes (1ms). Each subframe may include one or more time slots. A subframe may also include a mini-time slot, which may include 7, 4, or 2 symbols. Each time slot may include 14 or 12 symbols, depending on whether the cyclic prefix (CP) is a normal CP or an extended CP. For a normal CP, each time slot may include 14 symbols, and for an extended CP, each time slot may include 12 symbols. The symbols on the DL may be CP orthogonal frequency division multiplexing (OFDM) (CP-OFDM) symbols. The symbols on the UL may be CP-OFDM symbols (for high throughput scenarios) or discrete Fourier transform (DFT) extended OFDM (DFT-s-OFDM) symbols (also known as single carrier frequency division multiple access (SC-FDMA) symbols) (for power-limited scenarios; limited to single stream transmission). The number of slots within a subframe is based on the CP and parameter design. The parameter design defines the subcarrier spacing (SCS) and, in effect, the symbol length / duration, which is equal to 1 / SCS.

[0068]

[0069] For normal CP (14 symbols / slot), different parameter designs μ0 to 4 allow 1, 2, 4, 8, and 16 slots per subframe, respectively. For extended CP, parameter design 2 allows 4 slots per subframe. Accordingly, for normal CP and parameter design μ, there are 14 symbols / slot and 2 μ time slots / subframe. The subcarrier spacing can be equal to 2 μ *15kHz, where μ is parameter design 0 to 4. Thus, parameter design μ=0 has a subcarrier spacing of 15kHz, while parameter design μ=4 has a subcarrier spacing of 240kHz. The symbol length / duration is inversely related to the subcarrier spacing. FIG. 2A to FIG. 2D An example of a parameter design μ=2 with a normal CP of 14 symbols per slot and 4 slots per subframe is provided. The slot duration is 0.25 ms, the subcarrier spacing is 60 kHz, and the symbol duration is about 16.67 μs. Within a frame set, there may be one or more different bandwidth parts (BWPs) that are frequency division multiplexed (see Figure 2B ). Each BWP can have a specific parameter design and CP (normal or extended).

[0070] A resource grid may be used to represent the frame structure. Each slot includes a resource block (RB) (also called a physical RB (PRB)) extending over 12 consecutive subcarriers. The resource grid is divided into a number of resource elements (REs). The number of bits carried by each RE depends on the modulation scheme.

[0071] like Figure 2A As illustrated in FIG, some REs carry reference (pilot) signals (RS) for UEs. RSs may include demodulation RSs (DM-RSs) (indicated as R for one particular configuration, but other DM-RS configurations are possible) and channel state information reference signals (CSI-RSs) for channel estimation at the UE. RSs may also include beam measurement RSs (BRSs), beam refinement RSs (BRRSs), and phase tracking RSs (PT-RSs).

[0072] Figure 2B Examples of various DL channels within a subframe of a frame are illustrated. A physical downlink control channel (PDCCH) carries DCI within one or more control channel elements (CCEs) (e.g., 1, 2, 4, 8, or 16 CCEs), each CCE including 6 RE groups (REGs), each REG including 12 consecutive REs in an OFDM symbol of an RB. A PDCCH within a BWP may be referred to as a control resource set (CORESET). The UE is configured to monitor PDCCH candidates in a PDCCH search space (e.g., a common search space, a search space that varies with the UE) during a PDCCH monitoring opportunity on a CORESET, wherein the PDCCH candidates have different DCI formats and different aggregation levels. Additional BWPs may be located at higher and / or lower frequencies across the channel bandwidth. A primary synchronization signal (PSS) may be within symbol 2 of a particular subframe of a frame. PSS is used by UE104 to determine subframe / symbol timing and physical layer identity. A secondary synchronization signal (SSS) may be within symbol 4 of a particular subframe of a frame. The SSS is used by the UE to determine the physical layer cell identity group number and the radio frame timing. Based on the physical layer identity and the physical layer cell identity group number, the UE can determine the physical cell identifier (PCI). Based on the PCI, the UE can determine the location of the DM-RS. The physical broadcast channel (PBCH) carrying the master information block (MIB) can be logically grouped with the PSS and SSS to form a synchronization signal (SS) / PBCH block (also referred to as an SS block (SSB)). The MIB provides the number of RBs in the system bandwidth, and the system frame number (SFN). The physical downlink shared channel (PDSCH) carries user data, broadcast system information not transmitted through the PBCH (such as system information blocks (SIBs)), and paging messages.

[0073] As in Figure 2CAs explained in , some REs carry DM-RSs for channel estimation at the base station (indicated as R for one specific configuration, but other DM-RS configurations are possible). The UE may transmit DM-RSs for the physical uplink control channel (PUCCH) and DM-RSs for the physical uplink shared channel (PUSCH). The PUSCH DM-RS may be transmitted in the first or first two symbols of the PUSCH. The PUCCH DM-RS may be transmitted in different configurations depending on whether a short PUCCH or a long PUCCH is transmitted and depending on the specific PUCCH format used. The UE may transmit a sounding reference signal (SRS). The SRS may be transmitted in the last symbol of the subframe. The SRS may have a comb structure, and the UE may transmit the SRS on one of the comb teeth. The SRS may be used by the base station for channel quality estimation to enable frequency-dependent scheduling on the UL.

[0074] Figure 2D Examples of various UL channels within a subframe of a frame are illustrated. The PUCCH may be located at a position as indicated in one configuration. The PUCCH carries uplink control information (UCI), such as a scheduling request, a channel quality indicator (CQI), a precoding matrix indicator (PMI), a rank indicator (RI), and a hybrid automatic repeat request (HARQ) acknowledgement (ACK) (HARQ-ACK) feedback (i.e., one or more HARQ ACK bits indicating one or more ACKs and / or negative ACKs (NACKs)). The PUSCH carries data and may additionally be used to carry a buffer status report (BSR), a power headroom report (PHR), and / or UCI.

[0075] Figure 33 is a block diagram of a base station 310 in communication with a UE 350 in an access network. In the DL, IP packets from the EPC 160 may be provided to a controller / processor 375. The controller / processor 375 implements layer 3 and layer 2 functionality. Layer 3 includes a radio resource control (RRC) layer, and layer 2 includes a service data adaptation protocol (SDAP) layer, a packet data convergence protocol (PDCP) layer, a radio link control (RLC) layer, and a media access control (MAC) layer. The controller / processor 375 provides RRC layer functionality associated with broadcasting of system information (e.g., MIB, SIB), RRC connection control (e.g., RRC connection paging, RRC connection establishment, RRC connection modification, and RRC connection release), inter-radio access technology (RAT) mobility, and measurement configuration of UE measurement reports; PDCP layer functionality associated with header compression / decompression, security (ciphering, cipher decryption, integrity protection, integrity verification), and handover support functions; RLC layer functionality associated with delivery of upper layer packet data units (PDUs), error correction through ARQ, concatenation, segmentation and reassembly of RLC service data units (SDUs), re-segmentation of RLC data PDUs, and reordering of RLC data PDUs; and MAC layer functionality associated with mapping between logical channels and transport channels, multiplexing of MAC SDUs onto transport blocks (TBs), demultiplexing of MAC SDUs from TBs, scheduling information reporting, error correction through HARQ, priority handling, and logical channel prioritization.

[0076] The transmit (TX) processor 316 and the receive (RX) processor 370 implement layer 1 functionality associated with various signal processing functions. Layer 1, including the physical (PHY) layer, may include error detection on the transmission channel, forward error correction (FEC) encoding / decoding of the transmission channel, interleaving, rate matching, mapping to the physical channel, modulation / demodulation of the physical channel, and MIMO antenna processing. The TX processor 316 handles the mapping to the signal constellation based on various modulation schemes (e.g., binary phase shift keying (BPSK), quadrature phase shift keying (QPSK), M-phase shift keying (M-PSK), M-quadrature amplitude modulation (M-QAM)). The coded and modulated codewords can then be split into parallel streams. Each stream can then be mapped to an OFDM subcarrier, multiplexed with a reference signal (e.g., pilot) in the time domain and / or frequency domain, and then combined together using an inverse fast Fourier transform (IFFT) to produce a physical channel carrying a time domain OFDM codeword stream. The OFDM stream is spatially precoded to produce multiple spatial streams. Channel estimates from a channel estimator 374 may be used to determine coding and modulation schemes and for spatial processing. The channel estimate may be derived from a reference signal and / or channel condition feedback transmitted by the UE 350. Each spatial stream may then be provided to a different antenna 320 via a separate transmitter 318TX. Each transmitter 318TX may modulate a radio frequency (RF) carrier with a corresponding spatial stream for transmission.

[0077] At the UE 350, each receiver 354RX receives a signal through its corresponding antenna 352. Each receiver 354RX recovers the information modulated onto the RF carrier and provides the information to a receive (RX) processor 356. The TX processor 368 and the RX processor 356 implement layer 1 functionality associated with various signal processing functions. The RX processor 356 can perform spatial processing on the information to recover any spatial stream destined for the UE 350. If there are multiple spatial streams destined for the UE 350, they can be combined into a single OFDM symbol stream by the RX processor 356. The RX processor 356 then transforms the OFDM symbol stream from the time domain to the frequency domain using a fast Fourier transform (FFT). The frequency domain signal includes a separate OFDM symbol stream for each subcarrier of the OFDM signal. The symbol on each subcarrier and the reference signal are recovered and demodulated by determining the signal constellation point most likely transmitted by the base station 310. These soft decisions can be based on the channel estimate calculated by the channel estimator 358. These soft decisions are then decoded and deinterleaved to recover the data and control signals originally transmitted on the physical channel by the base station 310. These data and control signals are then provided to the controller / processor 359 which implements layer 3 and layer 2 functionality.

[0078] The controller / processor 359 may be associated with a memory 360 that stores program codes and data. The memory 360 may be referred to as a computer readable medium. In the UL, the controller / processor 359 provides demultiplexing between transport channels and logical channels, packet reassembly, cipher interpretation, header decompression, and control signal processing to recover IP packets from the EPC 160. The controller / processor 359 is also responsible for error detection using ACK and / or NACK protocols to support HARQ operations.

[0079] Similar to the functionality described in conjunction with DL transmissions performed by the base station 310, the controller / processor 359 provides RRC layer functionality associated with system information (e.g., MIB, SIB) acquisition, RRC connection, and measurement reporting; PDCP layer functionality associated with header compression / decompression, and security (encryption, decryption, integrity protection, integrity verification); RLC layer functionality associated with delivery of upper layer PDUs, error correction through ARQ, concatenation, segmentation, and reassembly of RLC SDUs, resegmentation of RLC data PDUs, and reordering of RLC data PDUs; and MAC layer functionality associated with mapping between logical channels and transport channels, multiplexing MAC SDUs onto TBs, demultiplexing MAC SDUs from TBs, scheduling information reporting, error correction through HARQ, priority handling, and logical channel prioritization.

[0080] Channel estimates derived by a channel estimator 358 from a reference signal or feedback transmitted by the base station 310 may be used by a TX processor 368 to select the appropriate coding and modulation schemes, and to facilitate spatial processing. The spatial streams generated by the TX processor 368 may be provided to different antennas 352 via separate transmitters 354TX. Each transmitter 354TX may modulate an RF carrier with a corresponding spatial stream for transmission.

[0081] UL transmissions are processed at the base station 310 in a manner similar to that described in conjunction with the receiver functionality at the UE 350. Each receiver 318RX receives a signal through its respective antenna 320. Each receiver 318RX recovers information modulated onto an RF carrier and provides the information to a RX processor 370.

[0082] The controller / processor 375 may be associated with a memory 376 that stores program codes and data. The memory 376 may be referred to as a computer readable medium. In the UL, the controller / processor 375 provides demultiplexing between transport channels and logical channels, packet reassembly, cipher decoding, header decompression, control signal processing to recover IP packets from the UE 350. The IP packets from the controller / processor 375 may be provided to the EPC 160. The controller / processor 375 is also responsible for error detection using ACK and / or NACK protocols to support HARQ operations.

[0083] At least one of the TX processor 368, the RX processor 356, and the controller / processor 359 may include a neural network capability / configuration component 198 configured to determine a processing time for at least one of: training a neural network for CSI derivation or reporting CSI based on a trained neural network; and providing the processing time to a network entity, such as in conjunction with Figure 1 As described.

[0084] At least one of the TX processor 316, the RX processor 370, and the controller / processor 375 may include a neural network configuration component 199 configured to receive a processing time from a UE for at least one of: training a neural network for CSF or reporting CSF based on the trained neural network; transmitting a configuration to the UE based on the processing time; and receiving CSI from the UE based on the configuration, such as in conjunction with Figure 1 As described.

[0085] Some aspects of wireless communication may be based on neural networks or machine learning. As an example, a UE may include a neural network component or a machine learning component. In other examples, a base station, a TRP, or another network component may include a neural network or a machine learning component. The UE and / or base station (e.g., including a (CU) and / or a distributed unit (DU)) may use machine learning algorithms, deep learning algorithms, neural networks, reinforcement learning, regression, boosting, or advanced signal processing methods for various aspects of wireless communication, such as with a base station, a TRP, another UE, etc.

[0086] In particular, examples of machine learning models or neural networks that may be included in a UE, TRP, base station, or network component include: artificial neural networks (ANN); decision tree learning; convolutional neural networks (CNN); deep learning architectures in which the outputs of a first layer of neurons become the inputs of a second layer of neurons, and so on; support vector machines (SVMs), for example, including a separating hyperplane (e.g., a decision boundary) for categorical data; regression analysis; Bayesian networks; genetic algorithms; deep convolutional networks (DCNs) configured with additional pooling and normalization layers; and deep belief networks (DBNs).

[0087] A machine learning model, such as an artificial neural network (ANN), may include a group of interconnected artificial neurons (e.g., a neuron model) and may be a computing device or may represent a method to be performed by a computing device. The connections of the neuron model may be modeled as weights. The machine learning model may provide predictive modeling, adaptive control, and other applications by training via a data set. The model may be adaptive based on external or internal information processed by the machine learning model. Machine learning may provide nonlinear statistical data models or decision making, and may model complex relationships between input data and output information.

[0088] The machine learning model may include multiple layers and / or operations, which may be formed by cascading one or more of the cited operations. Examples of operations that may be involved include extraction of various features of data, convolution operations, fully connected operations that can be activated or deactivated, compression, decompression, quantization, flattening, etc. As used herein, the "layer" of the machine learning model can be used to represent operations on input data. For example, convolution layers, fully connected layers, etc. can be used to refer to associated operations on data input to the layer. Convolution AxB operations refer to operations that convert several input features A into several output features B. "Kernel size" can refer to the number of adjacent coefficients combined in one dimension. As used herein, "weights" can be used to represent one or more coefficients used in operations for combining various rows and / or columns of input data in each layer. For example, a fully connected layer operation can have an output y, which is determined at least in part based on the product of the input matrix x and the weight A (which can be a matrix) and the sum of the bias value B (which can be a matrix). The term "weight" can be used herein to generally refer to both weights and bias values. Weights and biases are examples of parameters of a trained machine learning model. Different layers of a machine learning model can be trained independently.

[0089] Machine learning models can include various connectivity patterns, for example, including any of feedforward networks, hierarchies, recursive architectures, feedback connections, etc. The connections between the layers of a neural network can be fully connected or partially connected. In a fully connected network, a neuron in a first layer can communicate its output to each neuron in a second layer, and each neuron in a second layer can receive input from each neuron in the first layer. In a locally connected network, a neuron in a first layer can be connected to a limited number of neurons in a second layer. In some aspects, a convolutional network can be locally connected and configured with a shared connection strength associated with the input of each neuron in the second layer. The locally connected layers of a network can be configured so that each neuron in the layer has the same or similar connectivity pattern but with different connection strengths.

[0090] A machine learning model or neural network may be trained. For example, a machine learning model may be trained based on supervised learning. During training, a machine learning model may be provided with inputs that the model uses to calculate to produce an output. The actual output may be compared to the target output, and the difference may be used to adjust the parameters of the machine learning model (such as weights and biases) to provide an output that is closer to the target output. Before training, the output may be incorrect or less accurate, and the error or difference between the actual output and the target output may be calculated. The weights of the machine learning model may then be adjusted so that the output is aligned more closely with the target. In order to adjust the weights, the learning algorithm may calculate a gradient vector for the weights. The gradient may indicate the amount by which the error will increase or decrease if the weights are slightly adjusted. At the top layer, the gradient may directly correspond to the value of the weights connecting the activated neurons in the penultimate layer to the neurons in the output layer. In the lower layers, the gradient may depend on the value of the weights and the calculated error gradient of the higher layers. The weights may then be adjusted to reduce the error or bring the output closer to the target. This way of adjusting the weights may be referred to as back propagation through a neural network. This process may continue until the achievable error rate stops decreasing or until the error rate has reached a target level.

[0091] The machine learning model may include computational complexity and substantial processors for training the machine learning model. A neural network may include a network of interconnected nodes. The output of one node may be connected to another node as an input. The connections between the nodes may be referred to as edges, and weights may be applied to these connections / edges to adjust the output from one node applied as an input to another node. Each node may apply a threshold to determine whether or when to provide an output to a connected node. The output of each node may be calculated as a nonlinear function of the sum of the inputs to the node. A neural network may include any number of nodes and any type of connection between the nodes. A neural network may include one or more hidden nodes. The nodes may be clustered into layers, and different layers of the neural network may perform different kinds of transformations on the inputs. A signal may travel from an input at the first layer through multiple layers of the neural network to an output at the last layer of the neural network, and may traverse each layer multiple times.

[0092] Reinforcement learning is a type of machine learning that involves the concept of taking actions in an environment in order to maximize rewards. Reinforcement learning is a machine learning paradigm; other paradigms include supervised learning and unsupervised learning. Basic reinforcement can be modeled as a Markov decision process (MDP) with an environment and agent state set and an agent action set. The process may include a state transition probability based on an action and a reward representation after the transition. Agent action selection may be modeled as a strategy. Reinforcement learning enables an agent to learn an optimal or near-optimal strategy that maximizes rewards. Supervised learning may include learning a function that maps inputs to outputs based on example input-output pairs, which may be inferred from a training data set (which may be referred to as a training example). A supervised learning algorithm analyzes the training data and provides an algorithm to map to new examples. A federated learning (FL) procedure using an edge device as a client may rely on a client that is being trained based on supervised learning.

[0093] Regression analysis can include a statistical procedure for estimating the relationship between a dependent variable (e.g., which can be referred to as an outcome variable) and the independent variable(s). Linear regression is an example of regression analysis. Nonlinear models can also be used. Regression analysis can include inferring causal relationships between variables in a data set.

[0094] Boosting includes one or more algorithms for reducing bias and / or variance in supervised learning, such as a machine learning algorithm that transforms a weak learner (e.g., a classifier that is weakly correlated with the true classification) into a strong learner (e.g., a classifier that is more closely correlated with the true classification). Boosting may include iterative learning based on weak classifiers relative to a distribution added to a strong classifier. Weak learners may be weighted in relation to accuracy. Data weights may be readjusted through this process. In some aspects described herein, a coding device (e.g., a UE, a base station, or other network component) may train one or more neural networks to learn the dependence of each measured quality on an individual parameter.

[0095] The wireless receiver may provide various types of CSI to the transmitting device. Among other examples, the UE may perform measurements on a downlink signal (such as a reference signal) from a base station and may provide a CSI report including any combination of the following: channel quality indicator (CQI), precoding matrix indicator (PMI), rank indicator (RI), synchronization signal block / physical broadcast channel resource block indicator (SSBRI), layer indicator (LI). The UE may perform measurements and determine CSI based on one or more channel state information reference signals (CSI-RS), SSB, channel state information interference measurement (CSI-IM) resources, etc. received from the base station. The base station may, for example, configure the UE to perform CSI measurements using a CSI measurement configuration. The base station may configure the UE using a CSI resource configuration indicating the type of reference signal, such as a non-zero power CSI-RS (NZP CSI-RS), SSB, CSI-IM resources, etc. The base station may configure the UE with a CSI reporting configuration indicating a mapping between configured CSI measurements and configured CSI resources and indicating that the UE is to provide a CSI report to the base station.

[0096] There may be different types of CSI. A first type of CSI (which may be referred to as Type I CSI) may be used for beam selection, where the UE selects a set of one or more beam indices (e.g., beam 182' or 182") with better channel measurements and transmits CSI information for the beam set to the base station.

[0097] A second type of CSI (which may be referred to as type II CSI) may be used for beam combining of a beam set. The UE may determine optimal linear combination coefficients for individual beams (e.g., beam 182' or 182") and may transmit beam indices for the beam set and coefficients for combining the beams. The UE may provide coefficients for beam combining on a per-subband basis. For example, the UE may provide type II CSI for each configured subband.

[0098] The present application provides a type of CSI that uses machine learning or one or more neural networks to compress a channel and feed the channel back to a base station. The CSI may correspond to an additional type of CSI. The CSI may be referred to as, for example, neural network-based CSI or other names. The CSI may use machine learning or one or more neural networks to measure and provide feedback on interference observed at a UE. The feedback may be provided to a base station, for example, for communication on an access link. In other examples, the feedback may be provided to a transmit receive point (TRP) or another UE (e.g., for side link communication).

[0099] Figure 4AAn example architecture of components of an encoding device 400 and a decoding device 425 using previously stored CSI according to various aspects of the present disclosure is illustrated. In some examples, the encoding device 400 may be a UE (e.g., 104 or 350), and the decoding device 425 may be a base station (e.g., 102, 180, 310), a TRP (e.g., TRP 103), another UE (e.g., UE 104), etc. The encoding device 400 and the decoding device 425 may save and use previously stored CSI, and may encode and decode changes in CSI from previous instances. This may provide less CSI feedback overhead and may improve performance. The encoding device 400 may also be able to encode more accurate CSI, and neural network training may be performed with the more accurate CSI. The example architecture of the encoding device 400 and the decoding device 425 may be used for determination (e.g., calculation) of CSI and provision of feedback from the encoding device 400 to the decoding device 425, including processing based on neural networks or machine learning.

[0100] As illustrated at 402, the encoding device 400 measures downlink channel estimates based on downlink signals (such as CSI-RS, SSB, CSI-IM resources, etc.) input for encoding from a base station. The downlink channel estimate instance at time t is represented as H(t) and is provided to a CSI instance encoder 404, which encodes a single CSI instance at time t and outputs the encoded CSI instance at time t as m(t) to a CSI sequence encoder 406. The CSI sequence encoder 406 can take Doppler into account.

[0101] like Figure 4A As shown in , the CSI instance encoder 404 may encode the CSI instance into intermediate coded CSI for each DL channel estimate in the DL channel estimate sequence. The CSI instance encoder 404 (e.g., a feed-forward network) may use a neural network encoder weight θ. The intermediate coded CSI may be represented as The CSI sequence encoder 406 may be based on a long short-term memory (LSTM) network, while the CSI instance encoder 404 may be based on a feed-forward network. In other examples, the CSI sequence encoder 406 may be based on a gated recurrent unit network or a recurrent unit network. The CSI sequence encoder 406 (e.g., a long short-term memory (LSTM) network) may determine a previously encoded CSI instance h(t-1) from the memory 408 and compare the intermediate encoded CSI m(t) with the previously encoded CSI instance h(t-1) to determine a change n(t) in the encoded CSI. The change n(t) may be a new part of the channel estimate and may not be predicted by the decoding device. The encoded CSI at this time may be obtained by Representation. The CSI sequence encoder 406 may provide this change n(t) on a physical uplink shared channel (PUSCH) or a physical uplink control channel (PUCCH) 410, and the encoding device may transmit the change (e.g., information indicating the change) n(t) as encoded CSI to the decoding device on the UL channel. Because the change is smaller than the entire CSI instance, the encoding device may send a smaller payload for the encoded CSI on the UL channel while including more detailed information of the change in the encoded CSI. The CSI sequence encoder 406 may generate the encoded CSI h(t) based at least in part on the intermediate encoded CSI m(t) and at least a portion of the previous encoded CSI instance h(t-1). The CSI sequence encoder 406 may save the encoded CSI h(t) to the memory 408.

[0102] The CSI sequence decoder 414 may receive the encoded CSI on the PUSCH or PUCCH 412. The CSI sequence decoder 414 may determine that only a change n(t) of the CSI is received as the encoded CSI. The CSI sequence decoder 414 may determine the intermediate decoded CSI m(t) based at least in part on the encoded CSI and at least a portion of a previous intermediate decoded CSI instance h(t-1) from the memory 416 and the change. The CSI instance decoder 418 may decode the intermediate decoded CSI m(t) into decoded CSI 420. The CSI sequence decoder 414 and the CSI instance decoder 418 may use the neural network decoder weights φ. The intermediate decoded CSI may be determined by The CSI sequence decoder 414 may generate decoded CSI h(t) based at least in part on the intermediate decoded CSI m(t) and at least a portion of the previously decoded CSI instance h(t-1). The decoding device may reconstruct the DL channel estimate from the decoded CSI h(t), and the reconstructed channel estimate may be represented as The CSI sequence decoder 414 may save the decoded CSI h(t) into the memory 416 .

[0103] Because the change n(t) is less than the entire CSI instance, the encoding device can send a smaller payload on the UL channel. For example, if the DL channel has little change from the previous feedback due to the low Doppler or less movement of the encoding device, the output of the CSI sequence encoder can be quite compact. In this way, the encoding device 400 can take advantage of the correlation of the channel estimate over time. In some aspects, because the output is small, the encoding device 400 can include more detailed information for the change in the encoded CSI. In some aspects, the encoding device may transmit an indication (e.g., a flag) to the decoding device 425 that the encoded CSI is encoded in time (CSI change). Alternatively, the encoding device 400 may transmit an indication that the encoded CSI is encoded independently of any previously encoded CSI feedback. The decoding device 425 can decode the encoded CSI without using a previously decoded CSI instance. In some aspects, a device (which may include the encoding device 400 or the decoding device 425) can use a CSI sequence encoder and a CSI sequence decoder to train a neural network model.

[0104] In some aspects, CSI may be a function of a channel estimate (referred to as a channel response) H and interference N. There may be a variety of ways to convey H and N. For example, encoding device 400 may encode CSI as N -1 / 2 H. The encoding device 400 may encode H and N separately. The encoding device 400 may partially encode H and N separately, and then jointly encode the two partially encoded outputs. It may be advantageous to encode H and N separately. Interference and channel changes may occur on different time scales. In low Doppler scenarios, the channel may be stable, but interference may still change faster due to traffic or scheduler algorithms. In high Doppler scenarios, the channel may change faster than the scheduler grouping of the UE. In some aspects, a device (which may include the encoding device 400 or the decoding device 425) may use the separately encoded H and N to train a neural network model.

[0105] In some aspects, the reconstructed DL channel The DL channel H may be reflected truthfully, and this may be referred to as explicit feedback. In some aspects, Only the information required for the decoding device 425 to derive the rank and precoding may be captured. CQI may be fed back separately. In a time-coded scenario, CSI feedback may be expressed as m(t) or n(t). Similar to type II CSI feedback, m(t) may be constructed as a concatenation of a rank index (RI), a beam index, and a coefficient representing an amplitude or phase. In some aspects, m(t) may be a quantized version of a real-valued vector. The beam may be predefined (not obtained through training), or may be part of training (e.g., part of θ and φ and communicated to an encoding device or a decoding device).

[0106] In some aspects, the decoding device 425 and the encoding device 400 may maintain multiple encoder and decoder networks, each network targeting a different payload size (to achieve a trade-off of varying accuracy relative to UL overhead). For each CSI feedback, depending on the reconstruction quality and the uplink budget (e.g., PUSCH payload size), the encoding device 400 may select or the decoding device 425 may instruct the encoding device 400 to select one of the encoders to construct the encoded CSI. The encoding device 400 may send an encoder index and CSI based at least in part on the encoder selected by the encoding device. Similarly, the decoding device 425 and the encoding device 400 may maintain multiple encoder and decoder networks to cope with different antenna geometries and channel conditions. Note that although some operations are described for the decoding device 425 and the encoding device 400, these operations may also be performed by another device as part of the preconfiguration of the encoder and decoder weights and / or structures (e.g., parameters 422 or 424).

[0107] As indicated above, Figure 4A A non-limiting example is described to illustrate the concept. Other examples may differ from those described above. Figure 4A What is described and the concepts presented herein may be applied to other examples of encoding devices or decoding devices.

[0108] Based at least in part on using a neural network to encode and decode a data set for uplink communication, the encoding device can transmit the CSF with a reduced payload. This can save network resources that might otherwise have been used to transmit the full data set as sampled by the encoding device.

[0109] Figure 4B 4 is a diagram illustrating an example 450 associated with using a neural network to encode and decode a data set for uplink communication in accordance with various aspects of the present disclosure. An encoding device (e.g., UE 104, encoding device 400, etc.) may be configured to perform one or more operations on samples (e.g., data) received via one or more antennas of encoding device 400 to compress the samples. A decoding device 425 (e.g., base station 102 or 180, decoding device 425, etc.) may be configured to decode the compressed samples to determine information, such as a CSF.

[0110] In some aspects, the encoding device may identify features to be compressed. In some aspects, the encoding device may perform a first type of operation in a first dimension associated with the features to be compressed. The encoding device may perform a second type of operation in other dimensions (e.g., in all other dimensions). For example, the encoding device may perform a fully connected operation on a first dimension and perform convolution (e.g., point-by-point convolution) in all other dimensions.

[0111] In some aspects, reference numerals identify operations that include multiple neural network layers and / or operations. The neural networks of the encoding device and the decoding device may be formed by a cascade of one or more of the recited operations.

[0112] As indicated by reference numeral 455, the encoding device may perform spatial feature extraction on the data. As indicated by reference numeral 460, the encoding device may perform tap domain feature extraction on the data. In some aspects, the encoding device may perform tap domain feature extraction before performing spatial feature extraction. In some aspects, the extraction operation may include multiple operations. For example, the multiple operations may include one or more convolution operations that may be activated or inactive, one or more fully connected operations, etc. In some aspects, the extraction operation may include a residual neural network (ResNet) operation.

[0113] As indicated by reference numeral 465, the encoding device may compress the one or more features that have been extracted. In some aspects, the compression operation may include one or more operations, such as one or more convolution operations, one or more fully connected operations, etc. After compression, the bit count of the output may be less than the bit count of the input.

[0114] As indicated by reference numeral 470, the encoding device may perform a quantization operation. In some aspects, the encoding device may perform the quantization operation after flattening the output of the compression operation and / or performing a fully connected operation after flattening the output.

[0115] As indicated by reference numeral 475, the decoding device may perform feature decompression. As indicated by reference numeral 480, the decoding device may perform tap domain feature reconstruction. As indicated by reference numeral 485, the decoding device may perform spatial feature reconstruction. In some aspects, the decoding device may perform spatial feature reconstruction before performing tap domain feature reconstruction. After the reconstruction operation, the decoding device may output a reconstructed version of the input of the encoding device.

[0116] In some aspects, the decoding device may perform operations in the reverse order of the operations performed by the encoding device. For example, if the encoding device follows the operation (a, b, c, d), the decoding device may follow the reverse operation (D, C, B, A). In some aspects, the decoding device may perform operations that are completely symmetrical to the operations of the encoding device. This can reduce the number of bits required for the neural network configuration at the UE. In some aspects, the decoding device may perform additional operations in addition to the operations of the encoding device (e.g., convolution operations, fully connected operations, ResNet operations, etc.). In some aspects, the decoding device may perform operations that are asymmetrical to the operations of the encoding device.

[0117] Based at least in part on the use of a neural network by the coding device to encode a data set for uplink communication, the coding device (e.g., a UE) can transmit the CSF with a reduced payload. This can save network resources that might otherwise have been used to transmit the complete data set as sampled by the coding device.

[0118] As indicated above, Figure 4B These are provided as examples only. Other examples may differ from those described herein. Figure 4B Examples described.

[0119] Neural network-based CSI (such as combining Figure 4A The invention relates to a method for compressing downlink channels in a more comprehensive manner. For example, in type II CSI, the subband size may be fixed for all subbands for which the UE reports its CSI. For example, the subband granularity (e.g., subband) size may not be a function of the subband index within the bandwidth part (BWP). For some frequency bands, the subband size may provide the required higher granularity. In other frequency bands, the subband size may not provide sufficient granularity. Neural network-based CSI can solve the problem of fixed subband size by, for example, providing CSI over the entire channel. Neural network-based CSI can be configured to compress some subbands with higher or lower accuracy. Neural network-based CSI can also provide benefits for multi-user multiple input multiple output (MU-MIMO) wireless communications, such as at a base station. Neural network-based CSI provides direct information about the channel and interference, and allows decoding devices (such as base stations) to better group receivers (e.g., UEs).

[0120] Figure 5 5 is a diagram illustrating an example process 500 for an encoding device (e.g., UE 102, 350, encoding device 400, etc.) to perform one or more operations on data to compress the data. A decoding device (e.g., base station 102, 180, 310, decoding device 425, etc.) may be configured to decode the compressed data to determine information.

[0121] As used herein, a "layer" of a neural network is used to represent an operation on input data. For example, a convolutional layer, a fully connected layer, etc. represent associated operations on the data input into the layer. A convolution AxB operation refers to an operation that converts several input features A into several output features B. "Kernel size" refers to the number of adjacent coefficients that are combined in one dimension.

[0122] As used herein, "weights" are used to refer to one or more coefficients used in operations for combining individual rows and / or columns of input data in each layer. For example, a fully connected layer operation may have an output y that is determined at least in part based on the product of an input matrix x and a weight A (which may be a matrix) and a bias value B (which may be a matrix). The term "weight" may be used herein to refer generally to both weights and bias values.

[0123] like Figure 5 As shown in the example in , the encoding device may perform a convolution operation on the samples. For example, the encoding device may receive a set of bits structured as a 2x64x32 data set, which indicates IQ samples for tap features (e.g., associated with multipath timing offsets) and spatial features (e.g., associated with different antennas of the encoding device). The convolution operation can be a 2x2 operation with kernel sizes of 3 and 3 for the data structure. The output of the convolution operation can be input to a batch normalization (BN) layer, followed by a LeakyReLu activation, thereby giving an output data set with a size of 2x64x32. The encoding device may perform a flattening operation to flatten the bits into a 4096-bit vector. The encoding device may apply a fully connected operation with a size of 4096xM to the 4096-bit vector to output an M-bit payload. The encoding device may transmit the M-bit payload to the decoding device.

[0124] The decoding device may apply a fully connected operation with a size of Mx4096 to an M-bit payload to output a 4096-bit vector. The decoding device may reshape the 4096-bit vector to have a size of 2x64x32. The decoding device may apply one or more refinement network (RefineNet) operations to the reshaped bit vector. For example, the RefineNet operation may include: applying a 2x8 convolution operation (e.g., with kernel sizes of 3 and 3), whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set with a size of 8x64x32; applying an 8x16 convolution operation (e.g., with kernel sizes of 3 and 3), whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set with a size of 16x64x32; and / or applying a 16x2 convolution operation (e.g., with kernel sizes of 3 and 3), whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set with a size of 2x64x32. The decoding device may also apply 2x2 convolution operations with kernel sizes of 3 and 3 to generate decoded and / or reconstructed outputs.

[0125] As indicated above, Figure 5 These are provided as examples only. Other examples may differ from those described herein. Figure 5 Examples described.

[0126] As described herein, an encoding device operating in a network may measure reference signals, etc., to report to a decoding device. For example, a UE may measure reference signals during a beam management process to report channel state information feedback (CSF), may measure the received power of reference signals from a serving cell and / or a neighbor cell, may measure signal strength of an inter-radio access technology (e.g., WiFi) network, may measure sensor signals for detecting the location of one or more objects within an environment, etc. However, reporting this information to a network entity may consume communication and / or network resources.

[0127] In some aspects described herein, a coding device (e.g., a UE) may train one or more neural networks to learn the dependencies of these measured qualities on individual parameters, isolate these measured qualities through various layers (also referred to as "operations") of the one or more neural networks, and compress these measurements in a manner that limits compression losses.

[0128] In some aspects, the encoding device may use the properties of the number of bits being compressed to construct a process for extracting and compressing each feature (also referred to as a dimension) that affects the number of bits. In some aspects, the number of bits may be associated with sampling of one or more reference signals and / or may indicate channel state information.

[0129] Figure 6 600 is a diagram illustrating example operations 600 associated with using a neural network to encode and decode a data set for uplink communication in accordance with various aspects of the present disclosure. An encoding device (e.g., UE 120, encoding device 300, etc.) may be configured to perform one or more operations on samples (e.g., data) received via one or more antennas of the encoding device to compress the samples. A decoding device (e.g., base station 102 or 180, etc.) may be configured to decode the compressed samples to determine information, such as a CSF.

[0130] like Figure 6 As shown in the example of , the encoding device can receive samples from the antennas. For example, the encoding device can receive a data set of size 64x64 based at least in part on the number of antennas, the number of samples per antenna, and the tap characteristics.

[0131] The encoding device may perform spatial feature extraction, short-time (tap) feature extraction, etc. In some aspects, this may be achieved using a 1-dimensional convolution operation that is fully connected in the spatial dimension (to extract spatial features) and a simple convolution with a small kernel size (e.g., 3) in the tap dimension (to extract short tap features). The output from such a 64xW 1-dimensional convolution operation may be a Wx64 matrix.

[0132] The encoding device may perform one or more ResNet operations. The one or more ResNet operations may further refine the spatial features and / or temporal features. In some aspects, the ResNet operation may include multiple operations associated with the features. For example, the ResNet operation may include multiple (e.g., 3) 1-dimensional convolution operations, skip connections (e.g., between the input of the ResNet and the output of the ResNet to avoid applying a 1-dimensional convolution operation), a summation operation of a path through multiple 1-dimensional convolution operations and a path through a skip connection, etc. In some aspects, multiple 1-dimensional convolution operations may include: a Wx256 convolution operation with a kernel size of 3, whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set of size 256x64; a 256x512 convolution operation with a kernel size of 3, whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set of size 512x64; a 512xW convolution operation with a kernel size of 3, whose output is a BN data set of size Wx64. The output from one or more ResNet operations can be a Wx64 matrix.

[0133] The encoding device may perform a WxV convolution operation on the output from one or more ResNet operations. The WxV convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The WxV convolution operation may compress the spatial features into a reduced dimension for each tap. The WxV convolution operation has an input of W features and an output of V features. The output from the WxV convolution operation may be a Vx64 matrix.

[0134] The encoding device may perform a flattening operation to flatten the Vx64 matrix into a 64V element vector. The encoding device may perform a 64VxM fully connected operation to further compress the space-time feature data set into a low-dimensional vector of size M for transmission to the decoding device over the air. The encoding device may perform quantization before transmitting the low-dimensional vector of size M over the air to map the samples for transmission to discrete values ​​for the low-dimensional vector of size M.

[0135] The decoding device may perform an Mx64V fully connected operation to decompress a low-dimensional vector of size M into a space-time feature data set. The decoding device may perform a reshaping operation to reshape the 64V element vector into a 2-dimensional Vx64 matrix. The decoding device may perform a VxW (with a kernel of 1) convolution operation on the output from the reshaping operation. The VxW convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The VxW convolution operation may decompress the spatial features from the reduced dimension for each tap. The VxW convolution operation has an input of V features and an output of W features. The output from the VxW convolution operation may be a Wx64 matrix.

[0136] The decoding device may perform one or more ResNet operations. The one or more ResNet operations may further decompress spatial features and / or temporal features. In some aspects, the ResNet operation may include multiple (e.g., 3) 1-dimensional convolution operations, skip connections (e.g., to avoid applying 1-dimensional convolution operations), summation operations of paths through multiple convolution operations and paths through skip connections, etc. The output from one or more ResNet operations may be a Wx64 matrix.

[0137] The decoding device may perform spatial and temporal feature reconstruction. In some aspects, this may be achieved using a 1-dimensional convolution operation that is fully connected in the spatial dimension (to reconstruct spatial features) and a simple convolution with a small kernel size (e.g., 3) in the tap dimension (to reconstruct short tap features). The output from the 64xW convolution operation may be a 64x64 matrix.

[0138] In some aspects, the values ​​of M, W, and / or V may be configurable to adjust the weight of a feature, payload size, etc.

[0139] As indicated above, Figure 6 These are provided as examples only. Other examples may differ from those described herein. Figure 6 Examples described.

[0140] Figure 7 7 is a diagram illustrating example operations 700 associated with using a neural network to encode and decode a data set for uplink communication in accordance with various aspects of the present disclosure. An encoding device (e.g., UE 104, encoding device 300, etc.) may be configured to perform one or more operations on samples (e.g., data) received via one or more antennas of the encoding device to compress the samples. A decoding device (e.g., base station 102 or 180, etc.) may be configured to decode the compressed samples to determine information, such as a CSF. Figure 7 As shown in the example in , features can be compressed and decompressed in sequence. For example, an encoding device can extract and compress features associated with an input to produce a payload, and then a decoding device can extract and compress features associated with the payload to reconstruct the input. The encoding and decoding operations can be symmetric (as shown) or asymmetric.

[0141] like Figure 7 As shown, the encoding device may receive samples from the antennas. For example, the encoding device may receive a data set of size 256x64 based at least in part on the number of antennas, the number of samples per antenna, and the tap characteristics. The encoding device may reshape the data into a (64x64x4) data set.

[0142] The encoding device may perform 2-dimensional 64x128 convolution operations (with kernel sizes of 3 and 1). In some aspects, the 64x128 convolution operations may perform spatial feature extraction associated with the decoding device antenna dimension, short-time (tap) feature extraction associated with the decoding device (e.g., base station) antenna dimension, etc. In some aspects, this may be achieved using a 2D convolution layer that is fully connected in the decoding device antenna dimension, has a small kernel size (e.g., 3) in the tap dimension, and has a simple convolution operation with a small kernel size (e.g., 1) in the encoding device antenna dimension. The output from the 64xW convolution operation may be a matrix of size (128x64x4).

[0143] The encoding device may perform one or more ResNet operations. The one or more ResNet operations may further refine the spatial features associated with the decoding device and / or the temporal features associated with the decoding device. In some aspects, the ResNet operation may include multiple operations associated with the features. For example, the ResNet operation may include multiple (e.g., 3) 2D convolution operations, skip connections (e.g., between the input of the ResNet and the output of the ResNet to avoid applying a 2D convolution operation), a summation operation of a path through multiple 2D convolution operations and a path through a skip connection, etc. In some aspects, the plurality of 2D convolution operations may include: a Wx2W convolution operation with kernel sizes of 3 and 1, whose output is input to a BN layer followed by a LeakyReLU activation, which produces an output dataset of size 2Wx64xV; a 2Wx4W convolution operation with kernel sizes of 3 and 1, whose output is input to a BN layer followed by a LeakyReLU activation, which produces an output dataset of size 4Wx64xV; a 4WxW convolution operation with kernel sizes of 3 and 1, whose output is a BN dataset of size (128x64x4). The output from one or more ResNet operations may be a matrix of size (128x64x4).

[0144] The encoding device may perform a 2-dimensional 128xV convolution operation (with kernel sizes of 1 and 1) on the output from the one or more ResNet operations. The 128xV convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The WxV convolution operation may compress the spatial features associated with the decoding device into a reduced dimension for each tap. The output from the 128xV convolution operation may be a matrix of size (4x64xV).

[0145] The encoding device may perform a 2-dimensional 4x8 convolution operation (with kernel sizes of 3 and 1). In some aspects, the 4x8 convolution operation may perform spatial feature extraction associated with the encoding device antenna dimensions, short-term (tap) feature extraction associated with the encoding device antenna dimensions, etc. The output from the 4x8 convolution operation may be a matrix of size (8x64xV).

[0146] The encoding device may perform one or more ResNet operations. The one or more ResNet operations may further refine the spatial features associated with the encoding device and / or the temporal features associated with the encoding device. In some aspects, the ResNet operation may include multiple operations associated with the features. For example, the ResNet operation may include multiple (e.g., 3) 2D convolution operations, skip connections (e.g., to avoid applying 2D convolution operations), summation operations of paths through multiple 2D convolution operations and paths through skip connections, etc. The output from one or more ResNet operations may be a matrix of size (8x64xV).

[0147] The encoding device may perform a 2-dimensional 8xU convolution operation (with kernel sizes of 1 and 1) on the output from the one or more ResNet operations. The 8xU convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The 8xU convolution operation may compress the spatial features associated with the decoding device into a reduced dimension for each tap. The output from the 128xV convolution operation may be a matrix of size (Ux64xV).

[0148] The encoding device may perform a flattening operation to flatten a matrix of size (Ux64xV) into a 64UV element vector. The encoding device may perform a 64UVxM fully connected operation to further compress the 2-dimensional space-time feature data set into a low-dimensional vector of size M for transmission over the air to the decoding device. The encoding device may perform quantization before transmitting the low-dimensional vector of size M over the air to map the samples for transmission to discrete values ​​for the low-dimensional vector of size M.

[0149] The decoding device may perform an Mx64UV fully connected operation to decompress the low-dimensional vector of size M into a spatial-temporal feature data set. The decoding device may perform a reshaping operation to reshape the 64UV element vector into a matrix of size (Ux64xV). The decoding device may perform a 2-dimensional Ux8 (with a kernel of 1,1) convolution operation on the output from the reshaping operation. The Ux8 convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The Ux8 convolution operation may decompress the spatial features from the reduced dimension for each tap. The output from the Ux8 convolution operation may be a data set of size (8x64xV).

[0150] The decoding device may perform one or more ResNet operations. The one or more ResNet operations may further decompress the spatial features and / or temporal features associated with the encoding device. In some aspects, the ResNet operation may include multiple (e.g., 3) 2D convolution operations, skip connections (e.g., to avoid applying 2D convolution operations), summation operations of paths through multiple 2D convolution operations and paths through skip connections, etc. The output from the one or more ResNet operations may be a data set of size (8x64xV).

[0151] The decoding device may perform a 2-dimensional 8x4 convolution operation (with kernel sizes of 3 and 1). In some aspects, the 8x4 convolution operation may perform spatial feature reconstruction in the antenna dimension of the encoding device, as well as short-term feature reconstruction, etc. The output from the 8x4 convolution operation may be a data set of size (Vx64x4).

[0152] The decoding device may perform a 2-dimensional Vx128 (with a kernel of 1) convolution operation on the output from the 2-dimensional 8x4 convolution operation to reconstruct the tap features and spatial features associated with the decoding device. The Vx128 convolution operation may include a point-by-point (e.g., tap-by-tap) convolution operation. The Vx128 convolution operation may decompress the spatial features associated with the decoding device antenna from the reduced dimension for each tap. The output from the Ux8 convolution operation may be a matrix of size (128x64x4).

[0153] The decoding device may perform one or more ResNet operations. The one or more ResNet operations may further decompress the spatial features and / or temporal features associated with the decoding device. In some aspects, the ResNet operation may include multiple (e.g., 3) 2D convolution operations, skip connections (e.g., to avoid applying 2D convolution operations), summation operations of paths through multiple 2D convolution operations and paths through skip connections, etc. The output from the one or more ResNet operations may be a matrix of size (128x64x4).

[0154] The decoding device may perform a 2-dimensional 128x64 convolution operation (with kernel sizes of 3 and 1). In some aspects, the 128x64 convolution operation may perform spatial feature reconstruction associated with the decoding device antenna dimensions, short-term feature reconstruction, etc. The output from the 128x64 convolution operation may be a data set of size (64x64x4).

[0155] In some aspects, the values ​​of M, V, and / or U may be configurable to adjust feature weights, payload sizes, etc. For example, the value of M may be 32, 64, 128, 256, or 512, the value of V may be 16, and / or the value of U may be 1.

[0156] As indicated above, Figure 7 These are provided as examples only. Other examples may differ from those described herein. Figure 7 Examples described.

[0157] Figure 8 8 is a diagram illustrating example operations 800 associated with using a neural network to encode and decode a data set for uplink communication in accordance with various aspects of the present disclosure. An encoding device (e.g., UE 120, encoding device 300, etc.) may be configured to perform one or more operations on samples (e.g., data) received via one or more antennas of the encoding device to compress the samples. A decoding device (e.g., base station 110, etc.) may be configured to decode the compressed samples to determine information, such as a CSF. The encoding device and decoding device operations may be asymmetric. In other words, the decoding device may have a greater number of layers than the decoding device.

[0158] like Figure 8 As shown in the example of , the encoding device can receive samples from the antennas. For example, the encoding device can receive a data set of size 64x64 based at least in part on the number of antennas, the number of samples per antenna, and the tap characteristics.

[0159] The coding device may perform a 64xW convolution operation (with a kernel size of 1). In some aspects, the 64xW convolution operation may be fully connected in the antenna, may be a convolution in the tap, and the like. The output from the 64xW convolution operation may be a Wx64 matrix. The coding device may perform one or more WxW convolution operations (with a kernel size of 1 or 3). The output from one or more WxW convolution operations may be a Wx64 matrix. The coding device may perform a convolution operation (with a kernel size of 1). In some aspects, one or more WxW convolution operations may perform spatial feature extraction, short-time (tap) feature extraction, and the like. In some aspects, the WxW convolution operation may be a series of 1-dimensional convolution operations.

[0160] The encoding device may perform a flattening operation to flatten the Wx64 matrix into a 64W element vector. The encoding device may perform a 4096xM fully connected operation to further compress the space-time feature data set into a low-dimensional vector of size M for transmission to the decoding device over the air. The encoding device may perform quantization before transmitting the low-dimensional vector of size M over the air to map the samples for transmission to discrete values ​​for the low-dimensional vector of size M.

[0161] The decoding device may perform a 4096xM fully connected operation to decompress the low-dimensional vector of size M into a spatial-temporal feature data set. The decoding device may perform a reshape operation to reshape the 6W element vector into a Wx64 matrix.

[0162] The decoding device may perform one or more ResNet operations. The one or more ResNet operations may decompress spatial features and / or temporal features. In some aspects, the ResNet operation may include multiple (e.g., 3) 1-dimensional convolution operations, skip connections (e.g., between the input of the ResNet and the output of the ResNet to avoid applying a 1-dimensional convolution operation), a summation operation of a path through multiple 1-dimensional convolution operations and a path through a skip connection, etc. In some aspects, multiple 1-dimensional convolution operations may include: a Wx256 convolution operation with a kernel size of 3, whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set of size 256x64; a 256x512 convolution operation with a kernel size of 3, whose output is input to a BN layer, followed by a LeakyReLU activation, which produces an output data set of size 512x64; a 512xW convolution operation with a kernel size of 3, whose output is a BN data set of size Wx64. The output from one or more ResNet operations may be a Wx64 matrix.

[0163] The decoding device may perform one or more WxW convolution operations (with a kernel size of 1 or 3). The output from one or more WxW convolution operations may be a Wx64 matrix. The encoding device may perform a convolution operation (with a kernel size of 1). In some aspects, the WxW convolution operation may perform spatial feature reconstruction, short-term (tap) feature reconstruction, etc. In some aspects, the WxW convolution operation may be a series of 1-dimensional convolution operations.

[0164] The encoding device may perform a Wx64 convolution operation (with a kernel size of 1). In some aspects, the Wx64 convolution operation may be a 1-dimensional convolution operation. The output from the 64xW convolution operation may be a 64x64 matrix.

[0165] In some aspects, the values ​​of M and / or W may be configurable to adjust the weight of features, payload size, etc.

[0166] As indicated above, Figure 8 These are provided as examples only. Other examples may differ from those described herein. Figure 8 Examples described.

[0167] Fig. 9 9 is a diagram illustrating an example process 900 performed, for example, by a first device in accordance with various aspects of the present disclosure. Example process 900 is an example in which a first device (eg, encoding device, UE 104, etc.) performs operations associated with encoding a data set using a neural network.

[0168] like Fig. 9As shown in , in some aspects, process 900 may include encoding a data set using one or more extraction operations and compression operations associated with a neural network to produce a compressed data set, the one or more extraction operations and compression operations being based at least in part on a feature set of the data set (block 910). For example, a first device may encode a data set using one or more extraction operations and compression operations associated with a neural network to produce a compressed data set, the one or more extraction operations and compression operations being based at least in part on a feature set of the data set, as described above.

[0169] like Fig. 9 As further shown in FIG. 9 , in some aspects, process 900 may include transmitting the compressed data set to the second device (block 920). For example, the first device may transmit the compressed data set to the second device, as described above.

[0170] 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 herein.

[0171] The data set may be based at least in part on sampling of one or more reference signals.The compressed data set may include channel state information feedback transmitted to the second device.

[0172] The method may further include identifying a feature set of the data set, wherein the one or more extraction operations and compression operations include: a first type of operation performed in a dimension associated with a feature in the feature set of the data set and a second type of operation different from the first type of operation performed in the remaining dimensions associated with other features in the feature set of the data set. The first type of operation may include a one-dimensional fully connected layer operation, and the second type of operation includes a convolution operation. The one or more extraction operations and compression operations may include multiple operations, including one or more of convolution operations, fully connected layer operations, or residual neural network operations. The one or more extraction operations and compression operations may include a first extraction operation and a first compression operation performed on a first feature in the feature set of the data set, and a second extraction operation and a second compression operation performed on a second feature in the feature set of the data set.

[0173] The process may further include performing one or more additional operations on the intermediate data set output after performing the one or more extraction operations and compression operations. The one or more additional operations include one or more of a quantization operation, a flattening operation, or a full-connection operation.

[0174] The feature set of the data set may include one or more of spatial features or tap-domain features.

[0175] The one or more extraction operations and compression operations may include one or more of the following: spatial feature extraction using a one-dimensional convolution operation, temporal feature extraction using a one-dimensional convolution operation, a residual neural network operation for refining the extracted spatial features, a residual neural network operation for refining the extracted temporal features, a point-by-point convolution operation for compressing the extracted spatial features, a point-by-point convolution operation for compressing the extracted temporal features, a flattening operation for flattening the extracted spatial features, a flattening operation for flattening the extracted temporal features, or a compression operation for compressing one or more of the extracted temporal features or the extracted spatial features into a low-dimensional vector for transmission.

[0176] The one or more extraction operations and compression operations include: a first feature extraction operation associated with one or more features associated with the second device, a first compression operation for compressing one or more features associated with the second device, a second feature extraction operation associated with one or more features associated with the first device, and a second compression operation for compressing one or more features associated with the first device.

[0177] although Fig. 9 An example block diagram of process 900 is shown, but in some aspects, process 900 may include Fig. 9 Additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in the process 900. Additionally or alternatively, two or more blocks of the process 900 may be executed in parallel.

[0178] Fig.10 1 is a diagram illustrating an example process 1000 performed, for example, by a second device in accordance with various aspects of the present disclosure. Example process 1000 is an example in which a second device (eg, a decoding device, base station 102, 180, etc.) performs operations associated with decoding a data set using a neural network.

[0179] like Fig.10 As shown in , in some aspects, process 1000 may include receiving a compressed data set from a first device (block 1010). For example, a second device may receive a compressed data set from a first device, as described above.

[0180] like Fig.10As further shown in FIG. 1 , in some aspects, process 1000 may include decoding the compressed data set using one or more decompression operations and reconstruction operations associated with the neural network to produce a reconstructed data set, the one or more decompression operations and reconstruction operations based at least in part on the feature set of the compressed data set (block 1020). For example, the second device may decode the compressed data set using one or more decompression operations and reconstruction operations associated with the neural network to produce a reconstructed data set, the one or more decompression operations and reconstruction operations based at least in part on the feature set of the compressed data set, as described above.

[0181] 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 herein.

[0182] Decoding a compressed data set using one or more decompression operations and reconstruction operations may include performing the one or more decompression operations and reconstruction operations based at least in part on an assumption that the first device generated the compressed data set using a set of operations that are symmetric to the one or more decompression operations and reconstruction operations, or performing the one or more decompression operations and reconstruction operations based at least in part on an assumption that the first device generated the compressed data set using a set of operations that are asymmetric to the one or more decompression operations and reconstruction operations.

[0183] The compressed data set may be based at least in part on sampling of one or more reference signals by the first device.

[0184] The process may further include receiving the compressed data set includes receiving channel state information feedback from the first device.

[0185] The one or more decompression operations and reconstruction operations may include: a first type of operation performed in a dimension associated with a feature in the feature set of the compressed data set, and a second type of operation different from the first type of operation performed in remaining dimensions associated with other features in the feature set of the compressed data set.

[0186] The first type of operation may include a one-dimensional fully connected layer operation, and wherein the second type of operation includes a convolution operation.

[0187] The one or more decompression operations and reconstruction operations may include multiple operations, including one or more of a convolution operation, a fully connected layer operation, or a residual neural network operation. The one or more decompression operations and reconstruction operations may include: a first operation performed on a first feature in the feature set of the compressed data set, and a second operation performed on a second feature in the feature set of the compressed data set.

[0188] Process 1000 may further include performing a reshape operation on the compressed data set.

[0189] The feature set of the compressed data set may include one or more of spatial features or tap-domain features.

[0190] The one or more decompression operations and reconstruction operations may include one or more of a feature decompression operation, a temporal feature reconstruction operation, or a spatial feature reconstruction operation. The one or more decompression operations and reconstruction operations may include: a first feature reconstruction operation performed on one or more features associated with the first device, and a second feature reconstruction operation performed on one or more features associated with the second device.

[0191] although Fig.10 An example block diagram of process 1000 is shown, but in some aspects, process 1000 may include Fig.10 Additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in the process 1000. Additionally or alternatively, two or more blocks of the process 1000 may be executed in parallel.

[0192] Fig.11 An example timeline 1100 for CSI report triggering is illustrated. A base station may trigger an aperiodic CSI report from a UE by transmitting a PDCCH to the UE indicating that the UE is to measure CSI and transmit a CSI report. The UE not only uses available computational resources to compute the report, it also uses time to perform the computation and provide the report. Fig.11 Two timelines for aperiodic CSI reporting are illustrated. The first processing time may be based on the number (Z) of OFDM symbols (e.g., a minimum number) between the last symbol of the PDCCH that triggers the aperiodic CSI report and the first symbol of the PUSCH that carries the CSI report. During this time, the UE decodes the PDCCH, performs possible CSI-RS / IM measurements (if the UE does not already have the latest previous channel / interference measurements stored in its memory), performs possible channel estimation, calculates CSI reports, and performs UCI multiplexing with the UL-SCH.

[0193] However, if the reporting is for aperiodic CSI-RS / IM, this first processing time (e.g., which may be referred to as the first requirement) may not ensure that the UE has enough time to calculate the CSI, because the aperiodic CSI-RS may potentially be triggered close to the PUSCH transmission. Therefore, the second processing time may be based on the number (Z') of OFDM symbols (e.g., a minimum number) between the last symbol of the aperiodic CSI-RS / IM used to calculate the report and the first symbol of the PUSCH carrying the CSI report.

[0194] The difference between the Z processing time and the Z' processing time may be: Z is required to additionally cover the DCI decoding time of the UE. Thus, Z may be a few symbols larger than the corresponding Z' value, e.g. Fig.11 as shown in .

[0195] If the Z criterion (or Z' criterion) is not met and the base station is too close to the PDCCH that triggers the CSI report (or aperiodic CSI-RS / IM for measurement) to trigger the PUSCH for the UE to report CSI, the UE may ignore the scheduling DCI, for example, if the UE is also not scheduled for UL-SCH or HARQ-ACK, and the UE may not transmit CSI. However, if the UE is scheduled to multiplex UL-SCH or HARQ-ACK on PUSCH, the UE may transmit PUSCH but may fill the CSI report with dummy bits or transmit a stale CSI report. Table 1 illustrates an example of processing times to illustrate the concept that the processing times Z and Z' may be different based on different subcarrier spacings (μ). Table 1 is only an example, and the processing times may be different from the example illustrated in Table 1. Additionally or alternatively, the processing times Z and Z' may be different for different levels of latency (e.g., for low latency CSI and high latency CSI) or for beam reporting. Table 2 illustrates an example of processing time to illustrate the concept of processing time Z and Z' for ultra-low latency CSI reporting, and Table 3 illustrates an example of processing time to illustrate the concept of the difference in processing time Z' in milliseconds for high latency CSI and ultra-low latency CSI. Tables 2 and 3 are merely examples to illustrate the concept of different processing times for different subcarrier spacings and different latencies. The processing time may be different from the example illustrated in Table 1.

[0196] Table 1

[0197]

[0198] Table 2

[0199]

[0200]

[0201] Table 3

[0202] μ High latency CSI Ultra-low latency CSI 0 2.64 ms 0.57 ms 1 2.46 ms 0.4 ms 2 2.5 ms 0.375 ms 3 1.2 ms 0.32 ms

[0203] Ultra-low latency CSI may provide a different processing timeline that may be applied in some cases, such as when a single low latency CSI report is triggered without multiplexing with an uplink shared channel (UL-SCH) or HARQ-ACK, and when the UE has a certain amount of computational resources (e.g., all of its CPUs) that are not occupied. The UE may then allocate its computational resources to compute the CSI in a shorter amount of time. Thus, the CSI processing time for aperiodic CSI reporting may vary based on parameter design (subcarrier spacing), number of ports, latency, etc.

[0204] Another type of CSI may be semi-persistent CSI which may be triggered by downlink control information (DCI). Fig.12 Illustrated is an example semi-persistent CSI timeline 1200 in which a base station configures a MAC-CE 1202 that configures measurements of CSI-RS and / or CSI reports for a UE. The base station then transmits a DCI 1204 that activates the configuration and triggers the UE to measure a reference signal 1206 and transmit the report in a PUSCH 1208. Fig.11 , Fig.12 The processing times Z and Z' are explained. Fig.12 Illustrated is the additional processing time that may be required between the MAC-CE 1202 configuring the CSI-RS or CSI report and activating the DCI 1204. For example, at least 3 milliseconds may be provided from the PDSCH configuring the CSI-RS or CSI report to the DCI activating the configuration.

[0205] As combined Figure 4A-10 As described, a UE may perform measurements or provide reports (such as CSI) based on a neural network. As proposed herein, the UE and the base station may determine a processing time for the UE based at least in part on parameters of the neural network. The processing time may correspond to a time between a reference signal used for training in the neural network until a time when the UE has successfully trained the network. A successfully trained network may correspond to a neural network to which the UE is able to report CSI trained by it back, or to which the UE is able to use trained weights of the neural network to achieve a required / configured accuracy or QoS. The processing time may correspond to a time between a UE receiving a neural network to be trained and a time when the UE completes training.

[0206] Fig.13 An example communication flow 1300 between a UE 1302 and a network entity 1304 is illustrated. The network device may be a base station, a TRP, or another UE. Although the example is described with respect to a UE and a network entity, aspects may be applied to a first wireless device as an encoding device and a second wireless device as a decoding device.

[0207] As illustrated at 1306, the UE 1302 may determine a processing time for training a neural network (e.g., for CSI derivation or other wireless communication measurements or outputs).

[0208] For example, the UE may be based on the number of layers in a particular neural network (such as a combination of Figure 5-8 The UE may determine the amount of time to train the neural network or the physical resources to use to train the network based on the number of weights in the neural network (such as the combination of Figure 5-8 The UE may determine the processing time based on the type of layer in the neural network, such as whether the layer is a fully connected layer, a one-dimensional (1-D) convolutional layer, a residual neural network layer, or a point-by-point convolutional layer. The UE may determine the processing time based on whether information from a previously trained neural network is used as a starting point for training the neural network. For example, the UE may consider whether quasi-co-location (QCL) information is provided for training the neural network or whether a neural network state indication is provided for training the neural network. The UE may determine the processing time based on whether a single neural network or multiple neural networks are expected to be trained simultaneously. "Simultaneous" training may be used herein to refer to concurrent training of multiple neural networks in at least one of the same component carrier, the same frequency band, the same bandwidth portion, the same frequency band combination, the same frequency range, the same time slot, the same subframe, or the same frame (e.g., at least partially overlapping in time). Training may be performed, for example, in a TDM manner, where, for example, one neural network is trained after another until each neural network is trained. Simultaneous training may correspond to the number of training procedures to which the UE has not yet responded with a completion training message (e.g., a message indicating completion of training). In other words, after the UE receives a command to start training, this training procedure may be considered active until the UE reports back a completion training message. The completion training message may correspond to a report of a trained neural network. The UE may determine the processing time based on the procedure in which the neural network is being trained. For example, the UE may determine different times for training the neural network based on whether the neural network is being trained for CSI determination, demodulation, positioning determination, or waveform determination. The UE may determine the processing time based on the requested accuracy (e.g., a QoS that may be associated with a training level). The UE may determine the processing time based on whether a single layer or multiple layers within the neural network are expected to be trained. The layers to be trained may be, for example, a subset of several layers in the neural network. The UE may determine the processing time based on the sequence of the layers of the neural network (e.g., the type and order of the layers in the neural network). For example, a neural network having a fully connected layer followed by a fully connected layer may have a different processing time than a neural network having a residual neural network layer followed by a fully connected layer. The processing time for training the neural network may be based on any combination of the examples of parameters or features described herein.

[0209] The UE may also determine a processing time from the time the UE is triggered to provide a report until the UE provides a measurement, CSI, or result of the trained neural network. The processing time may be based on, for example, a fully trained neural network. For example, at 1308, Fig.13 In the example, the UE may determine the processing time for CSI derivation based on a fully trained neural network. Although the example is described with respect to determining the processing time for CSI reporting to illustrate the concept, the concept may be applied to other measurements or outputs of a neural network of a wireless communication device. The UE may determine the CSI processing time based on, for example, an encoder output vector of the neural network. In other examples, the UE may determine the CSI processing time based on an encoder input vector input to the neural network. The UE may determine the CSI processing time based on a vector calculated in the neural network or based on a combination of vectors calculated in the neural network. The UE may determine the CSI processing time based on the number of layers in the neural network. The UE may determine the CSI processing time based on the number of elements in the input vector, the output vector, and / or the intermediate vector(s) in the neural network. The UE may determine the CSI processing time based on the type of one or more layers in the neural network (e.g., fully connected, 1-D convolution, residual neural network, point-by-point convolution layer, etc.). The UE may determine the CSI processing time based on the sequence of layers in the neural network (e.g., a fully connected layer followed by a fully connected layer may have a different processing time than a neural network having a residual neural network layer followed by a fully connected layer). The UE may determine the CSI processing time based on any combination of the described factors.

[0210] As illustrated at 1310, after determining the processing time at 1306 and / or determining the processing time at 1308, the UE 1302 may provide the determined processing time to the network entity 1304. For example, the UE may report capabilities based on the training time used to train one or more neural networks and for deriving results by the neural network(s). The UE 1302 may report the processing time as a UE capability in higher layer signaling, in the MAC-CE, or in the UCI. The UE may report the processing time per frequency band, per frequency band combination, per frequency range, per BWP, per parameter design, per component carrier (CC), or on a combination basis. Additionally or alternatively, the UE may report the processing time for the different factors described in conjunction with the time determination. For example, the UE may report the processing time(s) per layer, per layer type, per layer combination, per input / output / intermediate vector length, per number of layers, or per layer sequence, etc.

[0211] The UE may report different levels of processing timelines, e.g., slower times and faster times, such as described in conjunction with the low latency, high latency, and ultra-low latency examples described with respect to Tables 1-3. The UE may report different levels of processing time for the same neural network, for the same layer, for the same combination of layers, for the same sequence of layers, and so on. The network entity 1304 may configure the UE to apply one of the reported timelines. In other examples, the UE may apply one of the reported timelines, e.g., based on a power saving feature at the UE. For example, the UE may apply a faster processing timeline associated with higher power consumption.

[0212] As illustrated at 1312, the network entity 1304 may configure the UE 1302 to measure the CSI-RS and / or report the CSI. In some examples, the network entity 1304 may configure the UE 1302 with one or more parameters for training a neural network for CSI derivation. The network entity 1304 may transmit a reference signal 1314 for the UE to measure, such as a CSI-RS, SSB, CSI-IM resource, etc. At 1316, the UE 1302 measures the CSI or some other measurement or output using the neural network and reports the CSI 1318 to the network entity.

[0213] Fig.14A is a flow chart 1400 of a wireless communication method. The method may be performed by a first wireless device. The first wireless device may correspond to Figure 4A In some examples, the method may be performed by a UE (eg, UE 104, 350) or by device 1502. Although various aspects of the method are described with respect to an example of a UE, various aspects may be applied to other wireless devices.

[0214] At 1402, the UE generates a message indicating a processing time at the UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network. For example, a processing time calculation component 1540 of the device 1502 can perform a determination of a processing time to be included in the message, and the message can be generated by the capability component 1542 based on the determined processing time. For example, Fig.13 An example of a UE determining a processing time at 1306 and 1308 is illustrated. The processing time may correspond to an amount of time between a first time associated with a reference signal for training in a neural network and a second time at which the UE has successfully trained the neural network. The neural network may be successfully trained when the UE is able to report CSI trained by it back to the neural network, or when the UE is able to achieve accuracy or QoS using trained weights of the neural network. The processing time may correspond to a time between the UE receiving a command to train the neural network and the UE completing training of the neural network.

[0215] The UE may determine the processing time for training the neural network based on at least one of: the number of layers in the neural network, the number of weights in the neural network, or the type of one or more layers of the neural network. The UE may determine the processing time for training the neural network based on the use of information from a previously trained neural network. The information from the previously trained neural network may include at least one of QCL information or a neural network state indication from the previously trained neural network. The UE may determine the processing time for training the neural network based on the amount of neural networks or layers to be trained.

[0216] The UE may determine the processing time based on whether a single neural network or multiple neural networks are to be trained simultaneously. The multiple neural networks may be trained simultaneously based on concurrent training of at least one of: same component carrier, same frequency band, same bandwidth portion, same frequency band combination, same frequency range, same time slot, same subframe, or same frame. A neural network in the multiple neural networks may be considered to be trained simultaneously until the UE responds with a completed training message.

[0217] The UE may determine the processing time based on whether a single layer or multiple layers of the neural network are to be trained simultaneously. The UE may determine the processing time for training the neural network based on the sequence order of the multiple layers of the neural network. The UE may determine the processing time for training the neural network based on the type of wireless signal procedure performed by the neural network. The type of wireless signal procedure includes at least one of: channel state information determination, demodulation, positioning determination, or waveform determination. The UE may determine the processing time for training the neural network based on the accuracy level. The accuracy level may be based on QoS, for example. The UE may determine the processing time based on any combination of the described parameters or factors.

[0218] The UE may determine the processing time for reporting the CSF based on at least one of the following or any combination thereof: an encoder output vector, an encoder input vector, one or more vectors determined at a neural network, the number of layers in the neural network, the number of first elements in the input of the neural network, the number of second elements in the output of the neural network, the number of third elements in an intermediate vector of the neural network, the layer type of one or more layers of the neural network, the amount of neural network reported for overlap, or the sequence order of multiple layers of the neural network.

[0219] At 1404, the UE transmits the processing time to the first network entity. For example, the UE may transmit a message indicating the processing time to the first network entity. The processing time may be transmitted, for example, by the capability component 1542 via the transmission component 1534 and / or the cellular RF transceiver 1522. The first network entity may be a base station, a TRP, or another UE. Although this example is described in conjunction with a UE and a network entity, aspects of the method may also be performed by a first wireless device that provides the processing time to a second wireless device. In some examples, the first wireless device may correspond to the encoding device 400, and the second wireless device may correspond to the decoding device 425.

[0220] The UE may report the processing time as a UE capability, MAC-CE, or UCI. The UE may report the processing time for at least one of a bandwidth part, a parameter design, a component carrier, a frequency band, a frequency band combination, a frequency range, or one or more timeline factors. The one or more timeline factors may include at least one of the following: a layer, a layer type, a combination of layers, an input vector length, an output vector length, an intermediate vector length, a number of layers, or a sequence of layers.

[0221] Fig. 14B Flowchart 1450 illustrates a method of wireless communication. The method may include incorporating Fig.14A The method may be performed by a UE (eg, UE 104, 350) or by a device 1502. Fig.14A The various aspects of the description are shown with the same reference numerals. As illustrated at 1406, the UE may receive a configuration based at least on the processing time from a second network entity. For example, the configuration may be received by a configuration component 1544 of device 1502. In some examples, the second network entity may be the same as the first network entity. As an example, the UE may provide processing time capabilities to a base station and may receive a configuration for measuring and / or reporting CSI from the base station. The UE may provide processing time capabilities to another UE and may receive an indication to report CSI from the other UE. In other examples, the second network entity may be different from the first network entity. For example, the UE may report processing time capabilities to a server, and may receive a configuration for reporting CSI from a base station and may report CSI to the base station.

[0222] In some examples, the UE may report the first processing time and the second processing time, e.g., the slower processing time and the faster processing time, at 1404. The UE may then receive a configuration from the second network entity to use the first processing time or the second processing time at 1406. Additionally or alternatively, the UE may apply the first processing time or the second processing time based on a power saving feature of the UE.

[0223] At 1408, the UE may transmit CSI to the second network entity based on the configuration. Transmission of the CSI may be performed, for example, by CSI reporting component 1546 of device 1502.

[0224] Fig.15 1500 is an example of a hardware implementation of an illustrated device 1502. The device 1502 may be a UE, a component of a UE, or may implement UE functionality. In some aspects, the device 1502 includes a cellular baseband processor 1504 (also referred to as a modem) coupled to a cellular RF transceiver 1522. In some aspects, the device 1502 may further include one or more of: a subscriber identity module (SIM) card 1520, an application processor 1506 coupled to a secure digital (SD) card 1508 and a screen 1510, a Bluetooth module 1512, a wireless local area network (WLAN) module 1514, a global positioning system (GPS) module 1516, and / or a power supply 1518. The cellular baseband processor 1504 communicates with the UE 104 and / or the BS 102 / 180 via the cellular RF transceiver 1522. The cellular baseband processor 1504 may include a computer-readable medium / memory. The computer-readable medium / memory may be non-transient. The cellular baseband processor 1504 is responsible for general processing, including the execution of software stored on a computer-readable medium / memory. The software, when executed by the cellular baseband processor 1504, causes the cellular baseband processor 1504 to perform the various functions described above. The computer-readable medium / memory may also be used to store data manipulated by the cellular baseband processor 1504 when executing the software. The cellular baseband processor 1504 further includes a receiving component 1530, a communication manager 1532, and a transmission component 1534. The communication manager 1532 includes the one or more illustrated components. The components within the communication manager 1532 may be stored in a computer-readable medium / memory and / or configured as hardware within the cellular baseband processor 1504. The cellular baseband processor 1504 may be a component of the UE 350 and may include a memory 360 and / or at least one of the following: a TX processor 368, an RX processor 356, and a controller / processor 359. In one configuration, the device 1502 may be a modem chip and include only the cellular baseband processor 1504, and in another configuration, the device 1502 may be an entire UE (e.g., see Figure 3 350) and includes additional modules of device 1502.

[0225] The communication manager 1532 includes a processing time calculation component 1540, which is configured to determine the processing time, for example, in conjunction with Fig.14A or 1402 in 14B. The communication manager 1532 further includes a capability component 1542 configured to provide processing time to the first network entity, for example, as described in conjunction with Fig.14A The communication manager 1532 may further include a CSI configuration component 1544 configured to receive a configuration based on the processing time from the second network entity, such as described in conjunction with 1406. The communication manager 1532 may further include a CSI reporting component 1546 configured to report CSI to the second network entity based on the configuration, such as described in conjunction with 1408.

[0226] The apparatus may include executing Fig.14A or additional components to each box of the algorithm in the flowchart of 14B. Thus, Fig.14A Each block in the flowchart of 14B or 14B may be executed by a component and the device may include one or more of those components. These components may be one or more hardware components specifically configured to execute the process / algorithm, implemented by a processor configured to execute the process / algorithm, stored in a computer-readable medium for implementation by a processor, or some combination thereof.

[0227] As shown, the device 1502 may include various components configured for various functions. In one configuration, the device 1502 and specifically the cellular baseband processor 1504 include: a means for generating a message indicating a processing time at the UE for at least one of: training a neural network for CSI derivation or reporting CSI based on the trained neural network; and a means for providing the processing time to the first network entity (e.g., the capability component 1542, the transmission component 1534, and / or the transceiver 1522). The device 1502 may further include a means for determining a processing time at the UE for at least one of: training a neural network for CSI derivation or reporting CSI based on the trained neural network. The device may further include a means for receiving a configuration based on the processing time from the second network entity (e.g., the CSI configuration component 1544 of the communication manager 1532). The apparatus may further include means for transmitting CSI to a second network entity based on the configuration (e.g., CSI reporting component 1546, transmission component 1534, and / or transceiver 1522). The means may be one or more of the components of the apparatus 1502 configured to perform the functions recited by the means. As described above, the apparatus 1502 may include a TX processor 368, an RX processor 356, and a controller / processor 359. Thus, in one configuration, the respective means may be the TX processor 368, the RX processor 356, and the controller / processor 359 configured to perform the functions recited by the respective means.

[0228] Fig.16is a flow chart 1600 of a wireless communication method. The method may be performed by a first wireless device. The first wireless device may correspond to Figure 4A In some examples, the method may be performed by a network entity (such as a base station, a TRP, or a UE). The method may be performed by device 1702. Fig.16 Example blocks of flowchart 1600 are shown, but in some aspects, the wireless communication method may include Fig.16 1600. In some embodiments, the flowchart 1600 may include additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in the flowchart 1600. Additionally or alternatively, two or more blocks of the flowchart 1600 may be executed in parallel.

[0229] At 1602, a network entity receives from a UE a processing time for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network. The receiving may be performed, for example, by a CSI processing time component 1740. The processing time may correspond to an amount of time between a first time associated with a reference signal for training in the neural network and a second time when the UE has successfully trained the neural network. The neural network may be successfully trained when the UE is able to report CSI trained by it back to the neural network, or when the UE is able to achieve accuracy or QoS using trained weights of the neural network. The processing time may correspond to a time between the UE receiving a command to train the neural network and the UE completing training of the neural network.

[0230] The processing time for training the neural network may be based on at least one of: the number of layers in the neural network, the number of weights in the neural network, or the type of one or more layers of the neural network. The processing time for training the neural network may be based on the use of information from previously trained neural networks. The information from previously trained neural networks may include at least one of QCL information from previously trained neural networks or neural network state indications. The processing time for training the neural network may be based on the amount of neural networks or layers to be trained.

[0231] The processing time for training the neural network may be based on whether a single neural network or multiple neural networks are to be trained simultaneously. The multiple neural networks may be trained simultaneously based on concurrent training of at least one of the following: same component carrier, same frequency band, same bandwidth portion, same frequency band combination, same frequency range, same time slot, same subframe, or same frame. A neural network in the multiple neural networks may be considered to be trained simultaneously until the UE responds with a completed training message.

[0232] The processing time for training the neural network may be based on whether a single layer or multiple layers of the neural network are to be trained simultaneously. The processing time for training the neural network may be based on the sequential order of the multiple layers of the neural network. The processing time for training the neural network may be based on the type of wireless signal procedures performed by the neural network. The type of wireless signal procedures includes at least one of: channel state information determination, demodulation, positioning determination, or waveform determination. The processing time for training the neural network may be based on the accuracy level. The accuracy level may be based on QoS, for example. The processing time for training the neural network may be based on any combination of the described parameters or factors.

[0233] The processing time for reporting CSI may be based on at least one of the following or any combination thereof: an encoder output vector, an encoder input vector, one or more vectors determined at a neural network, the number of layers in the neural network, the number of first elements in the input of the neural network, the second number of elements in the output of the neural network, the third number of elements in an intermediate vector of the neural network, the layer type of one or more layers of the neural network, the amount of neural network reported for overlap, or the sequential order of multiple layers of the neural network.

[0234] The network entity may receive the processing time as a UE capability, a MAC-CE, or a UCI. The UE may report the processing time for at least one of a bandwidth part, a parameter design, a component carrier, a frequency band, a frequency band combination, a frequency range, or one or more timeline factors. The one or more timeline factors may include at least one of: a layer, a layer type, a combination of layers, an input vector length, an output vector length, an intermediate vector length, a number of layers, or a sequence of layers.

[0235] At 1604, the network entity transmits a configuration based on the processing time to the UE. For example, the configuration can be transmitted by the CSI configuration component 1744 of the device 1702.

[0236] In some examples, the UE may report a first processing time and a second processing time, such as a slower processing time and a faster processing time, at 1602. Then at 1604, the base station may transmit a configuration for the UE to use the first processing time or the second processing time.

[0237] At 1606, the network entity receives CSI from the UE based on the configuration. The reception of CSI may be performed, for example, by a CSI reporting component 1748 of the device 1502. Fig.13 An example is illustrated of the network entity 1304 receiving CSI 1318 from the UE 1302 based on a configuration 1312 that is provided after the network entity 1304 receives processing time information from the UE 1302 at 1310 .

[0238] Fig.171700 is an example of a hardware implementation of an illustrated device 1702. In some aspects, the device 1702 may be a base station, a component of a base station, or may implement base station functionality. In some aspects, the device 1702 may be another network component. The device 1702 may include a baseband unit 1704. In some examples, the baseband unit 1704 may communicate with the UE 104, the TRP 103, and / or the base station 102 / 180 via a cellular RF transceiver. The baseband unit 1704 may include a computer-readable medium / memory. The baseband unit 1704 is responsible for general processing, including the execution of software stored on the computer-readable medium / memory. The software, when executed by the baseband unit 1704, causes the baseband unit 1704 to perform the various functions described above. The computer-readable medium / memory may also be used to store data manipulated by the baseband unit 1704 when executing the software. The baseband unit 1704 further includes a receiving component 1730, a communication manager 1732, and a transmission component 1734. The communication manager 1732 includes the one or more illustrated components. The components within the communication manager 1732 may be stored in a computer-readable medium / memory and / or configured as hardware within the baseband unit 1704. The baseband unit 1704 may be a component of the base station 310 and may include a memory 376 and / or at least one of the following: the TX processor 316, the RX processor 370, and the controller / processor 375.

[0239] The communication manager 1732 includes a CSI processing time component 1740 configured to receive, from a user equipment (UE), a processing time for at least one of training a neural network for channel state information derivation or reporting CSI based on the trained neural network, e.g., as described in conjunction with 1602. The communication manager 1732 further includes a CSI configuration component 1744 configured to transmit a configuration based on the processing time to the UE, e.g., as described in conjunction with 1604. The communication manager 1732 further includes a CSI reporting component 1748 configured to receive a CSI report based on the configuration, e.g., as described in conjunction with 1606.

[0240] The apparatus may include executing Fig.16 The additional components of each box of the algorithm in the flowchart. Fig.16 Each block in the flowchart of can be performed by a component and the device may include one or more of these components. These components can be one or more hardware components specially configured to perform the process / algorithm, implemented by a processor configured to perform the process / algorithm, stored in a computer-readable medium for implementation by a processor, or some combination thereof.

[0241] As shown, the device 1702 may include various components configured for various functions. In one configuration, the device 1702 and specifically the baseband unit 1704 include: a device for receiving a processing time for at least one of the following from a user equipment (UE) (e.g., a CSI processing time component 1740, a receiving component 1730, and / or a transceiver 1722): training a neural network for channel state information derivation or reporting CSI based on the trained neural network; a device for transmitting a configuration based on the processing time to the UE (e.g., a CSI configuration component 1744, a transmission component 1734, and / or a transceiver 1722); and a device for receiving CSI from the UE based on the configuration (e.g., a CSI reporting component 1748, a receiving component 1730, and / or a transceiver 1722). The device may be one or more of the components of the device 1702 configured to perform the functions recited by the device. As described above, the device 1702 may include a TX processor 316, an RX processor 370, and a controller / processor 375. As such, in one configuration, the various means may be the TX processor 316, RX processor 370, and controller / processor 375 configured to perform the functions recited by the various means.

[0242] It should be understood that the specific order or hierarchy of each box in the disclosed process / flowchart is an illustration of an example approach. It should be understood that the specific order or hierarchy of each box in these process / flowcharts can be rearranged based on design preferences. In addition, some boxes can be combined or omitted. The attached method claims present the elements of various boxes in an exemplary order and are not meant to be limited to the specific order or hierarchy presented.

[0243] The previous description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be easily understood by those skilled in the art, and the universal principles defined in this article can be applied to other aspects. Thus, the claims are not intended to be limited to the aspects shown in this article, but should be granted the full scope consistent with the claims in language, wherein the singular reference of the elements is not intended to represent "there is and only one", but "one or more", unless otherwise stated. Terms such as "if", "when..." and "when..." should be interpreted as meaning "under the condition", rather than implying a direct time relationship or reaction. That is, these phrases (e.g., "when...") do not imply an immediate action in response to the occurrence of an action or during the occurrence of an action, but only imply that an action will occur when the condition is met, and no specific or immediate time constraints are required for the action to occur. The wording "exemplary" is used herein to mean "used as an example, instance or explanation". Any aspect described as "exemplary" herein need not be interpreted as being superior to or superior to other aspects. Unless otherwise stated, the term "some / certain" refers to one or more. Combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B and C," "one or more of A, B, and C," and "A, B, C, or any combination thereof" include any combination of A, B, and / or C, and may include multiple A, multiple B, or multiple C. Specifically, combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B, and C," "one or more of A, B, and C," and "A, B, C, or any combination thereof" may be only A, only B, only C, A and B, A and C, B and C, or A and B and C, wherein any such combination may include one or more members of A, B, or C. All structural and functional equivalents of the elements of the various aspects described throughout the present disclosure that are currently or hereafter known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be covered by the claims. In addition, nothing disclosed herein is intended to be dedicated to the public, regardless of whether such disclosure is expressly recited in the claims. The terms “module,” “mechanism,” “element,” “device,” etc. may not be substitutes for the term “means.” Thus, no claim element should be construed as means-plus-function unless the element is explicitly recited using the phrase “means for….”

[0244] The following examples are merely illustrative, and aspects thereof may be combined with aspects of other examples or teachings described herein without limitation.

[0245] Aspect 1 is a method for wireless communication at a UE, comprising: generating a message indicating a processing time at the UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; and transmitting the message indicating the processing time to a first network entity.

[0246] In aspect 2, the method of aspect 1 further includes that the first network entity is a base station, a TRP or another UE.

[0247] In aspect 3, the method of aspect 1 or 2 further includes that the processing time corresponds to an amount of time between a first time associated with a reference signal used for training in the neural network and a second time when the UE has successfully trained the neural network.

[0248] In aspect 4, the method of any one of aspects 1-3 further includes that the neural network is successfully trained when the UE is able to report back CSI trained by the neural network or is able to use the trained weights of the neural network to achieve accuracy or QoS.

[0249] In aspect 5, the method of any one of aspects 1-4 further includes that the processing time corresponds to a time between the UE receiving a command to train the neural network and completing the training of the neural network.

[0250] In aspect 6, the method of any one of aspects 1-5 further comprises receiving a configuration based at least on the processing time from a second network entity; and transmitting the CSI to the second network entity based on the configuration.

[0251] In aspect 7, the method of any one of aspects 1-6 further includes that the second network entity is the same as the first network entity.

[0252] In aspect 8, the method of any one of aspects 1-7 further includes that the second network entity is different from the first network entity.

[0253] In aspect 9, the method of any one of aspects 1-8 further includes, the UE determining the processing time for training the neural network based on at least one of: the number of layers in the neural network, the number of weights in the neural network, or the type of one or more layers of the neural network.

[0254] In aspect 10, the method of any of aspects 1-9 further comprises that the processing time used to train the neural network is based on information from a previously trained neural network.

[0255] In aspect 11, the method of any of aspects 1-10 further includes that the information from the previously trained neural network includes at least one of QCL information from the previously trained neural network or an indication of a neural network state.

[0256] In aspect 12, the method of any of aspects 1-11 further comprises that the processing time for training the neural network is based on the amount of neural networks or layers to be trained.

[0257] In aspect 13, the method of any one of aspects 1-12 further includes that the processing time is based on whether a single neural network or multiple neural networks are to be trained simultaneously.

[0258] In aspect 14, the method of any one of aspects 1-13 further includes that the multiple neural networks are trained simultaneously based on concurrent training of at least one of the following: the same component carrier, the same frequency band, the same bandwidth part, the same frequency band combination, the same frequency range, the same time slot, the same subframe, or the same frame.

[0259] In aspect 15, the method of any one of aspects 1-14 further comprises that the neural networks of the plurality of neural networks are trained simultaneously until the UE responds with a training completion message.

[0260] In aspect 16, the method of any one of aspects 1-15 further includes that the processing time is based on whether a single layer or multiple layers of the neural network are to be trained simultaneously.

[0261] In aspect 17, the method of any one of aspects 1-16 further includes that the processing time for training the neural network is based on a sequential order of multiple layers of the neural network.

[0262] In aspect 18, the method of any of aspects 1-17 further comprises the processing time for training the neural network being based on a type of wireless signal procedure executed by the neural network.

[0263] In aspect 19, the method of any of aspects 1-18 further includes that the type of wireless signal procedure includes at least one of: channel state information determination, demodulation, positioning determination, or waveform determination.

[0264] In aspect 20, the method of any of aspects 1-19 further comprises that the processing time for training the neural network is based on an accuracy level.

[0265] In aspect 21, the method of any one of aspects 1-20 further includes that the accuracy level is based on QoS.

[0266] In aspect 22, the method of any of aspects 1-21 further includes that the processing time for reporting the CSF is based on at least one of: an encoder output vector, an encoder input vector, one or more vectors determined at the neural network, the number of layers in the neural network, the number of first elements in the input of the neural network, the number of second elements in the output of the neural network, the number of third elements in an intermediate vector of the neural network, the layer type of one or more layers of the neural network, the amount of neural network reported for overlap, or the sequential order of multiple layers of the neural network.

[0267] In aspect 23, the method of any one of aspects 1-22 further includes the UE reporting the processing time as UE capability, MAC-CE or UCI.

[0268] In aspect 24, the method of any one of aspects 1-23 further includes the UE reporting the processing time for at least one of: bandwidth part, parameter design, component carrier, frequency band, frequency band combination, frequency range, or one or more timeline factors.

[0269] In aspect 25, the method of any of aspects 1-24 further includes that the one or more timeline factors include at least one of the following: layer, layer type, layer combination, input vector length, output vector length, intermediate vector length, number of layers, or sequence of layers.

[0270] In aspect 26, the method of any one of aspects 1-25 further includes the UE reporting the first processing time and the second processing time.

[0271] In aspect 27, the method of any one of aspects 1-26 further comprises receiving, from the second network entity, a configuration to use the first processing time or the second processing time.

[0272] In aspect 28, the method of any one of aspects 1-27 further comprises applying the first processing time or the second processing time based on a power saving feature of the UE.

[0273] Aspect 29 is a device comprising one or more processors and one or more memories in electronic communication with the one or more processors, the one or more memories storing instructions executable by the one or more processors to cause the device to implement a method as in any one of Aspects 1-28.

[0274] Aspect 30 is a system or an apparatus comprising means for implementing the method of any one of aspects 1-28 or implementing the apparatus of any one of aspects 1-28.

[0275] Aspect 31 is a non-transitory computer-readable medium storing instructions, the instructions being executable by one or more processors to cause the one or more processors to implement the method of any one of aspects 1-28.

[0276] Aspect 32 is a method for wireless communication at a base station, comprising: receiving a processing time from a UE for at least one of: training a neural network for channel state information derivation or reporting CSI based on the trained neural network; transmitting a configuration to the UE based on the processing time; and receiving CSI from the UE based on the configuration.

[0277] In aspect 33, the method of aspect 32 further includes that the network entity is a base station, a TRP or another UE.

[0278] In aspect 34, the method of aspect 32 or aspect 33 further includes that the processing time corresponds to an amount of time between a first time associated with a reference signal for training in the neural network and a second time when the UE has successfully trained the neural network.

[0279] In aspect 35, the method of any one of aspects 32-34 further includes that the neural network is successfully trained when the UE is capable of reporting back CSI trained by the neural network or is capable of using trained weights of the neural network to achieve accuracy or QoS.

[0280] In aspect 36, the method of any one of aspects 32-35 further includes that the processing time corresponds to a time between the UE receiving a command to train the neural network and completing the training of the neural network.

[0281] In aspect 37, the method of any of aspects 32-36 further includes that the processing time for training the neural network is based on at least one of: the number of layers in the neural network, the number of weights in the neural network, or the type of one or more layers of the neural network.

[0282] In aspect 38, the method of any of aspects 32-37 further comprises that the processing time used to train the neural network is based on the use of information from a previously trained neural network.

[0283] In aspect 39, the method of any of aspects 32-38 further comprises that the information from the previously trained neural network comprises at least one of QCL information from the previously trained neural network or an indication of a neural network state.

[0284] In aspect 40, the method of any of aspects 32-39 further comprises that the processing time for training the neural network is based on the amount of neural networks or layers to be trained.

[0285] In aspect 41, the method of any one of aspects 32-40 further includes that the processing time is based on whether a single neural network or multiple neural networks are to be trained simultaneously.

[0286] In aspect 42, the method of any one of aspects 32-41 further includes that the multiple neural networks are trained simultaneously based on concurrent training of at least one of the following: the same component carrier, the same frequency band, the same bandwidth part, the same frequency band combination, the same frequency range, the same time slot, the same subframe, or the same frame.

[0287] In aspect 43, the method of any one of aspects 32-42 further comprises neural networks of the plurality of neural networks being trained simultaneously until the UE responds with a completed training message.

[0288] In aspect 44, the method of any one of aspects 32-43 further includes that the processing time is based on whether a single layer or multiple layers of the neural network are to be trained simultaneously.

[0289] In aspect 45, the method of any one of aspects 32-44 further comprises that the processing time for training the neural network is based on a sequential order of multiple layers of the neural network.

[0290] In aspect 46, the method of any one of aspects 32-45 further comprises the processing time for training the neural network being based on a type of wireless signal procedure executed by the neural network.

[0291] In aspect 47, the method of any one of aspects 32-46 further includes that the type of wireless signal procedure includes at least one of: channel state information determination, demodulation, positioning determination, or waveform determination.

[0292] In aspect 48, the method of any one of aspects 32-47 further comprises that the processing time for training the neural network is based on an accuracy level.

[0293] In aspect 49, the method of any one of aspects 32-48 further comprises that the accuracy level is based on QoS.

[0294] In aspect 50, the method of any of aspects 32-49 further includes that the processing time for reporting the CSF is based on at least one of: an encoder output vector, an encoder input vector, one or more vectors determined at the neural network, the number of layers in the neural network, the number of first elements in the input of the neural network, the number of second elements in the output of the neural network, the number of third elements in an intermediate vector of the neural network, the layer type of one or more layers of the neural network, the amount of neural network reported for overlap, or the sequential order of multiple layers of the neural network.

[0295] In aspect 51, the method of any one of aspects 32-50 further includes that the processing time is received as UE capability, MAC-CE or UCI.

[0296] In aspect 52, the method of any of aspects 32-51 further comprises that the processing time is received for at least one of: a bandwidth portion, a parameter design, a component carrier, a frequency band, a frequency band combination, a frequency range, or one or more timeline factors.

[0297] In aspect 53, the method of any of aspects 32-52 further includes that the one or more timeline factors include at least one of: a layer, a layer type, a layer combination, an input vector length, an output vector length, an intermediate vector length, a number of layers, or a sequence of layers.

[0298] In aspect 54, the method of any one of aspects 32-53 further includes the base station receiving the first processing time and the second processing time from the UE.

[0299] In aspect 55, the method of any one of aspects 32-54 further includes configuring the UE to use the first processing time or the second processing time.

[0300] Aspect 56 is a device comprising one or more processors and one or more memories in electronic communication with the one or more processors, the one or more memories storing instructions executable by the one or more processors to cause the device to implement a method as in any of Aspects 32-55.

[0301] Aspect 57 is a system or apparatus comprising means for implementing a method as in any one of Aspects 32-55 or implementing an apparatus as in any one of Aspects 32-55.

[0302] Aspect 58 is a non-transitory computer-readable medium storing instructions executable by one or more processors to cause the one or more processors to implement the method as in any of Aspects 32-55.

[0303] Aspect 59 is an apparatus for wireless communication, comprising a memory and at least one processor configured to perform the method of any one of aspects 32-55.

[0304] In aspect 60, the apparatus of aspect 59 further comprises at least one antenna and a transceiver coupled to the at least one antenna and the at least one processor.

[0305] Aspect 61 is an apparatus for wireless communication, comprising a memory and at least one processor configured to perform the method of any one of aspects 1-28.

[0306] In aspect 62, the apparatus of aspect 61 further comprises at least one antenna and a transceiver coupled to the at least one antenna and the at least one processor.

Claims

1. An apparatus for wireless communication at a user equipment (UE), include: Memory; as well as at least one processor coupled to the memory and configured to: generating a message indicating a processing time at the UE for training a neural network for channel state information (CSI) derivation, and the processing time may optionally include a time for reporting CSI based on the trained neural network, wherein the processing time corresponds to an amount of time between a first time associated with a reference signal used for training in the neural network and a second time at which the UE has successfully trained the neural network; as well as The message indicating the processing time is transmitted to a first network entity.

2. The apparatus of claim 1, wherein the first network entity is a base station, a transmission reception point (TRP), or another UE, and the apparatus further include: at least one antenna; as well as A transceiver is coupled to the at least one antenna and the at least one processor.

3. The apparatus of claim 1 , wherein the neural network is successfully trained when the UE is able to report back the CSI trained by the neural network or is able to use the trained weights of the neural network to achieve accuracy or quality of service (QoS).

4. The apparatus of claim 1 , wherein the processing time corresponds to a time between receiving a command to train the neural network and completing the training of the neural network.

5. The apparatus of claim 1 , wherein the memory and the at least one processor are further configured to: receiving, from a second network entity, a configuration based at least on the processing time; and Channel state information (CSI) is transmitted to the second network entity based on the configuration.

6. The apparatus of claim 1 , wherein the processing time for training the neural network is based on at least one of: The number of layers in the neural network, the number of weights in the neural network, or The type of one or more layers of the neural network.

7. The apparatus of claim 1, wherein the processing time for training the neural network is based on at least one of: at least one of quasi-co-location (QCL) information from a previously trained neural network or a neural network state indication.

8. The apparatus of claim 1, wherein the processing time for training the neural network is based on an amount of neural networks or layers to be trained.

9. The apparatus of claim 8, wherein the processing time is further based on whether a single neural network or multiple neural networks are to be trained simultaneously.

10. The apparatus of claim 9, wherein the processing time is based on training the plurality of neural networks, the plurality of neural networks being trained simultaneously based on concurrent training in at least one of: Same component carrier, Same frequency band, The same bandwidth part, Same frequency band combination, Same frequency range, Same time slot, The same subframe, or Same frame, Wherein the multiple neural networks are trained simultaneously until the UE responds with a completed training message.

11. The apparatus of claim 9, wherein the processing time is further based on whether a single layer or multiple layers of the neural network are to be trained simultaneously.

12. The apparatus of claim 1, wherein the processing time for training the neural network is based on a sequential order of a plurality of layers of the neural network.

13. The apparatus of claim 1 , wherein the processing time for training the neural network is based on a type of wireless signal procedure executed by the neural network, the type of wireless signal procedure comprising at least one of: Channel state information determination, demodulation, Positioning confirmed, or The waveform is determined.

14. The apparatus of claim 1, wherein the processing time used to train the neural network is based on a level of accuracy.

15. The apparatus of claim 1 , wherein the processing time for reporting the CSI is based on at least one of: The encoder output vector, Encoder input vector, one or more vectors determined by the neural network, The number of layers in the neural network, the number of first elements in the input of the neural network, the number of second elements in the output of the neural network, the number of third elements in the intermediate vector of the neural network, the layer type of one or more layers of the neural network, The amount of neural network reported for overlap, or The sequential order of the multiple layers of the neural network.

16. The apparatus of claim 1, wherein the processing time is for at least one of: Bandwidth part, Parameter design, Component carrier, frequency band, Band combination, Frequency range, or One or more timeline factors, and The one or more timeline factors include at least one of the following: layer, Layer type, The combination of layers, Input vector length, Output vector length, The length of the intermediate vector, The number of layers, or A sequence of layers.

17. The apparatus of claim 1, wherein the memory and the at least one processor are further configured to: reporting a first processing time and a second processing time, wherein the memory and the at least one memory are further configured to: receiving, from a second network entity, a configuration to use the first processing time or the second processing time; and The first processing time or the second processing time is applied based on a power saving feature of the UE.

18. A method of wireless communication at a user equipment (UE), include: determining a processing time at the UE for training a neural network for channel state information (CSI) derivation, and the processing time may optionally include a time for reporting CSI based on the trained neural network, wherein the processing time corresponds to an amount of time between a first time associated with a reference signal used for training in the neural network and a second time at which the UE has successfully trained the neural network; as well as The processing time is provided to the first network entity.

19. A computer readable medium storing computer executable code for wireless communication at a user equipment (UE), the code, when executed by a processor, causing the processor to: generating a message indicating a processing time at the UE for training a neural network for channel state information (CSI) derivation, and the processing time may optionally include a time for reporting CSI based on the trained neural network, wherein the processing time corresponds to an amount of time between a first time associated with a reference signal used for training in the neural network and a second time at which the UE has successfully trained the neural network; and The message indicating the processing time is transmitted to a first network entity.

20. An apparatus for wireless communication, include: Memory; as well as at least one processor coupled to the memory and configured to: receiving, from a user equipment (UE), a processing time for training a neural network for channel state information (CSI) derivation, and the processing time may optionally include a time for reporting CSI based on the trained neural network, wherein the processing time corresponds to an amount of time between a first time associated with a reference signal used for training in the neural network and a second time at which the UE has successfully trained the neural network; transmitting a configuration to the UE based on the processing time; as well as The CSI is received from the UE based on the configuration.

21. The apparatus of claim 20, wherein the wireless communication is at a base station, a transmission reception point (TRP), or another UE, the apparatus further comprising: include: at least one antenna; as well as A transceiver is coupled to the at least one antenna and the at least one processor.

22. The apparatus of claim 20, wherein the neural network is successfully trained when the UE is able to report back CSI trained by the neural network or is able to use trained weights of the neural network to achieve accuracy or quality of service (QoS).

23. The apparatus of claim 20, wherein the processing time corresponds to a time between receiving a command to train the neural network and completing the training of the neural network.

24. The apparatus of claim 20, wherein the processing time for training the neural network is based on at least one of: The number of layers in the neural network, the number of weights in the neural network, the type of one or more layers of the neural network, quasi-colocalization (QCL) information from a previously trained neural network, or An indication of a neural network state from the previously trained neural network.

25. The apparatus of claim 20, wherein the processing time for training the neural network is based on at least one of: an amount of neural networks or layers to be trained, and whether a single neural network or multiple neural networks are to be trained simultaneously, wherein the multiple neural networks are trained simultaneously based on concurrent training in at least one of: Same component carrier, Same frequency band, The same bandwidth part, Same frequency band combination, Same frequency range, Same time slot, The same subframe, or Same frame.

26. The apparatus of claim 20, wherein the processing time is based on at least one of whether a single layer or multiple layers of the neural network are to be trained simultaneously, a sequential order of the multiple layers of the neural network, a type of wireless signal procedure performed by the neural network, or a level of accuracy.

27. The apparatus of claim 20, wherein the processing time for reporting the CSI is based on at least one of: The encoder output vector, Encoder input vector, one or more vectors determined by the neural network, The number of layers in the neural network, the number of first elements in the input of the neural network, the number of second elements in the output of the neural network, the number of third elements in the intermediate vector of the neural network, the layer type of one or more layers of the neural network, The amount of neural network reported for overlap, or The sequential order of the multiple layers of the neural network.

28. The apparatus of claim 20, wherein the processing time is for at least one of: Bandwidth part, Parameter design, Component carrier, frequency band, Band combination, Frequency range, or One or more timeline factors, and wherein the one or more timeline factors include at least one of: layer, Layer type, The combination of layers, Input vector length, Output vector length, The length of the intermediate vector, The number of layers, or A sequence of layers.

29. The apparatus of claim 20, wherein the memory and the at least one processor are further configured to: receiving a first processing time and a second processing time from the UE; and The UE is configured to use the first processing time or the second processing time.

Images