Obtaining learning model parameters at a device

Abstract

Various aspects of the present disclosure relate to obtaining learning model parameters at a device. The device selects a subset of data samples from a set of data samples based on a determination to update one or more parameters of at least one normalization layer of an encoder learning model. The encoder learning model is part of a bilateral learning model that includes the encoder learning model and a decoder learning model. The device updates the parameters of the normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model. The device transmits a message to a second device for updating parameters of the decoder learning model. The message includes output associated with the subset of data samples from the encoder learning model.

Description

Related applications

[0001] This application claims priority to U.S. Provisional Application No. 63 / 632,786, filed April 11, 2024, entitled “OBTAINING LEARNING MODEL PARAMETERS AT A DEVICE”, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0002] This disclosure relates to wireless communication, and more specifically, to learning techniques for communication. Background Technology

[0003] A wireless communication system may include one or more network communication devices, which may otherwise be referred to as network equipment (NE), supporting wireless communication with one or more user communication devices, which may otherwise be referred to as user equipment (UE) or other suitable terms. The wireless communication system can support wireless communication with one or more user communication devices by utilizing the resources of the wireless communication system (e.g., time resources (e.g., symbols, time slots, subframes, frames, etc.) or frequency resources (e.g., subcarriers, carriers, etc.)). Furthermore, the wireless communication system can support wireless communication across various radio access technologies, including third-generation (3G), fourth-generation (4G), fifth-generation (5G), and other suitable radio access technologies above 5G (e.g., sixth-generation (6G)). Summary of the Invention

[0004] The article “a(a)” preceding an element is not limited and should be understood to mean “at least one” or “one or more” of those elements. The terms “a,” “at least one,” “one or more,” and “at least one of one or more” are interchangeable. As used herein (included in the claims), “or” as used in a list of items (for example, a list of items followed by phrases such as “at least one of…” or “one or more of…” or “one or both of…”) indicates a list of inclusion, such that (for example) a list of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (i.e., A and B and C). Furthermore, as used herein, the phrase “based on” should not be construed as a reference to a closed set of conditions. For example, an exemplary step described as “based on condition A” could be based on both condition A and condition B without departing from the scope of this disclosure. In other words, as used herein, the phrase “based on” should be interpreted in the same manner as the phrase “at least partially based on.” Furthermore, as used herein (included in the claims), a “set” may contain one or more elements.

[0005] A first device for wireless communication is described. The first device may be configured, enabled, or operable to perform one or more operations as described herein. For example, the first device may be configured, enabled, or operable to: select a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; update one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model; and transmit a message to a second device containing an output from the encoder learning model, wherein the output is associated with the subset of data samples.

[0006] A processor (e.g., a standalone processor chipset, or a component of a first device) is described. The processor may be configured, capable, or operable to perform one or more operations as described herein. For example, the processor may be configured, capable, or operable to: select a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; update one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model; and transmit a message to a second device containing the output from the encoder learning model, wherein the output is associated with the subset of data samples.

[0007] A method is described that is performed by or can be performed by a first device for wireless communication. The method may include: selecting a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; updating one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model; and transmitting a message containing an output from the encoder learning model, wherein the output is associated with the subset of data samples to a second device.

[0008] In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: update the one or more parameters on a periodic basis. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: receive from at least one of the second or third devices an instruction to the first device to update the one or more parameters. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: detect changes in one or more characteristics associated with Channel State Information (CSI). In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: detect changes in one or more channel conditions associated with the channel between the first device and the second device. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: determine that the quality of the output from the encoder learning model fails to meet a threshold.

[0009] In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, processor, and method may be further configured, capable of, or operable to: obtain the data sample set based on an estimated CSI. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, processor, and method may be further configured, capable of, or operable to: receive additional messages indicating the data sample set. In some embodiments of the first apparatus, processor, and method described herein, in order to select the data sample subset, the first apparatus, processor, and method may be further configured, capable of, or operable to: determine that the data sample subset satisfies one or more conditions, wherein the one or more conditions include a distance metric between corresponding data samples in the data sample subset being greater than a threshold dissimilarity value and the entropy value of the corresponding data samples in the data sample subset satisfying a threshold entropy value. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, processor, and method may be further configured, capable of, or operable to: determine one or more of the threshold dissimilarity value or the threshold entropy value. In some embodiments of the first apparatus, processor, and method described herein, the first apparatus, processor, and method may be further configured, capable of, or operable to: receive additional messages indicating one or more of the threshold dissimilarity value or the threshold entropy value. In some embodiments of the first apparatus, processor, and method described herein, in order to update the one or more parameters, the first apparatus, processor, and method may be further configured, capable of, or operable to: propagate the subset of data samples from a first layer associated with the encoder learning model to a last layer associated with the encoder learning model; and calculate one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. In some embodiments of the first apparatus, processor, and method described herein, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. In some embodiments of the first apparatus, processor, and method described herein, the at least one normalization layer is a batch normalization layer.

[0010] A first device for wireless communication is described. The first device may be configured, capable, or operable to perform one or more operations as described herein. For example, the first device may be configured, capable, or operable to: receive from a second device a message containing output from an encoder learning model, wherein: the output is associated with a subset of data samples in a set of data samples; and the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; and update one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model.

[0011] A processor (e.g., a standalone processor chipset, or a component of a first device) is described. The processor may be configured, capable, or operable to perform one or more operations as described herein. For example, the processor may be configured, capable, or operable to: receive from a second device a message containing output from an encoder learning model, wherein: the output is associated with a subset of data samples in a set of data samples; and the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; and update one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model.

[0012] A method is described that is performed by or can be performed by a first device for wireless communication. The method may include: receiving from a second device a message containing output from an encoder learning model, wherein: the output is associated with a subset of data samples in a set of data samples; and the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; and updating one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model.

[0013] In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: update the one or more parameters on a periodic basis. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: receive from at least one of the second or third devices an instruction to the first device to update the one or more parameters. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: detect changes in one or more channel conditions associated with the channel between the first device and the second device. In some embodiments of the first device, the processor, and the method described herein, in order to update the one or more parameters, the first device, the processor, and the method may be further configured, capable of, or operable to: determine that the quality of the output from the decoder learning model fails to meet a threshold.

[0014] In some embodiments of the first apparatus, processor, and method described herein, in order to update the one or more parameters, the first apparatus, processor, and method may be further configured, capable of, or operable to: propagate the subset of data samples from a first layer associated with the decoder learning model to a last layer associated with the decoder learning model; and compute one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. In some embodiments of the first apparatus, processor, and method described herein, the output comprises a set of information samples corresponding to feature vectors associated with the encoder learning model. In some embodiments of the first apparatus, processor, and method described herein, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, and wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. In some embodiments of the first apparatus, processor, and method described herein, the at least one normalization layer is a batch normalization layer. Attached Figure Description

[0015] Figure 1 Illustrated examples of wireless communication systems according to various aspects of this disclosure.

[0016] Figure 2 Illustrated examples of learning model diagrams based on various aspects of this disclosure.

[0017] Figure 3 Illustrated examples of wireless communication systems according to various aspects of this disclosure.

[0018] Figure 4 Illustrated examples of signaling diagrams according to various aspects of this disclosure.

[0019] Figure 5 Illustrated examples of UEs based on various aspects of this disclosure.

[0020] Figure 6 Illustrated examples of processors according to various aspects of this disclosure.

[0021] Figure 7 Illustrated examples of NEs according to various aspects of this disclosure.

[0022] Figure 8 The diagram illustrates a flowchart of a method performed by a UE according to various aspects of this disclosure.

[0023] Figure 9 The diagram illustrates a flowchart of a method performed by an NE according to various aspects of this disclosure. Detailed Implementation

[0024] A wireless communication system may include one or more devices, such as a UE and an NE, and other devices for transmitting and receiving signaling. The devices may transmit and receive signaling via one or more channels between the devices, where a channel refers to the medium through which information is transmitted between the devices. The devices may exchange channel-related information, such as CSI, in time, frequency, and space. The devices use this information to perform coherent reception and / or transmission of control information and / or data between the devices. Coherent reception and / or transmission refers to maintaining phase synchronization between transmitted and received signals, which enables more accurate signal detection and demodulation, and other advantages. The devices may use an uplink reference signal (e.g., transmitted from the UE to the NE) to obtain the downlink CSI, where uplink-downlink channel reciprocity exists. Uplink-downlink channel reciprocity occurs when the channel characteristics observed for uplink transmission from the UE to the NE (e.g., channel gain, phase shift, and fading conditions) are approximately equivalent to the channel characteristics observed for downlink transmission from the NE to the UE. Alternatively, in the absence of uplink-downlink channel reciprocity, the apparatus may measure one or more downlink reference signals (e.g., transmitted from the NE to the UE) to obtain CSI, wherein the UE may transmit a message indicating the CSI to the NE. To reduce signaling overhead, the UE may compress the CSI to be transmitted to the NE. That is, the UE may compress the CSI report (e.g., minimize the complexity of the CSI report and / or encode it into a more compact representation) to include the CSI information provided in the CSI report.

[0025] In some variations, one or more devices (e.g., UE and NE) in a wireless communication system may implement one or more learning models (which may be instances of machine learning (ML) models and / or artificial intelligence (AI) models) to perform CSI compression. For example, the devices may implement a two-sided CSI compression model comprising an encoder component at the UE that generates compressed CSI data and a decoder component at the NE that generates estimated CSI data from signaling indicating compressed CSI data. In some instances, the NE or other devices train the learning model on a training dataset containing CSI data samples, wherein training involves obtaining one or more parameters of the learning model by inputting the training dataset (e.g., for the encoder and decoder components). For example, the NE obtains one or more parameters of the learning model that minimize a loss function for the training dataset, wherein the loss function calculates the error between the predicted output and the true output of the learning model. One or more devices use these parameters to perform (e.g., process, implement) the learning model. However, when executing a learning model, the device may input data with different statistical properties than the dataset used to train the learning model, leading to a performance loss in the learning model (e.g., incorrect inferences and predictions). This performance loss can result in increased processing resource usage at the device executing the model, as well as increased signaling overhead due to the correction of errors in the learning model's output caused by distribution mismatch.

[0026] In some scenarios, the NE can train multiple different learning models for a given domain, which infer or predict the domain, where the domain represents data on parameters, physical propagation media, or device behavior for a scenario, configuration, or network. The NE can store parameters of the different learning models for transmission to one or more other devices (e.g., a UE or other device with reduced memory storage capacity compared to the NE). Alternatively, one or more other devices can receive and store parameters of different learning models. Other devices can use the appropriate learning model for different domains. However, transmitting and / or storing parameters of learning models for different domains leads to increased signaling overhead and memory usage, as well as inefficient use of time-frequency resources, because the learning models have a relatively large (e.g., greater than a threshold) number of parameters.

[0027] As described herein, to reduce performance loss without increasing processing and signaling overhead due to the transmission of parameters from multiple learning models, one or more devices may generate parameters for the learning models to adapt them to different data distributions of the input data. For example, a transmitting device, acting as a signaling transmission device, may execute an encoder component of a two-sided CSI compression model, and a receiving device, acting as a signaling reception device, may execute a decoder component of the two-sided CSI compression model. The encoder and decoder components may comprise multiple layers, such as hidden layers (e.g., neural layers) and normalization layers. Each layer of the learning model is defined by one or more parameters, which may include weights, biases, and / or other parameters. The normalization layer may contain a relatively small number of parameters (e.g., less than a threshold) compared to the hidden layers. The transmitting and / or receiving devices may determine to update one or more parameters of the normalization layer of the encoder component (e.g., based on periodicity, by detecting changes in the characteristics of the estimated CSI, by detecting changes in the performance of the encoder component, and / or upon receiving signaling instructing the transmitting device to update the encoder neural network, and other instances).

[0028] In some instances, the transmitting device can update the parameters of the normalization layer of the encoder component by providing a subset of data samples obtained by the transmitting device as input to the encoder component. The transmitting device can transmit the output from the encoder component to the receiving device. Alternatively or concurrently, the receiving device can use the output from the encoder component to update the parameters of the normalization layer of the decoder component. For example, the receiving device can provide the output from the encoder component as input to the decoder learning model to update the parameters of the normalization layer. The transmitting and / or receiving devices can execute the encoder and decoder components to perform CSI compression on the remaining data samples obtained by the transmitting device. This reduces the performance loss of the encoder and decoder components by adapting the parameters to the statistical characteristics of the subset of data samples, without increasing processing and signaling overhead due to transmitting parameters of multiple learning models.

[0029] This document refers to the receiving, transmission, or communication of data or information, such as signaling communication resources and / or communication transmitted or received between devices. It should be understood that other terms may be used interchangeably with "communication," such as signaling, transmission, receiving, output, forwarding, retrieval, acquisition, etc. Similarly, other terms may be used interchangeably with "transmission" (e.g., communication, signaling, output, forwarding, etc.), and other terms may be used interchangeably with "receive" (e.g., communication, retrieval, acquisition, etc.).

[0030] Various aspects of this disclosure are described in the context of wireless communication systems.

[0031] Figure 1The illustration depicts examples of a wireless communication system 100 according to various aspects of this disclosure. The wireless communication system 100 may include one or more NEs 102, one or more UEs 104, and a core network (CN) 106. The wireless communication system 100 may support various radio access technologies. In some embodiments, the wireless communication system 100 may be a 4G network, such as an LTE network or an LTE-A advanced network. In some other embodiments, the wireless communication system 100 may be an NR network, such as a 5G network, an 5G-A advanced network, or a 5G ultra-wideband (5G-UWB) network. In other embodiments, the wireless communication system 100 may be a combination of 4G and 5G networks, or other suitable radio access technologies, including IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), and IEEE 802.20. The wireless communication system 100 may support radio access technologies beyond 5G, such as 6G. In addition, the wireless communication system 100 can support technologies such as Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), or Code Division Multiple Access (CDMA).

[0032] One or more NEs 102 may be distributed throughout a geographic area to form a wireless communication system 100. One or more of the NEs 102 described herein may be, include, or be referred to as a network node, base station, access point (AP), network element, network function, network entity, network infrastructure (or infrastructure), radio access network (RAN), NodeB, eNodeB (eNB), next-generation NodeB (gNB), or other suitable terms. NEs 102 and UEs 104 may communicate via a communication link, which may be a wireless or wired connection. For example, NEs 102 and UEs 104 may perform wireless communication (e.g., receiving signaling, transmitting signaling) via a Uu interface.

[0033] NE 102 can provide a geographic coverage area, and NE 102 can support service to one or more UEs 104 within the geographic coverage area. For example, NE 102 and UE 104 can support wireless communication of signals related to services (e.g., voice, video, packet data, message sending and receiving, broadcasting, etc.) according to one or more radio access technologies. In some embodiments, NE 102 can be mobile, such as a satellite associated with a non-terrestrial network (NTN). In some embodiments, different geographic coverage areas associated with the same or different radio access technologies can overlap, but different geographic coverage areas can be associated with different NEs 102.

[0034] One or more UEs 104 may be distributed throughout the geographic area of ​​the wireless communication system 100. UE 104 may include or be referred to as a remote unit, mobile device, wireless device, remote device, subscriber device, transmitter device, receiver device, or some other suitable term. In some embodiments, UE 104 may be referred to as a unit, station, terminal, or client, and other instances thereof. Additionally or alternatively, UE 104 may be referred to as an Internet of Things (IoT) device, an Internet of Everything (IoE) device, or a Machine Type Communication (MTC) device, and other instances thereof.

[0035] UE 104 may be able to support direct wireless communication with other UE 104 via a communication link. For example, UE 104 may support direct wireless communication with another UE 104 via a device-to-device (D2D) communication link. In some implementations, such as vehicle-to-vehicle (V2V) deployments, vehicle-to-everything (V2X) deployments, or cellular V2X deployments, the communication link may be referred to as a sidelink. For example, UE 104 may support direct wireless communication with another UE 104 via a PC5 interface.

[0036] NE 102 may support communication with CN 106, another NE 102, or both. For example, NE 102 may interface with other NE 102 or CN 106 via one or more backhaul links (e.g., S1, N2, N6, or another network interface). In some embodiments, NE 102 may communicate directly with each other. In other embodiments, NE 102 may communicate indirectly with each other (e.g., via CN 106). In some embodiments, one or more NE 102 may include sub-components such as access network entities, which may be instances of access node controllers (ANCs). The ANC may communicate with one or more UE 104s via one or more other access network transport entities, which may be referred to as radio headends, smart radio headends, or transmit-receive points (TRPs).

[0037] CN 106 can support user authentication, access authorization, tracking, connectivity, and other access, routing, or mobility functions. CN106 can be an evolved packet core (EPC) or a 5G core (5GC), which may include control plane entities that manage access and mobility (e.g., a mobility management entity (MME), access and mobility management functions (AMF)) and user plane entities that route packets or interconnections to external networks (e.g., a serving gateway (S-GW), a packet data network (PDN) gateway (P-GW), or a user plane function (UPF)). In some implementations, the control plane entities may manage non-access plane (NAS) functions, such as mobility, authentication, and bearer management (e.g., data bearers, signaling bearers, etc.) for one or more UEs 104 served by one or more NEs 102 associated with CN106.

[0038] CN 106 may communicate with the packet data network via one or more backhaul links (e.g., via S1, N2, N6, or another network interface). The packet data network may contain an application server. In some implementations, one or more UEs 104 may communicate with the application server. UE 104 may establish a session (e.g., a Protocol Data Unit (PDU) session, etc.) with CN 106 via NE 102. CN 106 may use the established session (e.g., an established PDU session) to route services (e.g., control information, data, etc.) between UE 104 and the application server. A PDU session may be an instance of a logical connection between UE 104 and CN 106 (e.g., one or more network functions of CN 106).

[0039] In the wireless communication system 100, NE 102 and UE 104 can use the resources of the wireless communication system 100 (e.g., time resources (e.g., symbols, time slots, subframes, frames, etc.) or frequency resources (e.g., subcarriers, carriers)) to perform various operations (e.g., wireless communication). In some embodiments, NE 102 and UE 104 can support different resource structures. For example, NE 102 and UE 104 can support different frame structures. In some embodiments, such as in 4G, NE 102 and UE 104 can support a single frame structure. In some other embodiments, such as in 5G and other suitable radio access technologies, NE 102 and UE 104 can support various frame structures (i.e., multiple frame structures). NE 102 and UE 104 can support various frame structures based on one or more parameter sets (numerologies).

[0040] The wireless communication system 100 may support one or more parameter sets, and the parameter sets may include subcarrier spacing and cyclic prefixes. A first parameter set (e.g., μ=0) may be associated with a first subcarrier spacing (e.g., 15 kHz) and a normal cyclic prefix. In some embodiments, the first parameter set (e.g., μ=0) associated with the first subcarrier spacing (e.g., 15 kHz) may utilize one time slot per subframe. A second parameter set (e.g., μ=1) may be associated with a second subcarrier spacing (e.g., 30 kHz) and a normal cyclic prefix. A third parameter set (e.g., μ=2) may be associated with a third subcarrier spacing (e.g., 60 kHz) and a normal cyclic prefix or an extended cyclic prefix. A fourth parameter set (e.g., μ=3) may be associated with a fourth subcarrier spacing (e.g., 120 kHz) and a normal cyclic prefix. A fifth parameter set (e.g., μ=4) may be associated with a fifth subcarrier spacing (e.g., 240 kHz) and a normal cyclic prefix.

[0041] Time intervals for resources (e.g., communication resources) can be organized according to frames (also referred to as radio frames). Each frame may have a certain duration, such as 10 milliseconds (ms). In some embodiments, each frame may contain multiple subframes. For example, each frame may contain 10 subframes, and each subframe may have a certain duration, such as 1 ms. In some embodiments, each frame may have the same duration. In some embodiments, each subframe of a frame may have the same duration.

[0042] Alternatively or concurrently, the time intervals of resources (e.g., communication resources) can be organized according to time slots. For example, a subframe may contain a certain number (e.g., quantity) of time slots. The number of time slots in each subframe may also depend on one or more parameter sets supported in the wireless communication system 100. For example, a first parameter set, a second parameter set, a third parameter set, a fourth parameter set, and a fifth parameter set (i.e., μ=0, μ=1, μ=2, μ=3, μ=4) associated with corresponding subcarrier intervals of 15 kHz, 30 kHz, 60 kHz, 120 kHz, and 240 kHz may utilize one time slot per subframe, two time slots per subframe, four time slots per subframe, eight time slots per subframe, and 16 time slots per subframe, respectively. Each time slot may contain a certain number (e.g., quantity) of symbols (e.g., Orthogonal Frequency Division Multiplexing (OFDM) symbols). In some embodiments, the number (e.g., quantity) of time slots in a subframe may depend on the parameter set. For a normal cyclic prefix, a time slot may contain 14 symbols. For an extended cyclic prefix (e.g., for a 60 kHz subcarrier spacing), a time slot may contain 12 symbols. The relationships between the number of symbols per time slot, the number of time slots per subframe, and the number of time slots per frame for both normal and extended cyclic prefixes may depend on the parameter set. It should be understood that a reference to the first parameter set (e.g., μ=0) associated with the first subcarrier spacing (e.g., 15 kHz) may be used interchangeably between subframes and time slots.

[0043] In the wireless communication system 100, the electromagnetic (EM) spectrum can be divided into various categories, frequency bands, channels, etc., based on frequency or wavelength. By way of example, the wireless communication system 100 can support one or more operating frequency bands, such as frequency ranges specified as FR1 (410 MHz to 7.125 GHz), FR2 (24.25 GHz to 52.6 GHz), FR3 (7.125 GHz to 24.25 GHz), FR4 (52.6 GHz to 114.25 GHz), FR4a or FR4-1 (52.6 GHz to 71 GHz), and FR5 (114.25 GHz to 300 GHz). In some embodiments, NE 102 and UE 104 can perform wireless communication on one or more of the said operating frequency bands. In some embodiments, FR1 can be used by NE 102 and UE 104, as well as other equipment or devices for cellular communication services (e.g., control information, data). In some implementations, FR2 may be used by NE 102 and UE 104, as well as other equipment or devices for short-range, high-data-rate capabilities.

[0044] FR1 may be associated with one or more parameter sets (e.g., at least three parameter sets). For example, FR1 may be associated with a first parameter set containing a 15 kHz subcarrier spacing (e.g., μ=0), a second parameter set containing a 30 kHz subcarrier spacing (e.g., μ=1), and a third parameter set containing a 60 kHz subcarrier spacing (e.g., μ=2). FR2 may be associated with one or more parameter sets (e.g., at least two parameter sets). For example, FR2 may be associated with a third parameter set containing a 60 kHz subcarrier spacing (e.g., μ=2) and a fourth parameter set containing a 120 kHz subcarrier spacing (e.g., μ=3).

[0045] In some instances, the wireless communication system 100 may include one or more transmitting devices that transmit signaling to one or more devices (e.g., receiving devices) that receive signaling (including data and / or control signaling). Exemplary transmitting devices include, but are not limited to, NE 102 and / or UE 104. Alternatively, exemplary receiving devices include, but are not limited to, NE 102 and / or UE 104. The transmitting and receiving devices may communicate via one or more channels (e.g., wireless channels). A channel contains the medium through which signaling propagates between the transmitting and receiving devices. In some variations, the transmitting and / or receiving devices may have multiple antennas. For example, UE 104 may have r antennas and the NE (e.g., a base station or gNB) may have t antennas. The channel between NE 102 and UE 104 has a total of rt paths. A path defines the route or trajectory of electromagnetic waves from the transmitting device to the receiving device. Due to propagation (e.g., reflection, diffraction, and scattering), signals can follow one or more paths from the transmitting device to the receiving device.

[0046] In the downlink communication where NE 102 transmits signaling to UE 104, the discrete-time channel can be represented as an r×t dimensional complex-valued matrix H, where the elements h of H are... ij This represents the complex channel gain between the i-th receiving antenna and the j-th transmitting antenna. A discrete-time channel is a channel in which transmitted and received signals are represented and processed in discrete moments or intervals. That is, the wireless communication system 100 uses discrete-time samples of the signal rather than continuous-time signals. The channel gain, or channel matrix H, depends on the physical propagation medium. Due to the dynamic nature of the physical propagation medium, the wireless channel is a time-varying channel. Furthermore, the channel gain depends on the frequency used by the device for signaling. For example, for a multi-carrier waveform such as OFDM, the channel matrix can have different values ​​for different subcarriers (e.g., frequencies) at the same time. In other words, the wireless channel matrix H is random and varies across time, frequency, and spatial dimensions.

[0047] In some instances, one or more devices in the wireless communication system 100 may adapt the transmission method and / or preprocess the signaling to be transmitted according to the current channel implementation to increase the signaling throughput on the communication link while improving the reliability of the communication link. The transmission device may use CSI to implement adaptive transmission or perform preprocessing at the transmission device. For example, when the characteristics of the channel change, the transmission device may determine the channel matrix H within the operating frequency range (e.g., at the corresponding subcarrier for OFDM and / or multicarrier waveforms).

[0048] The receiving device can estimate the channel using reference signals and / or pilot signals transmitted by the transmitting device. The transmitting device can obtain (e.g., acquire, determine) the CSI by receiving signaling indicating the CSI from the receiving device. For example, the receiving device can measure the CSI and transmit the CSI to the transmitting device via a feedback channel. Therefore, in downlink communication (e.g., from NE102 to UE 104), UE 104 estimates the downlink CSI by performing one or more measurements on the pilot signals and / or reference signals transmitted from NE 102. The CSI estimation may include a channel matrix and / or a channel covariance matrix, wherein the elements of the channel matrix include channel gain and phase shift. In some variations, UE 104 may measure one or more pilot signals and / or reference signals (e.g., from NE 102) to obtain the CSI. In some cases, the CSI includes information describing the characteristics and conditions of the communication channel between the transmitting device (e.g., UE 104 and / or NE 102) and the receiving device (e.g., NE 102 and / or UE 104) in the wireless communication system 100. For example, the CSI data and corresponding data samples may include, but are not limited to, channel gain data, phase shift data, frequency response data, delay spread data, and Doppler shift data, as well as other information. The UE transmits the estimated CSI to NE 102.

[0049] Alternatively, for uplink communication from UE 104 to NE 102, NE 102 uses uplink reference signals and / or pilot signals from UE 104 to estimate the CSI. NE 102 then transmits the CSI to UE 104. However, the transmission of CSI by NE 102 and / or UE 104 to UE 104 or NE 102 respectively results in relatively high signaling overhead (e.g., exceeding a threshold) because the CSI does not contain data. To reduce the signaling overhead associated with CSI transmission, while ensuring that NE 102 obtains a sufficiently high quality CSI to enable NE 102 to adapt communication for the link, UE 104 may compress the CSI before transmitting it to NE 102. For example, once the receiving device has estimated the CSI, it may compress the estimated CSI and transmit the compressed version of the CSI to the transmitting device.

[0050] In some variations, the transmitting and receiving devices may implement or execute a two-sided learning model for CSI compression. The two-sided learning model comprises two neural networks: an encoder neural network (e.g., an encoder component) and a decoder neural network (e.g., a decoder component). The encoder component computes a low-dimensional feature vector (e.g., a low-dimensional latent representation) of the input CSI. The decoder component uses the low-dimensional feature vector to reconstruct the CSI. By deploying the encoder component at the receiving device and the decoder component at the transmitting device, the feedback from the receiving device to the CSI is equivalent to feedback to the low-dimensional feature vector computed by the encoder at the receiving device. Therefore, the two-sided learning model implements CSI compression for transmitting CSI from the receiving device to the transmitting device, as described above. Figure 2 To describe in further detail.

[0051] In some instances, one or more devices executing a learning model, whether at one of the transmitting or receiving devices (e.g., unilateral) or both (e.g., bilateral), are expected to perform the learning model with the same level of inference and / or prediction accuracy and / or precision as during the training and testing phases of the learning model. However, in real-world wireless networks, the learning model may make predictions and / or inferences based on input data with different statistical properties compared to the dataset used to train the learning model. That is, after training, testing, and deploying the learning model at the device, the learning model may output inferences and / or predictions based on input data with one or more different statistical data compared to the input data used during the training and testing phases of the learning model. Therefore, the learning model may output erroneous and / or incorrect inferences and / or predictions. To reduce errors generated by the learning model, a relatively general dataset can be used to train the learning model. However, due to the generalized data, the performance (e.g., accuracy or precision) of the learning model may degrade for the one or more domains in which the learning model is implemented.

[0052] In some instances, the means of implementing an encoder component (e.g., an autoencoder or other two-sided neural network) for a two-sided learning model for CSI compression can provide the estimated CSI as input to the encoder component. In some cases, the input is a function of the estimated CSI matrix H, such as the covariance matrix of the channel matrix, the eigenvectors of the CSI matrix, the singular vectors of the CSI matrix, the pre-encoder matrix calculated from the CSI, or some other function of the estimated CSI. One or more channel characteristics may be time-varying and affected by various factors, such as physical environment (e.g., urban, densely populated, rural), device mobility (e.g., stationary UE 104, low-speed UE 104, or high-speed UE), device location (e.g., indoor, outdoor), weather and / or seasonal conditions (fog, rain, vegetation, etc.) and / or the antenna array configuration used by NE 102 and UE 104, and other factors. Due to the factors affecting the channel characteristics or implementation, one or more marginal distributions of the estimated CSI and the marginal distributions of the function of the estimated CSI may change over time. The training dataset used by the device to train a two-sided model for CSI compression may not contain one or more possible channel variations. Therefore, the device that performs the two-sided learning model trained for CSI compression can obtain input data with a distribution different from that of the training data.

[0053] In some instances, one or more devices can adapt a two-sided CSI compressed learning model when the distribution of the input data changes relative to the distribution of the training data. Unlike one-sided learning models, two-sided learning models are deployed or executed on geographically separated devices after training. For example, the encoder component is deployed at UE 104 and the decoder component is deployed at NE 102, and vice versa. However, without using additional network resources (e.g., for sending the output from the decoder component to the encoder component), the device adapting the learning model to the distribution of the input data at the encoder component may not be able to access the output of the decoder component.

[0054] According to the implementation scheme, one or more of NE 102 and UE 104 are operable to implement various aspects of the technology described with reference to this disclosure. For example, NE 102 (e.g., a base station) and / or UE 104 may execute a two-sided learning model that adapts to a changing distribution of input data without using additional network resources. Model adaptation of a learning model refers to when the apparatus changes or adjusts one or more parameters or weights of a trained learning model. Alternatively, the process by which the apparatus adapts a learning model from one domain to another is referred to as domain adaptation and / or domain generalization. The apparatus (e.g., NE 102 and / or UE 104) may adapt and / or adjust one or more parameters of the learning model to improve the performance (e.g., accuracy and / or precision) of the learning model in response to a changing distribution of input data. The apparatus may adapt and / or adjust parameters without accessing the output of the learning model (e.g., from a decoder component) and does not involve calculating a loss function, calculating gradients, and / or performing backpropagation and other computationally complex tasks. Although the learning model is described in the context of CSI compression, the device may implement a learning model that can be adapted to the changing distribution of input data for other tasks in the wireless communication system 100.

[0055] Figure 2 The diagram illustrates an example of a learning model diagram 200 according to various aspects of this disclosure. In some instances, the learning model diagram 200 implements various aspects of the wireless communication system 100. For example, the learning model diagram 200 may implement, or be implemented by, a UE and / or NE, which may be a reference. Figure 1 Examples of UE 104 and NE 102 are described. Learning model diagram 200 illustrates one or more instances of learning model 202. In some cases, learning model 202 may comprise a single learning model 202. In other cases, learning model 202 may comprise multiple learning models 202.

[0056] In some instances, the learning model 202 may be an instance of an ML model and / or an AI model. The learning model may be an algorithm containing learnable parameters, such as a support vector machine or a decision tree. Alternatively or concurrently, the learning model may be part of a neural network, where neuron weights are learned parameters. For example, one or more learning models 202 may be instances of neural networks with multiple layers. The learning model may contain a bilateral deep neural network (DNN), such as an autoencoder, which will refer to... Figure 3A further detailed description is provided. A neural network is a computational model containing multiple layers of artificial neurons, which may be referred to as nodes or units, organized into an input layer 204, one or more hidden layers 206, and an output layer 208. A DNN is a type of learning model 202 with multiple hidden layers 206 between the input layer 204 and the output layer 208. The input layer 204 passes input data to subsequent layers and may not contain learnable parameters. The output layer 208 represents the predictions generated by the neural network and may not contain learnable parameters. Nodes in the hidden layers 206 receive input signals, perform mathematical operations on the input data, and produce output signals, which are then passed to other nodes in subsequent layers. Connections between nodes are represented by weighted edges, which determine the strength of the connections between nodes.

[0057] In some instances, one or more parameters of hidden layer 206 are updated during the training process. Exemplary parameters of hidden layer 206 include, but are not limited to, weights, biases, activation function parameters, and other parameters. Weight parameters represent the connection strength between neurons in different layers of the learning model. Bias is an additional parameter added to neurons in the learning model, ensuring that the learning model captures offsets or biases in the input data. Activation function parameters may include slope parameters or parameters defining the shape of the activation function in a parameterized activation function. Determining one or more values ​​for the model parameters for a defined use case is called training the model or learning the model. A general procedure for developing a learning model involves updating one or more parameters based on a training dataset containing labeled or unlabeled samples to minimize a cost function, resulting in supervised learning or training or unsupervised learning or training, respectively. For example, weights are adjusted using input-output pairs from the training dataset with the goal of minimizing a defined loss or error function. Neural networks are capable of learning complex patterns and representations from data, enabling them to perform a wide range of tasks, including classification, regression, clustering, pattern recognition, and sequence generation. Exemplary neural networks include, but are not limited to, autoencoders (e.g., regarding...). Figure 3 (For further detailed description), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Generative Adversarial Networks (GAN), Long Short-Term Memory (LSTM) networks, or any other type of neural network.

[0058] Alternatively or alternatively, the learning model 202 may include one or more normalization layers 210. Normalization or standardization refers to scaling and shifting a set of data samples to have zero mean and unit variance. Alternatively or alternatively, normalization may refer to scaling a set of data samples to maintain the values ​​of the set of data samples within a defined range (e.g., 0 to 1 or -1 to 1) without changing the mean and variance of the data samples. Standardization may refer to modifying a set of data samples to have zero mean and unit variance. However, the terms normalization and standardization are used synonymously, wherein normalization or standardization is defined herein as transforming a set of data samples to have zero mean and unit variance. In some cases, the device performs input data normalization or standardization independently of the neural network (e.g., outside the neural network). Alternatively or alternatively, normalization occurs within the neural network (e.g., inside a deep neural network).

[0059] In some cases, such as for a four-layer learning model 202 with one input layer 204, one output layer 208, and two hidden layers 206, when a normalization layer 210 is introduced, there may be one normalization layer 210 between the input layer 204 and the first hidden layer 206, another normalization layer 210 between the first hidden layer and the second hidden layer 206, and yet another normalization layer 210 between the second hidden layer 206 and the output layer 208. Therefore, excluding the possibility of using a normalization layer 210 before the input layer 204 of the learning model 202, three normalization layers 210 may exist in the four-layer learning model 202. Therefore, for a total number N... L The learning model 202 has multiple layers, with a normalization layer 210 between each pair of hidden layers 206. The number of normalized layers is 210.

[0060] For normalization occurring within the neural network (e.g., learning model 202), the device may feed input data samples to the first layer of the neural network at different times, which may be referred to as input layer 204. If the input data samples have a relatively high dynamic range, where the number of input samples with values ​​greater than a threshold value is less than a threshold value compared to other data samples in the dataset, then the learned weights and instantaneous gradient values ​​of the layers of learning model 202 fluctuate. Fluctuations in instantaneous gradient values ​​cause the neural network to converge to one or more optimal weights (e.g., weights that maximize the performance of the learning model (including precision and accuracy)) for a relatively long time (e.g., a duration greater than the threshold). In some cases, if the device normalizes the input weights to have zero mean and unit variance, then learning model 202 converges faster during training. Hidden layer 206 (e.g., hidden layer 206 after input layer 204) receives input from the previous layer. The inputs to the second and higher layers of learning model 202 (e.g., the layer containing the second layer to the output layer 208) are the outputs from other layers. Therefore, normalizing the output sample values ​​of input layer 204 and subsequent hidden layer 206 to have zero mean and unit variance can reduce the convergence time. For example, the inputs of the corresponding layers after input layer 204 (e.g., hidden layer 206 and / or output layer 208) can be normalized to zero mean and unit variance.

[0061] In some variations, normalization within the learning model 202 is also referred to as feature normalization or activation normalization within the learning model 202. The apparatus can perform batch normalization, layer normalization, group normalization, instance normalization, and / or root mean square normalization, among others. Batch normalization reduces the training time of the learning model 202 and stabilizes model weights in a relatively short time (e.g., less than a threshold duration) by addressing internal covariate shifts. Batch normalization scales the input data to have a consistent range or distribution, such that features in a portion of a batch of training samples (e.g., a subset of a batch of data samples) are scaled by the mean and variance computed within said portion of the batch of training data samples. In other words, batch normalization normalizes the corresponding features within a portion of a batch of training data samples.

[0062] Batch normalization is performed between two or more layers of the learning model 202 by inserting another layer called a batch normalization layer. The batch normalization layer receives input from the preceding network layer and provides (e.g., output) a batch-normalized (e.g., normalized) version of the input to the subsequent network layer. The learning model 202 can use the mean β and variance γ. 2(For example, instead of using zero mean and unit variance) to perform well, the learning model 202 can learn and optimize the mean and variance during the training process. For instance, the batch normalization layer can introduce two parameters, β and γ, which are learned by the learning model 202 during training. In some instances, x i The feature is computed by the layer, and i is the index. In the case of a two-dimensional input (e.g., an image with non-zero width and non-zero height), the index i is four-dimensional, where i = (i B i C i H i W ), where i B i C i H andi W These represent the batch index, channel index, height index, and width index, respectively. In the case of a one-dimensional input (e.g., an array of length H), i is a three-dimensional vector, where i = (i B i C i H Batch normalization is defined according to Equations 1, 2, and 3:

[0063]

[0064] In some instances, sets Defined as , where i C (and j) C ) represents the sub-indexes of i and j along the C-axis (e.g., the channel axis). In some variations, Represents a set cardinality and These are small (e.g., less than a threshold) positive constants. For example, in the case where the input is a two-dimensional image, pixels with the same channel index are normalized together, meaning that batch normalization calculates μ and σ along the (B, H, W) axes. Finally, Transform into ,in and It is i C Learnable and / or trainable parameters of the index. In some variations, and Multiplication is element-wise multiplication and Addition is also element-wise addition. Therefore, and With The same length as the parameter. This can be referred to as a learnable scaling parameter and parameter The learnable shift parameter can be referred to as feature i. The learnable scaling parameter and the learnable shift parameter can be referred to as the affine parameters of the normalization layer 210. In some cases, the i-th batch normalization layer has one or more learnable and / or trainable parameters, including... , ,in and The length is equal to the feature x i The length.

[0065] During training, features are normalized across corresponding subsets of the batch (e.g., mini-batch) using the mean and variance calculated based on corresponding subsets of the batch of training samples. During training, the device can normalize features across the batch for each batch normalization layer. Subsets of representation are computed continuously The expected value, mean and / or average value, and the values ​​spanning the said batch by Subsets of representation are computed continuously The expected value, mean, and / or average value. The device will and The value is stored in the i-th batch normalization layer for inference and / or prediction, for example, when the learning model 202 is implemented at another device and / or the device. In some variations, M B This indicates the total number of values ​​in the subset of the batch used during training.

[0066] To reduce computation and The computational complexity involved in calculating the true value can be reduced by the device using computationally efficient methods. and The estimated value. For example, the device can use the exponential moving average (EMA) method to calculate... and The true value, in the EMA method, is assigned exponentially decreasing weights to past observations. The EMA of the mean and the variance of the mini-batch using parameter τ. Calculate based on equations 4 and 5:

[0067]

[0068] Where M B It is the total number of subsets of batches (e.g., mini-batches) during training and .Notice, and These are the estimates of the true mean and true variance of the training data samples, respectively. In some instances, in addition to the EMA value (e.g., ...), and In addition to, or as an alternative to, the device may use, the EMA value. and ,in yes The estimated value (e.g., calculated using a method different from EMA) and yes The estimated value (calculated using a procedure different from EMA). Used for estimation and Exemplary methods include, but are not limited to, Simple Moving Average (SMA) and Weighted Moving Average (WMA), as well as other methods. In some instances, SMA involves calculating the average of a fixed number of recent data points, and WMA involves assigning different weights to the corresponding data points within a window, thereby giving more recent observations a greater influence on the average.

[0069] At the end of training, the j-th batch normalization layer is stored. and F j It is the number of features at the input of the j-th batch normalization layer. yes The estimated value and yes The estimated value. In some variations, when EMA is used during training to calculate... and When estimating the value, and Therefore, when using EMA, the j-th batch normalization layer stores... and F j This represents the number of features at the input of the j-th batch normalization layer. In some instances, parameters can be learned without calculating gradients and backpropagation. and Calculated via EMA and In the case of simple exponential averaging, the normalization layer for the j-th batch can be learned. and This eliminates the need to calculate gradients and backpropagation. Parameters and The batch normalization layer statistics are typically referred to as those for the j-th batch normalization layer. When EMA is used during training to calculate estimates of the mean and variance, and It is the batch normalized statistical data of the j-th batch normalization layer.

[0070] In some instances, the learning model 202 (e.g., DNN) may contain N norm The number of normalized layers is 210, by It is represented as θ. In some variations, θ A This represents a neural network (e.g., learning model 202) f θ The set of learnable and / or trainable parameters of the normalized layer 210 in the model, where θ A It is a subset of θ (e.g., This is because θ represents the neural network f. θ The set of all learnable and / or trainable parameters. When normalization layer 210 is a batch normalization layer, the j-th batch normalization layer is ,in The j-th normalization layer will introduce F. j A number of learnable and / or trainable scaling parameters and F j A number of learnable and / or trainable shift parameters In some variations, F j It is the total number of features at the input of the j-th batch normalization layer and the corresponding Further including L j The number of scalar parameters (e.g., for) , and ), where L j Is The length of the feature vector at the input, or equivalently, is The number of nodes or neurons in the system.

[0071] At the end of training, the j-th batch normalization layer storage and F j Is The number of features at the input, yes The estimated value and yes The estimated value. In some variations, when EMA is used during training to calculate... and When estimating the value, and Therefore, the learnable and / or trainable parameters introduced by the j-th batch normalization layer also include and (For example, when using EMA to calculate) and In the case of time, they are respectively equivalent to and Therefore, for batch normalization layers... The set of learnable and / or trainable parameters is given by Equation 6:

[0072]

[0073] Where F j Is The total number of features at the input and L j Is The length of the feature vector at the input, or equivalently, is The number of neurons in the brain.

[0074] In some variations, devices in a wireless communication system (e.g., UE and / or NE) may implement one or more learning models 202 for CSI compression, as referenced. Figure 1 As described, and as about Figure 3 Further details are provided. The UE and / or NE can train one or more learning models 202 on a training dataset to perform CSI compression. For example, the training dataset may contain labeled and / or unlabeled data. A dataset is called a labeled dataset when it contains both input data samples and corresponding output data samples for data samples in the dataset. A dataset is called an unlabeled dataset when it contains input data samples but does not have corresponding output data samples.

[0075] In some instances, the device may implement supervised learning techniques to train the learning model 202. For example, the NE102 may train the learning model 202 to map input data to output labels based on exemplary input-output pairs provided during training. In supervised learning, the learning model 202 computes a mapping function from input features to output labels by observing a dataset containing labeled instances, such that the learning model 202 can generalize the mapping to make accurate predictions on new, unseen data. While supervised learning techniques have been described, the device may additionally or alternatively implement any other type of training technique to train the learning model 202, including but not limited to unsupervised and semi-supervised learning techniques, and others. In unsupervised learning, the learning model 202 detects patterns, structures, or relationships within the training data. Unlike supervised learning, no explicit output labels are provided during training. Semi-supervised learning utilizes both labeled and unlabeled data during the training process. The learning model 202 uses labeled instances while also using the structure of the unlabeled data to improve the performance of the learning model 202.

[0076] In some instances, the source domain refers to the set of data samples on which the learning model 202 is trained. For example, when in the data samples When training model 202, This is called the source domain. In some variations, It is a scalar, a one-dimensional vector, or a multi-dimensional vector and It can be a scalar, a one-dimensional vector, or a multi-dimensional vector. For and Joint probability density function (PDF) This represents the source distribution. In some cases, the learning model is trained on more than one source domain and / or on a dataset containing data samples from more than one source domain, resulting in multiple source domains. , where j is a positive integer. The target domain refers to the set of data samples on which the learning model 202 makes inferences and / or predictions. Therefore, when the learning model 202... Perform inference to predict the corresponding label At that time, among them , Called the target domain and PDF This is referred to as the target distribution. In some instances, different target domains may exist for different network conditions, device parameters, channel characteristics, or any other factors affecting the data distribution of one or more data samples. In some instances, the device may obtain input data from the target domain that is not included in the training dataset. Therefore, the learning model can make inferences and / or predictions in domains not used during training (e.g., by compressing CSI derived from a probability distribution different from the distribution of the training dataset).

[0077] Conventionally, to train the learning model 202 to make inferences and / or predictions across different data distributions, the device may implement domain adaptation techniques. For example, the device may perform domain adaptation during the training of the learning model 202 by using domain-invariant feature learning, data mixing and / or data augmentation techniques across multiple source domains during training, and other domain adaptation techniques. In some variations, the domain adaptation techniques implemented during training may involve training the learning model 202 on multiple source domains (e.g., multiple datasets rich in diversity), providing a learning model 202 that generalizes to a target domain not included in the training dataset. However, using domain adaptation techniques implemented during training may not yield a generalized learning model and may utilize a relatively large amount of data (e.g., greater than a threshold) from multiple source domains, increasing the use of computational resources for training.

[0078] Alternatively, the apparatus may perform domain adaptation during the inference and / or testing of the learning model 202 by adapting and / or generalizing the learning model 202 after it has been trained and deployed. The apparatus may obtain labeled or unlabeled data from the target domain for the adaptation and / or generalization of the learning model 202. For example, the apparatus may obtain data in batches, such as by obtaining multiple data samples simultaneously. In some other instances, the apparatus may obtain data from samples one at a time or sequentially. However, conventional techniques for domain adaptation include calculating the value of an end-to-end loss function and performing backpropagation through the network. To calculate the end-to-end loss function for a two-sided learning model, the apparatus hosting the encoder component may use the decoder component output, or the apparatus hosting the decoder component may use the encoder component input. However, the transmission and / or reception of signaling containing decoder component output and / or encoder component input can lead to increased signaling overhead. In some cases, domain adaptation during inference and / or testing of the learning model 202 includes backpropagation, which results in increased signaling overhead between the devices of the managed encoder component and the managed decoder component. During backpropagation, the algorithm works backward through the learning model 202 to compute the gradient of the loss function with respect to the corresponding weights in the learning model 202, thereby determining the contribution of the weights to the total error.

[0079] In some variations, domain adaptation and / or domain generalization techniques adapt and / or adjust the parameters of the learning model 202 by minimizing the loss function, where the calculated loss function value depends on the output of the learning model 202. For a two-sided learning model, the device provides input data samples (e.g., the estimated CSI) to the encoder component, and the output of the learning model 202 is generated by the decoder component. The encoder component and the decoder component may be located at two geographically separate devices. Therefore, the adaptation of the two-sided learning model may involve signaling between the devices to exchange information. That is, when adapting the two-sided learning model to the data distribution of the input data samples at the encoder component, the encoder component may not be able to access the output of the decoder component. To make the decoder component output available at the encoder component, the device may use additional network resources to transmit the decoder component output to the encoder component.

[0080] In some instances, to adapt the learning model 202 (e.g., a one-sided and / or two-sided learning model) without increasing signaling overhead while maintaining relatively low (e.g., below a threshold) processing, power, and memory resource usage, the device may develop a learning model 202 with a normalization layer 210. The device may adapt the parameters of the normalization layer 210 (e.g., rather than adapting all learnable parameters of the learning model 202, including the parameters of the hidden layer 206 and the normalization layer 210) based on the distribution of the input data. Therefore, the device can perform domain adaptation of the learning model 202 (e.g., a learning model 202 implemented at devices in different geographical locations) without calculating the value of the end-to-end loss function, without additional signaling overhead, and without backpropagation through the learning model 202, as per [reference to...]. Figure 3 To describe in further detail.

[0081] Figure 3 The illustrations illustrate examples of a wireless communication system 300 according to various aspects of this disclosure. In some examples, the wireless communication system 300 implements aspects of the wireless communication system 100 and the learning model diagram 200. For example, the wireless communication system 300 includes a UE 104 and an NE 102, which may be as shown in the reference... Figure 1 Examples of UE 104 and NE 102 described herein. In some instances, NE 102 may wirelessly communicate with one or more other devices in the wireless communication system 300. For example, NE 102 may transmit signaling to and / or receive signaling from UE 104. UE 104 may transmit signaling to NE 102 via feedback channel 302, which may be an example of a communication link. Signaling between UE 104 and NE 102 may include control signaling and / or data transmission. For example, signaling may include data samples, such as CSI data samples measured by UE 104.

[0082] In some instances, CSI data samples contain information about the time, frequency, and spatial channels between two nodes or devices in the wireless communication system 300 (e.g., UE 104 and NE 102 (e.g., a base station, gNB, or other network entity)). Through effective combination and precoding, the devices use this information to coherently receive and / or transmit data between themselves. A device receiving data may be referred to as a receiving device, and a device transmitting data may be referred to as a transmitting device. The receiving device may use combination techniques to maximize the quality of the received signal by combining one or more signals received from multiple antennas. The transmitting device may use precoding techniques to maximize the quality of the received signal at the receiver, or to reduce interference at other devices by assigning precoding weights according to the channel gain of the respective antennas and / or by using CSI data to eliminate interference in data transmission.

[0083] UE 104 can use an uplink reference signal (e.g., transmitted from UE 104 to NE 102) to obtain downlink CSI, thereby achieving uplink-downlink channel reciprocity. Uplink-downlink channel reciprocity occurs when the channel characteristics (e.g., channel gain, phase shift, and fading conditions) observed for uplink transmissions from UE 104 to NE 102 are approximately equivalent to the channel characteristics observed for downlink transmissions from NE 102 to UE 104. In the absence of this reciprocity, UE 104 may perform one or more measurements on the downlink reference signal from NE 102 and may transmit CSI feedback (which may be referred to as CSI data) indicating said measurements to NE 102. However, transmitting CSI data to NE 102 results in relatively high signaling overhead on one or more uplink communication resources (e.g., time-frequency resources) allocated for said feedback.

[0084] To reduce signaling overhead, UE 104 can compress CSI data using a quantization scheme that minimizes feedback overhead while maintaining CSI data quality. Exemplary quantization schemes may include, but are not limited to, scalar quantization, vector quantization, adaptive quantization, non-uniform quantization, differential quantization, entropy decoding, and others. In scalar quantization, the elements of the CSI matrix (e.g., channel gain, phase shift) are independently quantized into a finite set of discrete levels. Vector quantization groups multiple elements of the CSI matrix into vectors and performs joint quantization on these vectors. Adaptive quantization dynamically adjusts the quantization levels based on instantaneous channel conditions, providing adaptation to changes in the channel environment (e.g., changes in signal-to-noise ratio (SNR) or channel fading). Adaptive quantization minimizes the trade-off between quantization error and data rate. Non-uniform quantization provides quantization levels that are not uniformly spaced based on the statistical properties of the CSI data, which can improve the accuracy of CSI data representation by allocating additional quantization levels to regions of the CSI matrix with higher variability or importance. Differential quantization encodes the difference rather than the absolute value between consecutive CSI measurements, which reduces the dynamic range of the quantized data and improves the representation efficiency of CSI data, for example, for relatively slow changes in the channel. Entropy decoding techniques, such as Huffman decoding or arithmetic decoding, can be applied to further compress the quantized CSI data by using statistical redundancy. Entropy decoding assigns shorter codewords to more likely CSI values, resulting in a more compact representation.

[0085] In some instances, one or more devices in the wireless communication system 300 (e.g., UE 104 and NE 102) implement an autoencoder. The autoencoder may be an example of a neural network architecture or an ML architecture, wherein data is compressed from its original form to a lower-dimensional representation at encoder component 304, and the data is reconstructed back to its original form at decoder component 306. Encoder component 304 and decoder component 306 may be components of an autoencoder and may implement a corresponding learning model (e.g., as referenced). Figure 2 The described learning model 202) performs data compression and reconstruction. In addition to signal compression, or as an alternative to it, NE 102 and UE 104 may also use an autoencoder for feature extraction and / or noise reduction. The device can train the autoencoder on a large wireless signal dataset, and the autoencoder can learn a compact (e.g., low-dimensional) representation of the signal for subsequent tasks such as modulation classification, channel equalization, and / or interference detection. The autoencoder may additionally or alternatively be used to reduce or eliminate noise in wireless signals by learning to reconstruct a noise-free signal from one or more measurements of a noisy signal, resulting in improved robustness and reliability of the wireless communication system 300 in noisy environments.

[0086] In some instances, such as for signal compression, encoder component 304 receives input data and maps the input data to a low-dimensional representation. The input data may comprise a data sample set 308, including, but not limited to, CSI data obtained from measurements performed by UE 104. For example, UE 104 may measure one or more reference signals (e.g., from NE 102) to obtain CSI data. In some cases, the CSI data contains information describing the characteristics and conditions of the communication channel between the transmitting device (e.g., UE 104 and / or NE 102) and the receiving device (e.g., NE 102 and / or UE 104) in the wireless communication system 300. For example, the CSI data and corresponding data samples may include, but are not limited to, channel gain data, phase shift data, frequency response data, delay spread data, and Doppler shift data, as well as other information. Mapping the input data to a lower-dimensional representation may involve transforming the input data into a compressed representation using multiple layers of a learning model. The output of encoder component 304 is the compressed representation of the input data.

[0087] UE 104 may transmit a compressed representation of input data to NE 102 via feedback channel 302. For example, UE 104 may include a compressed representation of the input data in control signaling sent to NE 102, wherein one or more fields in the control signaling contain the compressed representation of the input data. NE 102 receives the compressed representation of the input data and reconstructs the original input data at decoder component 306. Decoder component 306 uses multiple layers of a neural network to map the compressed representation of the input data back to the original input data. That is, the output of decoder component 306 is a reconstructed version of the input data. For example, NE 102 may reconstruct CSI data to determine one or more characteristics and / or conditions of the communication channel between UE 104 (e.g., a receiver device) and NE 102 (e.g., a transmitter device). NE 102 may use the CSI data for adaptive modulation and decoding, beamforming, precoding, equalization and interference cancellation, and other processes and procedures. Therefore, NE 102 can adapt one or more transmissions between NE 102 and UE 104 to changing channel conditions, thereby improving spectral efficiency, reliability and overall performance.

[0088] In some variations, although encoder component 304 is illustrated as being implemented by UE 104 and decoder component 306 is illustrated as being implemented by NE 102, UE 104 may additionally or alternatively implement decoder component 306 and NE 102 may additionally or alternatively implement encoder component 304.

[0089] In some instances, the automatic encoder includes encoder component 304 E ω This is an instance of a DNN with learnable and / or trainable parameters denoted by ω. An autoencoder may contain a decoder component 306 D. ψThis is an example of a DNN that uses ψ as its set of trainable and / or learnable parameters. Encoder component 304 learns a representation of the input signal and / or data (e.g., encodes the input signal and / or data) such that one or more features of the input signal and / or data are captured as a low-dimensional feature vector 310. Encoder component 304 can obtain a set of data samples 308 and can output a feature vector 310 containing characteristics, attributes, and / or properties (e.g., features) of the data sample set 308. UE 104 can transmit the feature vector 310 to NE 102 via feedback channel 302. For example, UE 104 can transmit control signaling and / or data transmission containing one or more parameters having values ​​including the feature vector 310. Decoder component 306 verifies the feature vector 310 (e.g., the encoding of the data sample set 308) and improves the encoding by regenerating the input signal and / or data from the feature vector 310 generated by encoder component 304. For example, decoder component 306 can generate an estimated set of data samples 312 as an estimate of the data sample set 308. Therefore, encoder component 304 and decoder component 306 are trained and developed together such that the signal and / or data at the input to encoder component 304 is reconstructed at the output of decoder component 306. Thus, two learning models 202 (e.g., E...) ω and D ψ Together, they constitute an automatic encoder, represented as... .

[0090] In some instances, such as for downlink communication from NE 102 to UE 104, the apparatus can train an autoencoder to encode and decode channel matrices. The training dataset may contain a relatively large (e.g., greater than a threshold) number of channel matrices, which may be collected by the apparatus or generated through simulation. The apparatus trains encoder component 304 to generate a low-dimensional latent representation of the input channel matrices, and decoder component 306 reconstructs the channel matrices based on the latent representation generated by encoder component 304. After training, the encoder component 304 of the autoencoder... ω The decoder component 306 D of the autoencoder is deployed at UE 104. ψ Deployed at NE 102. UE 104 uses reference and / or pilot signals received from NE 102 to estimate the channel matrix, using encoder component 304E. ω The channel matrix is ​​encoded, and the encoded output (e.g., the eigenvectors and / or latent representations of the channel matrix) calculated by the encoder component 304 is transmitted to NE102 via feedback channel 302. NE102 uses decoder component 306 D. ψThe channel matrix is ​​decoded and / or reconstructed based on the feature vectors received from UE 104. Because the autoencoder reduces the dimensionality of the data (e.g., CSI), the data is compressed before transmission via feedback channel 302. The compressed CSI data at the output of encoder component 304 contains features and / or feature vectors 310 calculated by encoder component 304.

[0091] In some instances, the transmission device can determine the left singular vector of H and / or H The eigenvectors of H. In addition to implementing an autoencoder for CSI compression, or as an alternative to implementing an autoencoder for CSI compression, the apparatus may also use an autoencoder to transmit the singular vectors or eigenvectors of the channel matrix. The apparatus may train the autoencoder to represent and / or compress a matrix containing singular vectors and / or eigenvectors.

[0092] In some instances, the apparatus (e.g., NE 102 and / or UE 104) can develop and / or train bilateral DNN models or autoencoders (by... This refers to a model that includes normalization layers (e.g., batch normalization layers) between pairs of hidden layers. The batch normalization layer adjusts and / or modifies the statistical distribution of the data samples received from the previous layer before feeding them as input to the next layer. In some cases, a four-layer learning model with one input layer, one output layer, and two hidden layers may contain three normalization layers, as shown in the reference. Figure 2 As described. Therefore, in a total number N L In a multi-layered learning model, with a normalization layer between each pair of neural layers, there will be... Number of normalized layers.

[0093] In some cases, the device can be made of The provided training dataset, or a two-sided learning model for CSI compression, can be used. The source domain dataset. For example, the device can use a training dataset. and self-supervised learning and / or training algorithms to develop and / or train bilateral learning models. The apparatus can determine the architecture and / or structure (number of layers, how layers are connected, etc.) of the learning model, wherein the learning model can be an instance of CNN, RNN, or LSTM, or other types of learning models. The apparatus can determine the architecture and / or structure of the learning model by applying a loss function to the training dataset. Minimize and determine the values ​​of learnable and / or trainable parameters that maximize the performance of the learned model (e.g., accuracy and precision, and other performance metrics).

[0094] In some instances, the learning model (e.g., a two-sided DNN) may include a normalization layer. For example, encoder component 304 and decoder component 306 may include a batch normalization layer such that the batch normalization layer is located between the hidden layer pairs of encoder component 304 and decoder component 306. Alternatively, encoder component 304 may include a batch normalization layer prior to the input layer to normalize the input data before it is provided to encoder component 304. Alternatively, decoder component 306 may include a batch normalization layer located between the output layer of encoder component 304 and the input layer of decoder component 306. After training the two-sided model, encoder component 304 is deployed at a device (e.g., UE 104) and decoder component 306 is deployed at another device (e.g., NE 102). During the training phase, a batch normalization layer existing between the output layer of encoder component 304 and the input layer of decoder component 306 may be located at the end of encoder component 304 and / or before the input layer of decoder component 306 (e.g., when a two-sided learning model is split to deploy encoder component 304 at one device and decoder component 306 at another device). In some instances, the learning model may include a normalization layer located between the output of encoder component 304 and the input of decoder component 306. While the learning model may be described as containing normalization layers located between hidden layers (e.g., and not before the input layer), the learning model may contain any number of normalization layers located anywhere within the learning model.

[0095] In some cases, sets The set includes the learnable parameters of encoder component 304 and the learnable parameters of the batch normalization layer within encoder component 304. If, during training of the learning model, the batch normalization layer is located between encoder component 304 and decoder component 306 and at the output of encoder component 304, then the parameters of the batch normalization layer belong to the set ω. Alternatively, the set... This includes the learnable parameters of decoder component 306, and the learnable parameters of the batch normalization layer within decoder component 306. If, during the training of the learning model, the batch normalization layer is located between encoder component 304 and decoder component 306 and at the input of decoder component 306, then the parameters of the batch normalization layer belong to the set ψ. In some instances, This represents the set of learnable and / or trainable parameters or weights of a two-sided learning model. For example, the set includes learnable and / or trainable parameters for hidden layers and batch normalization layers.

[0096] Once the device selects and / or determines the network architecture for the learning model that includes the normalization layer, the device then uses the training dataset... The model is trained using self-supervised learning by minimizing the loss function. For example, the parameter set is determined by solving Equation 7 using a chosen optimization algorithm (Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (ADAM), etc.). :

[0097]

[0098] in In some instances, the device can handle more than one dataset. Training a two-sided learning model or autoencoder ,in It is a positive integer and The corresponding values ​​are different from each other. Alternatively, the device can train a two-sided learning model on a dataset containing data samples from more than one data distribution.

[0099] In some instances, the device can adapt a trained learning model (e.g., a two-sided DNN) by updating normalized statistics for different normalization layers. In some cases, one or more devices (e.g., UE 104 and / or NE 102) may execute, implement, or deploy a trained and / or developed two-sided learning model or autoencoder. For example, an encoder component is deployed at one device (e.g., UE 104) and a decoder component 306 is deployed at another device (e.g., NE 102), wherein the two devices are geographically separated. After deployment, when operating in a real wireless network (e.g., wireless communication system 300) to compress CSI, the devices can provide input data to the learning model, the input data having a distribution different from that of the training data. When the distribution of the input data changes from... (For example, or in cases where training is performed on multiple data distributions,) (any one of them) is changed into another distribution. At that time, the device can enable two-sided learning models or autoencoders. The device adapts to new input data distributions to meet the threshold performance requirements of the learning model (e.g., a threshold for the learning model's precision and / or accuracy). In some instances, to adapt the learning model, the device may update (change, adjust, replace, modify, etc.) one or more parameters of the learning model, including one or more weights.

[0100] Using techniques involving calculating end-to-end loss or adjusting the parameters of the learning model based on its final output, the device may struggle to adjust and / or retrain the learning model. For example, the final output of a two-sided learning model is the output of decoder component 306, located at a device geographically separate from the device hosting encoder component 304. In some variations, the output of decoder component 306 can be made available at encoder component 304 by transmitting signaling containing the output of decoder component 306, resulting in inefficient use of time-frequency resources due to increased signaling overhead. In some instances, the device can adjust the learning model without calculating the loss function or accessing the output of decoder component 306.

[0101] In some instances, and This indicates the number (e.g., number of values) of batch normalization layers at the encoder component 304 and decoder component 306 that have been trained. At the end of training, the j-th batch normalization layer is stored... and , F j It is the number of features at the input of the j-th batch normalization layer. yes The estimated value and / or the actual value and yes The estimated and / or true values. In some variations, when EMA is used during training to calculate... and When estimating the value, and Parameters can be learned without calculating the loss function, gradients, or backpropagation. and In some cases, such as for calculations via EMA... and The device can determine (e.g., learn) the j-th batch normalization layer using exponential averaging (e.g., without calculating gradients and backpropagation). and .parameter and This includes the batch normalization layer statistics for the j-th batch normalization layer. If the device uses EMA to calculate estimates of the mean and variance during training, then... and It is the batch normalized statistical data of the j-th batch normalization layer.

[0102] In some instances, the distribution of the input data changes to a different distribution than that of the training data. (For example, or in cases where training is performed on multiple data distributions,) (any one of them) Then, the device can adjust the bilateral learning model or the autoencoder. For example, the device can discard batch normalized statistics (e.g., mean and variance values) calculated from source domain data or training data and stored in the corresponding batch normalization layer. and The device can be used according to the new target domain distribution. Distributed data samples During the forward pass, the batch normalization statistics (e.g., mean and variance) of the corresponding batch normalization layer are calculated and / or recalculated. In some variations, adaptations may be performed separately for encoder component 304 and decoder component 306.

[0103] For encoder component 304, the device (e.g., UE 104) can use target domain data samples. Calculate new batch normalization statistics for the corresponding batch normalization layer in encoder component 304. The device can divide data samples into subsets of data samples called mini-batches, where the number of mini-batches is equal to... For encoder component 304, the device performs forward pass-through on the corresponding mini-batches of data samples from the target domain and calculates the mean at the corresponding batch normalization layer for different features. and variance Then, according to ( and variance ( The device calculates the value of ). and The encoder component 304 calculates new values ​​(e.g., using EMA or any other method) and stores these new values ​​in the corresponding batch normalization layer of the encoder component 304 as adapted, adjusted, and / or new values ​​of the batch normalization statistics. For the decoder component 306, the device uses the output from the encoder component 304 to calculate new values ​​of the batch normalization statistics for the corresponding batch normalization layer in the decoder component 306.

[0104] The device provides input data samples to the decoder component 306 for adaptation. These input data samples contain latent and / or feature vectors computed by the encoder component 304 after adapting batch normalized statistics. The device partitions the input data samples received from the encoder component 304 into subsets of data samples called mini-batches, where the number of mini-batches is equal to... For decoder component 306, the device forwards the corresponding mini-batches of data samples received from encoder component 304 and calculates the mean at the corresponding batch normalization layer for different features. and variance Then, the device according to and variance Value calculation and The new value (e.g., via EMA or any other method) and will and The new values ​​are stored in the corresponding batch normalization layer of the decoder component 306. In some instances, the device can select a subset of data samples for a small batch containing a diverse set of data samples by calculating the distances between the data samples. For example, if the data samples are obtained sequentially, the device can use reservoir sampling to select the samples that are most likely to be independent and identically distributed (IID). The IID data samples come from the same underlying population and are obtained independently of each other.

[0105] The device can adapt to new input data distributions Obtain a data sample set The device can obtain according to Independent and identically distributed data samples are used to provide reliable estimates of the new data distribution using the calculated mean and variance of the data samples. If the dataset... If the number of relevant data samples exceeds a threshold, the calculated mean and variance may be biased, and the adaptation may not effectively correct the learning model to the new data distribution. Therefore, the device can use IID data samples to perform the adaptation procedure. The device can perform this adaptation by using data samples from the dataset. IID data samples are obtained by selecting irrelevant data samples. In other words, the device can extract a sample set. ,in It contains diverse samples or data samples with relatively low correlation (e.g., less than a threshold correlation, zero correlation). Alternatively, the device can be configured to... Generate a dataset by generating and / or selecting a set of samples. The resulting set It has relatively high entropy (e.g., greater than the threshold entropy). For example, the device can produce a set. Different subsets can be selected, and the entropy of the corresponding subsets can be calculated. Then, the subset with the largest or highest entropy can be selected.

[0106] In some instances, the device (e.g., UE 104) can obtain data samples. The encoder component 304 is input into the learning model, wherein the encoder component contains Batch normalization layer. For value j, where The j-th batch normalization layer storage in encoder component 304 and F j This is the number of features at the input of the j-th batch normalization layer. In some instances, the decoder component 306 contains... Batch normalization layer. For the value of j, where The j-th batch normalization layer storage in decoder component 306 and F j This is the number of features at the input of the j-th batch normalization layer. UE 104 from the sample set Obtain the sample set ,in yes A subset of , containing data samples that are different, dissimilar, distant, and / or unrelated to each other. Set Can be equal to a set (For example, ).

[0107] In some instances, the device (e.g., UE 104) can be from Generate a set For example, the device can select a non-negative, non-zero distance threshold δ and select a distance metric and / or measure the distance (x). i ,x j ), where distance measurement and / or distance measurement (x i ,x j Returns any two arbitrary data samples x i With x j The measurement of distance between. In some instances, the distance metric and / or measurement is the Euclidean distance, where distance (x) is the distance between two points. i ,x j ) = In some other instances, the distance (x) i ,x j ) = ,in Let x represent the p-th norm. In some other instances, the distance metric and / or measure is the cosine distance between data samples (e.g., by comparing multidimensional data samples x). i ,x j (Consider it as a row vector or column vector). The device can be accessed from a set. Random selection Initialization The device can be set. ,in It is a collection Not in the set The set of elements in set A. The symbol A \ B represents set A minus set B and / or the difference between the two sets A and B, where A \ B is the set containing elements of set A that are not in set B. The apparatus can be derived from the set... Another data sample is randomly selected from the data. The device can use the selected distance metric and / or measure distance (x). i ,x j To calculate With sets The distance between samples in the set. If any of the calculated distances is less than a distance threshold δ, then the device does not include the data sample in the set. And from the set (For example, sets) The sample is removed from the set. If the calculated distance is greater than or equal to a distance threshold δ, the device may include the data sample in the set. (For example, sets) and In some instances, if the set If it is empty, then the device can terminate the use from Generate a set The program. If the set If it is not empty, then the device can return from the set. Another data sample is randomly selected from the data. .

[0108] In some other instances, the device (e.g., UE 104) can be from a set The apparatus generates one or more distinct subsets and calculates the entropy of said subsets. The apparatus may select the subset that produces the highest entropy as the set. The device can collect... The data samples are divided and / or grouped as follows: A number of small batches, wherein the small batches contain A number of data samples, and among them Represents a set The number of elements (e.g., data samples) in the data and This represents the lower bound of the value of g. In some variations, Can be from unit value to set Any non-zero value and / or any integer value of the number of samples (e.g., ).when At that time, the data sample is in a small batch, and when At that time, there exists a small batch containing a sample of available data for tuning the learning model.

[0109] For encoder component 304, the device can perform forward passes on corresponding mini-batches of data samples and can calculate the mean at the corresponding batch normalization layer for different features. and variance Then, according to ( and variance ( The device can calculate the value of ). and The new value (e.g., using EMA or any other method) and can and The new value is stored in the corresponding batch normalization layer of encoder component 304 as an adapted value, adjusted value, and / or new value of batch normalization statistics. In some cases, for the corresponding j-th batch normalization layer at encoder component 304, where As the data samples are passed forward in small batches, the device calculates the mean. ( and variance ( The value of ). The device uses and Calculate the value and , F j This is the number of features at the input of the j-th batch normalization layer. The device stores this information at the j-th batch normalization layer in the encoder assembly 304. and ( The value of ), where .

[0110] For decoder component 306, the device (e.g., NE 102) can provide data samples as input to the decoder neural network for adaptation, the data samples containing latent and / or feature vectors computed by encoder component 304 for the corresponding input data samples, by... This indicates that, in some instances, the output of encoder component 304... Contains latent representations and / or vectors (e.g., feature vectors) for the corresponding input data samples, and the set The data samples included during the adaptation of the two-sided learning model are input to the encoder component 304. The device will collect... The data samples are divided and / or grouped as follows: A number of small batches, wherein the corresponding small batch contains Number of data samples, among which This represents the lower bound of the value of g. In some instances, It can be any nonzero value from a unit value to a value of r, and any integer value (e.g., ).when At that time, the corresponding data sample is considered to be a small batch and when At that time, there exists a small batch containing a sample of available data for tuning the learning model.

[0111] The device can perform forward passes on corresponding mini-batches of data samples in decoder component 306. The device can calculate the mean at the corresponding batch normalization layer for different features. and variance Then, according to ( and variance ( The device can calculate the value of ). and The new value (e.g., using EMA or any other method) and will and The new value is stored in the corresponding batch normalization layer of the decoder component 306 as an adapted value, adjusted value, and / or new value of the batch normalization statistics. With each forward pass of the corresponding mini-batch containing the data sample, the device calculates the mean of the corresponding batch normalization layer of the j-th batch normalization layer at the decoder component 306. ( and variance ( The value of ), where The device is based on and Calculate the value and , F j This is the number of features at the input of the j-th batch normalization layer. The device stores this information at the j-th batch normalization layer in the decoder component 306. and ( The value of ), where .

[0112] Figure 4The diagram illustrates examples of signaling diagram 400 according to various aspects of this disclosure. In some instances, signaling diagram 400 may implement aspects of wireless communication system 100, learning model diagram 200, and wireless communication system 300. Signaling diagram 400 may illustrate examples of devices 402-a and 402-b adapting a bilateral learning model using normalized layer parameters. In some cases, device 402-a may be an example of UE 104 and device 402-b may be an example of NE 102, as referenced. Figures 1 to 3 The following alternative instances may be implemented, some of which may be performed in a different order than described, or not at all. In some cases, processes may include additional features not mentioned below, or other processes may be added.

[0113] In some instances, at 404, devices 402-a and / or 402-b may transmit instructions to update learning model parameters. The learning model parameters may define layers of the learning model implemented at devices 402-a and / or 402-b, for example, referencing... Figure 2 and 3 The described autoencoder. For example, the learning model can be a two-sided learning model that includes an encoder learning model and a decoder learning model.

[0114] In some cases, at 406, device 402-b can transmit data samples to device 402-a. For example, device 402-b can obtain data samples by performing one or more measurements (e.g., measurements for obtaining CSI) and / or by performing data samples via an analog channel. Alternatively, device 402-a can obtain data samples, for example, via analog and / or measurement. Devices 402-b and / or device 402-a can obtain data samples by estimating CSI.

[0115] At 408, device 402-a may select a subset of data samples based on the determination of one or more parameters of at least one normalization layer of the learning model to be updated. The parameters may include one or more normalization statistics of at least one normalization layer (e.g., a batch normalization layer). Normalization statistics may include, but are not limited to, values ​​of a mean parameter and a variance parameter. In some instances, device 402-a may determine that a subset of data samples satisfies one or more conditions. These conditions may include a distance metric between corresponding data samples in the subset being greater than a threshold dissimilarity value and the entropy values ​​of corresponding data samples in the subset satisfying a threshold entropy value. Device 402-a may determine one or more of the threshold dissimilarity value or the threshold entropy value. For example, device 402-a may select a threshold dissimilarity value and / or a threshold entropy value. In some other instances, device 402-a may (e.g., from device 402-b or from another device) receive signaling indicating or configuring the threshold dissimilarity value and / or the threshold entropy value.

[0116] In some variations, device 402-a may determine to update parameters according to periodicity. For example, device 402-a may be configured to and / or determine the periodicity for updating parameters. Device 402-a may update parameters periodically. In some other variations, device 402-a may determine to update parameters after receiving a message at 404. In other variations, device 402-a may detect changes in one or more characteristics of CSI. Device 402-a may determine to update parameters immediately after detecting a change. For example, changes in CSI characteristics may include, but are not limited to, changes in channel gain, phase shift, delay spread, and / or Doppler shift. CSI characteristics may change due to fading, interference, device mobility, environmental factors, and / or antenna configuration changes. In other variations, device 402-a may detect changes in one or more channel conditions for the channel between device 402-a and device 402-b. Device 402-a can immediately determine to update one or more parameters upon detecting the change. Changes in channel conditions may be due to changes in channel fading, changes in channel interference, changes in device mobility of device 402-a and / or device 402-b, changes in environmental factors of the channel (e.g., weather conditions, obstructions, and other environmental factors), and / or changes in the antenna configuration at device 402-a and / or device 402-b. In other variations, device 402-a may determine to update parameters if the quality of the output from the encoder learning model fails to meet a threshold. For example, device 402-a may determine to update parameters if the accuracy and / or precision of the encoder learning model and other quality metrics fail to meet corresponding thresholds.

[0117] At 410, device 402-a can generate one or more updated parameters by providing a subset of data samples as input to the encoder learning model. The updated parameters replace one or more parameters of the normalization layer of the encoder learning model. In some instances, device 402-a can propagate a subset of data samples from the first layer to the last layer of the encoder learning model. Device 402-a can compute one or more normalized statistics for at least one normalization layer. The updated parameters may include, but are not limited to, normalized statistics for at least one normalization layer. Exemplary normalized statistics include, but are not limited to, values ​​for the mean parameter and the variance parameter.

[0118] At 412, device 402-a can transmit the output from the encoder learning model to device 402-b. In some instances, the output is a compressed (e.g., encoded) representation of a subset of data samples. Device 402-b can provide the output from the encoder learning model as input to the decoder learning model.

[0119] At 414, device 402-b can generate updated parameters for the decoder learning model by providing the output from the encoder learning model as input to the decoder learning model. For example, device 402-b can determine one or more parameters of at least one normalization layer (e.g., a batch normalization layer) of the decoder learning model to be updated. The output may contain a set of information samples of feature vectors generated by the encoder learning model. The updated parameters may include, but are not limited to, normalized statistics for at least one normalization layer. The normalized statistics may include values ​​for a mean parameter and a variance parameter.

[0120] In some variations, device 402-b may determine to update parameters based on periodicity (e.g., the same periodicity used by device 402-a to determine whether to update the parameters of the encoder learning model). For example, device 402-b may be configured to and / or determine the periodicity for updating parameters. Device 402-b may update parameters periodically. In some other variations, device 402-b may determine to update parameters after receiving a message at 404 and / or after receiving a message at 412. In other variations, device 402-b may detect changes in one or more characteristics of CSI. Device 402-b may determine to update parameters immediately upon detecting a change. For example, changes in CSI characteristics may include, but are not limited to, changes in channel gain, phase shift, delay spread, and / or Doppler shift. CSI characteristics may change due to fading, interference, device mobility, environmental factors, and / or changes in antenna configuration. In other variations, device 402-b can detect changes in one or more channel conditions for the channel between devices 402-a and 402-b. Device 402-a can immediately determine to update one or more parameters upon detecting the change. Changes in channel conditions may be due to changes in channel fading, changes in channel interference, changes in device mobility of devices 402-a and / or 402-b, changes in environmental factors of the channel (e.g., weather conditions, obstructions, and other environmental factors), and / or changes in antenna configuration at devices 402-a and / or 402-b. In other variations, device 402-b can determine to update parameters if the quality of the output from the decoder learning model fails to meet a threshold. For example, device 402-b can determine to update parameters if the accuracy and / or precision of the decoder learning model and other quality metrics fail to meet corresponding thresholds.

[0121] Device 402-b can propagate a subset of data samples from the first layer of the decoder learning model to the last layer of the decoder learning model. Device 402-b can compute one or more normalized statistics for at least one normalization layer. The normalized statistics may include, but are not limited to, values ​​of the mean parameter and the variance parameter.

[0122] Figure 5 The illustration depicts an example of a UE 500 according to various aspects of this disclosure. The UE 500 may include a processor 502, a memory 504, a controller 506, and a transceiver 508. The processor 502, memory 504, controller 506, or transceiver 508, or various combinations thereof, or various components thereof, may be examples of components for performing the various aspects of this disclosure as described herein. These components may be coupled via one or more interfaces (e.g., operational ground, communication ground, functional ground, electronic ground, electrical ground).

[0123] Processor 502, memory 504, controller 506, or transceiver 508, or various combinations or components thereof, may be implemented in hardware (e.g., a circuit system). The hardware may include a processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), or other programmable logic device, or any combination thereof, configured to or otherwise supporting components for performing the functions described in this disclosure.

[0124] Processor 502 may include intelligent hardware devices (e.g., a general-purpose processor, DSP, CPU, ASIC, field-programmable gate array (FPGA), or any combination thereof). In some embodiments, processor 502 may be configured to operate memory 504. In some other embodiments, memory 504 may be integrated into processor 502. Processor 502 may be configured to execute computer-readable instructions stored in memory 504, thereby enabling UE 500 to perform various functions of this disclosure.

[0125] Memory 504 may include volatile or non-volatile memory. Memory 504 may store computer-readable, computer-executable code containing instructions that, when executed by processor 502, cause UE 500 to perform the various functions described herein. The code may be stored in a non-transitory computer-readable medium, such as memory 504 or another type of memory. Computer-readable medium includes both non-transitory computer storage media and communication media, wherein the communication media includes any medium that facilitates the transfer of a computer program from one location to another. Non-transitory storage media may be any available medium accessible by a general-purpose or special-purpose computer.

[0126] In some implementations, processor 502 and memory 504 coupled to processor 502 may be configured or operable to cause UE 500 to perform one or more of the functions described herein (e.g., instructions stored in memory 504 are executed by processor 502). For example, according to the examples disclosed herein, processor 502 may support wireless communication at UE 500. UE 500 may be configured or operable to support components for performing: selecting a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; updating one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model; and transmitting a message containing the output from the encoder learning model, wherein the output is associated with the subset of data samples, to a second means.

[0127] Additionally, UE 500 may be configured to support any or a combination of the following: updating one or more parameters periodically in order to update the one or more parameters. Alternatively, in order to update the one or more parameters, UE 500 may be configured to support receiving an additional message from at least one of the second or third devices instructing the first device to update the one or more parameters. Alternatively, in order to update the one or more parameters, UE 500 may be configured to support detecting changes in one or more characteristics associated with CSI. Alternatively, in order to update the one or more parameters, UE 500 may be configured to support detecting changes in one or more channel conditions associated with the channel between the first device and the second device. Alternatively, in order to update the one or more parameters, UE 500 may be configured to support determining that the quality of the output from the encoder learning model fails to meet a threshold.

[0128] Alternatively or alternatively, UE 500 may be configured to support obtaining the data sample set based on estimated CSI. Alternatively or alternatively, UE 500 may be configured to support receiving additional messages indicating the data sample set. Alternatively or alternatively, selecting the data sample subset includes determining that the data sample subset satisfies one or more conditions, wherein the one or more conditions include a distance metric between corresponding data samples in the data sample subset being greater than a threshold dissimilarity value and the entropy value of the corresponding data samples in the data sample subset satisfying a threshold entropy value. Alternatively or alternatively, UE 500 may be configured to support determining one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, UE 500 may be configured to support receiving additional messages indicating one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, updating the one or more parameters includes: propagating the subset of data samples from a first layer associated with the encoder learning model to a last layer associated with the encoder learning model; and calculating one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, and wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the at least one normalization layer is a batch normalization layer.

[0129] Alternatively, UE 500 may support at least one memory (e.g., memory 504) and at least one processor (e.g., processor 502), the at least one processor being coupled to the at least one memory and configured or operable to cause the UE to: select a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; update one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing the subset of data samples as input to the encoder learning model; and transmit a message containing the output from the encoder learning model to a second device, wherein the output is associated with the subset of data samples.

[0130] Additionally, the UE 500 may be configured to support any or a combination of the following: To update the one or more parameters, the at least one processor is configured to update the one or more parameters periodically. Alternatively, to update the one or more parameters, the at least one processor is configured to receive an additional message from at least one of the second or third devices instructing the first device to update the one or more parameters. Alternatively, to update the one or more parameters, the at least one processor is configured to detect changes in one or more characteristics associated with the CSI. Alternatively, to update the one or more parameters, the at least one processor is configured to detect changes in one or more channel conditions associated with the channel between the first and second devices. Alternatively, to update the one or more parameters, the at least one processor is configured to determine that the quality of the output from the encoder learning model fails to meet a threshold.

[0131] Alternatively or alternatively, the at least one processor is configured to obtain the data sample set based on the estimated CSI. Alternatively or alternatively, the at least one processor is configured to receive additional messages indicating the data sample set. Alternatively or alternatively, in order to select the data sample subset, the at least one processor is configured to determine that the data sample subset satisfies one or more conditions, wherein the one or more conditions include a distance metric between corresponding data samples in the data sample subset being greater than a threshold dissimilarity value and the entropy value of the corresponding data samples in the data sample subset satisfying a threshold entropy value. Alternatively or alternatively, the at least one processor is configured to determine one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, the at least one processor is configured to receive additional messages indicating one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, to update the one or more parameters, the at least one processor is configured to: propagate the subset of data samples from a first layer associated with the encoder learning model to a last layer associated with the encoder learning model; and calculate one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the at least one normalization layer is a batch normalization layer.

[0132] Controller 506 manages the input and output signals of UE 500. Controller 506 can also manage peripheral devices not integrated into UE 500. In some embodiments, controller 506 may utilize an operating system such as iOS®, ANDROID®, WINDOWS®, or other operating systems. In some embodiments, controller 506 may be implemented as part of processor 502.

[0133] In some embodiments, UE 500 may include at least one transceiver 508. In other embodiments, UE 500 may have more than one transceiver 508. Transceiver 508 may represent a wireless transceiver. Transceiver 508 may include one or more receiver chains 510, one or more transmitter chains 512, or a combination thereof.

[0134] Receiver chain 510 may be configured to receive signals (e.g., control information, data, packets) via wireless media. For example, receiver chain 510 may include one or more antennas for receiving signals via air or wireless media. Receiver chain 510 may include at least one amplifier (e.g., a low-noise amplifier (LNA)) configured to amplify the received signal. Receiver chain 510 may include at least one demodulator configured to demodulate the received signal and obtain transmitted data by reversing the modulation technique applied during signal transmission. Receiver chain 510 may include at least one decoder for decoding the demodulated signal to receive transmitted data.

[0135] Transmitter chain 512 can be configured to generate and transmit signals (e.g., control information, data, packets). Transmitter chain 512 may include at least one modulator for modulating data onto a carrier signal, thereby preparing the signal for transmission over a wireless medium. The at least one modulator may be configured to support one or more techniques, such as amplitude modulation (AM), frequency modulation (FM), or digital modulation schemes like phase shift keying (PSK) or quadrature amplitude modulation (QAM). Transmitter chain 512 may also include at least one power amplifier configured to amplify the modulated signal to an appropriate power level suitable for transmission over a wireless medium. Transmitter chain 512 may also include one or more antennas for transmitting the amplified signal over the air or in the wireless medium.

[0136] Figure 6The illustration depicts an example of a processor 600 according to various aspects of this disclosure. Processor 600 may be an example of a processor configured to perform various operations according to the examples described herein. Processor 600 may include a controller 602 configured to perform various operations according to the examples described herein. Processor 600 may optionally include at least one memory 604, which may be, for example, an L1 / L2 / L3 cache memory. Additionally or alternatively, processor 600 may optionally include one or more arithmetic logic units (ALUs) 606. One or more of these components may be electronically communicated or otherwise coupled (e.g., operative ground, communicative ground, functional ground, electronic ground, electrical ground) via one or more interfaces (e.g., buses).

[0137] Processor 600 may be a processor chipset and includes a protocol stack (e.g., a software stack) executed by the processor chipset to perform various operations (e.g., receiving, obtaining, retrieving, transmitting, outputting, forwarding, storing, determining, identifying, accessing, writing, and reading) according to the examples described herein. The processor chipset may include one or more cores, one or more cache memories (e.g., memory native to or contained within the processor chipset (e.g., processor 600), or other memories (e.g., random access memory (RAM), read-only memory (ROM), dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), static RAM (SRAM), ferroelectric RAM (FeRAM), magnetic RAM (MRAM), resistive RAM (RRAM), flash memory, phase-change memory (PCM), and others)).

[0138] The controller 602 can be configured to manage and coordinate various operations of the processor 600 (e.g., signaling, receiving, acquiring, retrieving, transmitting, outputting, forwarding, storing, determining, identifying, accessing, writing, and reading) to enable the processor 600 to support various operations according to the examples described herein. For example, the controller 602 can operate as a control unit of the processor 600, thereby generating control signals that manage the operation of various components of the processor 600. These control signals include enabling or disabling functional units, selecting data paths, initiating memory accesses, and coordinating operation timing.

[0139] Controller 602 may be configured to fetch (e.g., obtain, retrieve, receive) instructions from memory 604 and determine subsequent instructions to be executed to enable processor 600 to support various operations according to the examples described herein. Controller 602 may be configured to track the memory addresses of instructions associated with memory 604. Controller 602 may be configured to decode instructions to determine the operation to be performed and the operands involved. For example, controller 602 may be configured to interpret instructions and determine control signals to be output to other components of processor 600, thereby enabling processor 600 to support various operations according to the examples described herein. Alternatively or additionally, controller 602 may be configured to manage data flow within processor 600. Controller 602 may be configured to control data transfers between registers, ALU 606, and other functional units of processor 600.

[0140] Memory 604 may include one or more cache memories (e.g., memory native to or contained therein of processor 600, or other memory such as RAM, ROM, DRAM, SDRAM, SRAM, MRAM, flash memory, etc.). In some embodiments, memory 604 may reside within or on the processor chipset (e.g., native to processor 600). In some other embodiments, memory 604 may reside external to the processor chipset (e.g., remote from processor 600).

[0141] Memory 604 may store computer-readable, computer-executable code containing instructions that, when executed by processor 600, cause processor 600 to perform the various functions described herein. The code may be stored in a non-transitory computer-readable medium, such as system memory or another type of memory. Controller 602 and / or processor 600 may be configured to execute the computer-readable instructions stored in memory 604, thereby causing processor 600 to perform various functions. For example, processor 600 and / or controller 602 may be coupled together with or to memory 604, and processor 600 and controller 602 may be configured to perform the various functions described herein. In some instances, processor 600 may include multiple processors and memory 604 may include multiple memories. One or more of the multiple processors may be coupled to one or more of the multiple memories, which may be individually or jointly configured to perform the various functions described herein.

[0142] One or more ALUs 606 may be configured to support various operations according to the examples described herein. In some embodiments, one or more ALUs 606 may reside within or on a processor chipset (e.g., processor 600). In some other embodiments, one or more ALUs 606 may reside external to the processor chipset (e.g., processor 600). One or more ALUs 606 may perform one or more operations on data, such as addition, subtraction, multiplication, and division. For example, one or more ALUs 606 may receive input operands and an opcode that determines the operation to be performed. One or more ALUs 606 may be configured with various logic and arithmetic circuits, including adders, subtractors, shifters, and logic gates, to process and manipulate data according to the operation. Alternatively, one or more ALU 606s may support logical operations such as AND, OR, XOR, NOR, and NAND, thereby enabling one or more ALU 606s to handle conditional operations, comparisons, and bitwise operations.

[0143] According to the examples disclosed herein, processor 600 may support wireless communication. Processor 600 may be configured or operable to support at least one controller (e.g., controller 602), said at least one controller being coupled to at least one memory (e.g., memory 604) and configured or operable to cause the processor to: select a subset of data samples from a set of data samples associated with an encoder learning model, wherein said encoder learning model is associated with a two-sided learning model comprising said encoder learning model and decoder learning model; update one or more parameters corresponding to at least one normalization layer of said encoder learning model based on providing said data sample subset as input to said encoder learning model; and transmit a message containing the output from said encoder learning model to a second device, wherein said output is associated with said data sample subset.

[0144] Additionally, the processor 600 may be configured or operable to support any or a combination of the following: In order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to update the one or more parameters periodically. Alternatively, in order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to receive an additional message from at least one of the device or additional devices instructing the processor to update the one or more parameters. Alternatively, in order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to detect changes in one or more characteristics associated with CSI. Alternatively, in order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to detect changes in one or more channel conditions associated with the channel between the processor and the device. Alternatively, in order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to determine that the quality of the output from the encoder learning model fails to meet a threshold.

[0145] Alternatively or alternatively, the at least one controller is configured or operable to cause the processor to obtain the data sample set based on the estimated CSI. Alternatively or alternatively, the at least one controller is configured or operable to cause the processor to receive additional messages indicating the data sample set. Alternatively or alternatively, in order to select the data sample subset, the at least one controller is configured or operable to cause the processor to determine that the data sample subset satisfies one or more conditions, wherein the one or more conditions include a distance metric between corresponding data samples in the data sample subset being greater than a threshold dissimilarity value and the entropy value of the corresponding data samples in the data sample subset satisfying a threshold entropy value. Alternatively or alternatively, the at least one controller is configured or operable to cause the processor to determine one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, the at least one controller is configured or operable to cause the processor to receive additional messages indicating one or more of the threshold dissimilarity value or the threshold entropy value. Alternatively or alternatively, in order to update the one or more parameters, the at least one controller is configured or operable to cause the processor to: propagate the subset of data samples from a first layer associated with the encoder learning model to a last layer associated with the encoder learning model; and calculate one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the at least one normalization layer is a batch normalization layer.

[0146] Figure 7 The illustration shows an example of an NE 700 according to various aspects of this disclosure. The NE 700 may include a processor 702, a memory 704, a controller 706, and a transceiver 708. The processor 702, memory 704, controller 706, or transceiver 708, or various combinations thereof, or various components thereof, may be examples of components for performing the various aspects of this disclosure as described herein. These components may be coupled via one or more interfaces (e.g., operational ground, communication ground, functional ground, electronic ground, electrical ground).

[0147] Processor 702, memory 704, controller 706, or transceiver 708, or various combinations or components thereof, may be implemented in hardware (e.g., a circuit system). The hardware may include processors, DSPs, ASICs, or other programmable logic devices, or any combination thereof, configured or otherwise supporting components for performing the functions described in this disclosure.

[0148] Processor 702 may include intelligent hardware devices (e.g., a general-purpose processor, DSP, CPU, ASIC, FPGA, or any combination thereof). In some embodiments, processor 702 may be configured to operate memory 704. In some other embodiments, memory 704 may be integrated into processor 702. Processor 702 may be configured to execute computer-readable instructions stored in memory 704, thereby enabling NE 700 to perform various functions of this disclosure.

[0149] Memory 704 may include volatile or non-volatile memory. Memory 704 may store computer-readable, computer-executable code containing instructions that, when executed by processor 702, cause NE 700 to perform the various functions described herein. The code may be stored in a non-transitory computer-readable medium, such as memory 704 or another type of memory. Computer-readable media includes both non-transitory computer storage media and communication media, wherein the communication media includes any medium that facilitates the transfer of a computer program from one location to another. Non-transitory storage media may be any available medium accessible by a general-purpose or special-purpose computer.

[0150] In some embodiments, processor 702 and memory 704 coupled to processor 702 may be configured to cause NE 700 to perform one or more of the functions described herein (e.g., processor 702 executing instructions stored in memory 704). For example, according to the examples disclosed herein, processor 702 may support wireless communication at NE 700. NE 700 may be configured or operable to support components that perform the following operations: receiving from a second means a message containing output from an encoder learning model, wherein: the output is associated with a subset of data samples in a set of data samples; and the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; and updating one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model.

[0151] Additionally, the NE 700 may be configured or operable to support any or a combination of the following: To update the one or more parameters, the method further includes updating the one or more parameters based on periodicity. Alternatively, to update the one or more parameters, the NE 700 may be configured or operable to support receiving from at least one of the second or third devices an additional message instructing the first device to update the one or more parameters. Alternatively, to update the one or more parameters, the NE 700 may be configured or operable to support detecting changes in one or more channel conditions associated with the channel between the first and second devices. Alternatively, to update the one or more parameters, the NE 700 may be configured or operable to support determining that the quality of the output from the decoder learning model fails to meet a threshold.

[0152] Alternatively or alternatively, updating the one or more parameters includes: propagating the subset of data samples from a first layer associated with the decoder learning model to a last layer associated with the decoder learning model; and calculating one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the output includes a set of information samples corresponding to feature vectors associated with the encoder learning model. Alternatively or alternatively, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the at least one normalization layer is a batch normalization layer.

[0153] Alternatively, the NE 700 may support at least one memory (e.g., memory 704) and at least one processor (e.g., processor 702), the at least one processor being coupled to the at least one memory and configured or operable to cause the NE to: receive from a second means a message containing output from an encoder learning model, wherein: the output is associated with a subset of data samples in a set of data samples; and the encoder learning model is associated with a two-sided learning model comprising the encoder learning model and a decoder learning model; and update one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model.

[0154] Additionally, the NE 700 can be configured to support any or a combination of the following: To update the one or more parameters, the at least one processor is configured or operable to cause the NE to update the one or more parameters periodically. Alternatively, to update the one or more parameters, the at least one processor is configured to receive an additional message from at least one of the second or third devices instructing the first device to update the one or more parameters. Alternatively, to update the one or more parameters, the at least one processor is configured to detect changes in one or more channel conditions associated with the channel between the first and second devices. Alternatively, to update the one or more parameters, the at least one processor is configured to determine that the quality of the output from the decoder learning model fails to meet a threshold.

[0155] Alternatively or alternatively, to update the one or more parameters, the at least one processor is configured to: propagate the subset of data samples from a first layer associated with the decoder learning model to a last layer associated with the decoder learning model; and compute one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the output includes a set of information samples corresponding to feature vectors associated with the encoder learning model. Alternatively or alternatively, the one or more parameters include one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include values ​​for a mean parameter and a variance parameter. Alternatively or alternatively, the at least one normalization layer is a batch normalization layer.

[0156] Controller 706 manages the input and output signals of NE 700. Controller 706 can also manage peripheral devices not integrated into NE 700. In some embodiments, controller 706 may utilize an operating system such as iOS®, ANDROID®, WINDOWS®, or other operating systems. In some embodiments, controller 706 may be implemented as part of processor 702.

[0157] In some embodiments, the NE 700 may include at least one transceiver 708. In other embodiments, the NE 700 may have more than one transceiver 708. The transceiver 708 may represent a wireless transceiver. The transceiver 708 may include one or more receiver chains 710, one or more transmitter chains 712, or a combination thereof.

[0158] Receiver chain 710 may be configured to receive signals (e.g., control information, data, packets) via wireless media. For example, receiver chain 710 may include one or more antennas for receiving signals via air or wireless media. Receiver chain 710 may include at least one amplifier (e.g., a low-noise amplifier (LNA)) configured to amplify the received signal. Receiver chain 710 may include at least one demodulator configured to demodulate the received signal and obtain transmitted data by reversing the modulation technique applied during signal transmission. Receiver chain 710 may include at least one decoder for decoding the demodulated signal to receive transmitted data.

[0159] Transmitter chain 712 can be configured to generate and transmit signals (e.g., control information, data, packets). Transmitter chain 712 may include at least one modulator for modulating data onto a carrier signal, thereby preparing the signal for transmission over a wireless medium. The at least one modulator may be configured to support one or more techniques, such as amplitude modulation (AM), frequency modulation (FM), or digital modulation schemes like phase shift keying (PSK) or quadrature amplitude modulation (QAM). Transmitter chain 712 may also include at least one power amplifier configured to amplify the modulated signal to an appropriate power level suitable for transmission over a wireless medium. Transmitter chain 712 may also include one or more antennas for transmitting the amplified signal over the air or in the wireless medium.

[0160] Figure 8 The diagram illustrates a flowchart of a method 800 according to various aspects of this disclosure. The operation of the method can be implemented by a UE as described herein. In some embodiments, the UE can execute a set of instructions to control functional elements of the UE to perform the described functions. It should be noted that the methods described herein describe possible implementations, and the operation and steps can be rearranged or otherwise modified, and other implementations are possible.

[0161] At 802, the method may include selecting a subset of data samples from a set of data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model that includes both an encoder learning model and a decoder learning model. The operation of 802 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 802 may be provided by reference to [reference needed]. Figure 5 The described UE is used to execute.

[0162] At 804, the method may include updating one or more parameters corresponding to at least one normalization layer of the encoder learning model based on providing a subset of data samples as input to the encoder learning model. The operation of 804 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 804 may be provided by reference to... Figure 5 The described UE is used to execute.

[0163] At 806, the method may include transmitting to a second device a message containing the output from the encoder's learning model, wherein the output is associated with a subset of data samples. The operation of 806 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 806 may be provided by reference to... Figure 5 The described UE is used to execute.

[0164] Figure 9 The diagram illustrates a flowchart of a method 900 according to various aspects of this disclosure. The operation of the method can be implemented by an NE as described herein. In some embodiments, the NE may execute a set of instructions to control the functional elements of the NE to perform the described functions. It should be noted that the methods described herein describe possible implementations, and the operation and steps may be rearranged or otherwise modified, and other implementations are possible.

[0165] At 902, the method may include receiving from a second means a message containing output from an encoder learning model, wherein the output is associated with a subset of data samples in a set of data samples and the encoder learning model is associated with a two-sided learning model comprising an encoder learning model and a decoder learning model. The operation of 902 may be performed according to examples as described herein. In some embodiments, aspects of the operation of 902 may be provided by reference to [reference needed]. Figure 7 The NE described is used to execute.

[0166] At 904, the method may include updating one or more parameters corresponding to at least one normalization layer of the decoder learning model based on providing the output as input to the decoder learning model. The operation of 904 may be performed according to the examples described herein. In some embodiments, aspects of the operation of 904 may be provided by reference to... Figure 7 The NE described is used to execute.

[0167] The description herein is provided to enable those skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is given the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A first device comprising: At least one memory; as well as At least one processor, coupled to the at least one memory and operable to enable the first device to: Select a subset of data samples from multiple data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model that includes the encoder learning model and the decoder learning model; At least in part, one or more parameters corresponding to at least one normalization layer of the encoder learning model are updated based on providing the subset of data samples as input to the encoder learning model. and A message including the output from the encoder's learning model is transmitted to the second device, wherein the output is associated with the subset of data samples.

2. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to update the one or more parameters at least in part based on periodicity.

3. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to receive from at least one of the second or third apparatus an additional message instructing the first apparatus to update the one or more parameters.

4. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is operable to enable the first apparatus to detect a change in one or more characteristics associated with the channel state information (CSI).

5. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is further operable to enable the first apparatus to detect a change in one or more channel conditions associated with the channel between the first apparatus and the second apparatus.

6. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to determine that the quality of the output from the encoder learning model fails to meet a threshold.

7. The first apparatus of claim 1, wherein the at least one processor is further operable to enable the first apparatus to obtain the plurality of data samples at least in part based on estimated channel state information (CSI).

8. The first apparatus of claim 1, wherein the at least one processor is further operable to enable the first apparatus to receive additional messages indicating the plurality of data samples.

9. The first apparatus of claim 1, wherein, in order to select the subset of data samples, the at least one processor is operable to cause the first apparatus to determine that the subset of data samples satisfies one or more conditions, wherein the one or more conditions include a distance metric between corresponding data samples in the subset of data samples being greater than a threshold dissimilarity value and the entropy values ​​of corresponding data samples in the subset of data samples satisfying a threshold entropy value.

10. The first apparatus of claim 9, wherein the at least one processor is further operable to enable the first apparatus to determine one or more of the threshold dissimilarity value or the threshold entropy value.

11. The first apparatus of claim 9, wherein the at least one processor is further operable to enable the first apparatus to receive an additional message indicating one or more of the threshold dissimilarity value or the threshold entropy value.

12. The first apparatus of claim 1, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to: The subset of data samples is propagated from the first layer associated with the encoder learning model to the last layer associated with the encoder learning model; and Calculate one or more normalized statistics associated with the at least one normalization layer, wherein the one or more normalized statistics include the values ​​of the mean parameter and the variance parameter.

13. The first apparatus according to claim 1, wherein: The one or more parameters include one or more normalized statistics associated with the at least one normalization layer; The one or more normalized statistical data include the values ​​of the mean parameter and the variance parameter; and The at least one normalization layer is a batch normalization layer.

14. A method performed by a first device, the method comprising: Select a subset of data samples from multiple data samples associated with an encoder learning model, wherein the encoder learning model is associated with a two-sided learning model that includes the encoder learning model and the decoder learning model; At least in part, one or more parameters corresponding to at least one normalization layer of the encoder learning model are updated based on providing the subset of data samples as input to the encoder learning model; and A message including the output from the encoder's learning model is transmitted to the second device, wherein the output is associated with the subset of data samples.

15. A first device comprising: At least one memory; as well as At least one processor, coupled to the at least one memory and operable to enable the first device to: Receive a message from the second device that includes the output of the encoder's learned model, wherein: The output is associated with a subset of data samples from multiple data samples; and The encoder learning model is associated with a two-sided learning model that includes the encoder learning model and the decoder learning model; and At least in part, one or more parameters corresponding to at least one normalization layer of the decoder learning model are updated based on providing the output as input to the decoder learning model.

16. The first apparatus of claim 15, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to update the one or more parameters at least in part based on periodicity.

17. The first apparatus of claim 15, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to receive from at least one of the second or third apparatus an additional message instructing the first apparatus to update the one or more parameters.

18. The first apparatus of claim 15, wherein, in order to update the one or more parameters, the at least one processor is operable to cause the first apparatus to detect a change in one or more channel conditions associated with the channel between the first apparatus and the second apparatus.

19. The first apparatus according to claim 15, wherein: The one or more parameters include one or more normalized statistics associated with the at least one normalization layer; The one or more normalized statistical data include the values ​​of the mean parameter and the variance parameter; and The at least one normalization layer is a batch normalization layer.

20. A method performed by a first device, the method comprising: Receive a message from the second device that includes the output of the encoder's learned model, wherein: The output is associated with a subset of data samples from multiple data samples; and The encoder learning model is associated with a two-sided learning model that includes both the encoder learning model and the decoder learning model; and At least in part, one or more parameters corresponding to at least one normalization layer of the decoder learning model are updated based on providing the output as input to the decoder learning model.

Images